Method: Uses a single-stream Transformer that processes text, video, and audio within a unified token sequence via self-attention only. Combines this with model distillation, latent-space super-resolution, and a Turbo VAE decoder for efficient inference.

Result: Achieves highest visual quality and text alignment among leading open models, with lowest word error rate (14.60%) for speech intelligibility. Wins 80.0% against Ovi 1.1 and 60.9% against LTX 2.3 in human evaluations. Generates 5-second 256p video in 2 seconds on single H100 GPU.

Conclusion: daVinci-MagiHuman demonstrates that a simple single-stream Transformer architecture can effectively generate synchronized audio-video content with strong human-centric performance, multilingual capabilities, and efficient inference.

Abstract: We present daVinci-MagiHuman, an open-source audio-video generative foundation model for human-centric generation. daVinci-MagiHuman jointly generates synchronized video and audio using a single-stream Transformer that processes text, video, and audio within a unified token sequence via self-attention only. This single-stream design avoids the complexity of multi-stream or cross-attention architectures while remaining easy to optimize with standard training and inference infrastructure. The model is particularly strong in human-centric scenarios, producing expressive facial performance, natural speech-expression coordination, realistic body motion, and precise audio-video synchronization. It supports multilingual spoken generation across Chinese (Mandarin and Cantonese), English, Japanese, Korean, German, and French. For efficient inference, we combine the single-stream backbone with model distillation, latent-space super-resolution, and a Turbo VAE decoder, enabling generation of a 5-second 256p video in 2 seconds on a single H100 GPU. In automatic evaluation, daVinci-MagiHuman achieves the highest visual quality and text alignment among leading open models, along with the lowest word error rate (14.60%) for speech intelligibility. In pairwise human evaluation, it achieves win rates of 80.0% against Ovi 1.1 and 60.9% against LTX 2.3 over 2000 comparisons. We open-source the complete model stack, including the base model, the distilled model, the super-resolution model, and the inference codebase.

Method: Proposes OmniCodec with hierarchical multi-codebook design featuring semantic-acoustic decoupling by leveraging pre-trained audio encoder from understanding models, plus self-guidance strategy to improve codebook utilization and reconstruction.

Result: Outperforms Mimi codec at same bitrate, delivering superior reconstruction quality while providing more semantically informative representations that benefit downstream generation tasks.

Conclusion: OmniCodec offers a universal neural audio codec solution for low frame rate across diverse audio domains with improved semantic representation quality for generation tasks.

Abstract: Large Language Models (LLMs) have advanced audio generation through discrete representation learning. However, most existing neural codecs focus on speech and emphasize reconstruction fidelity, overlooking unified low frame rate modeling across diverse audio domains, including speech, music, and general sound. Moreover, high reconstruction quality does not necessarily yield semantically informative representations, limiting effectiveness in downstream generation tasks. We propose OmniCodec, a universal neural audio codec tailored for low frame rate. It adopts a hierarchical multi-codebook design with semantic-acoustic decoupling by leveraging the audio encoder of the pre-trained understanding model, along with a self-guidance strategy to improve codebook utilization and reconstruction. Compared with the Mimi codec, experiments show that OmniCodec achieves outstanding performance at the same bitrate, delivering superior reconstruction quality while also providing more semantically informative representations that benefit downstream generation tasks. Our model and code will be open-sourced. Our demo page is available.

Method: Proposes AcoustEmo with a novel Utterance-Aware Acoustic Q-Former that uses timestamp-synchronized sliding windows to dynamically extract segment-level audio tokens, enabling explicit tracing of temporal evolution of acoustic clues and capturing deep contextual dependencies in dialogues.

Result: Experiments on Explainable Multimodal Emotion Recognition (EMER) show AcoustEmo significantly enhances complex emotion reasoning, outperforming baselines while maintaining robust contextual accuracy.

Conclusion: The proposed time-sensitive approach with fine-grained acoustic modeling improves emotion recognition in multimodal LLMs by better capturing local temporal dynamics.

Abstract: Multimodal Large Language Models (MLLMs) excel in Open-Vocabulary (OV) emotion recognition but often neglect fine-grained acoustic modeling. Existing methods typically use global audio encoders, failing to capture subtle, local temporal dynamics like micro-prosody and intonation shifts within individual utterances. To address this, we propose AcoustEmo, a time-sensitive MLLM featuring a novel Utterance-Aware Acoustic Q-Former. Our approach utilizes a timestamp-synchronized sliding window to dynamically extract segment-level audio tokens instead of coarse global representations. This enables the model to explicitly trace the temporal evolution of subtle acoustic clues and capture deep contextual dependencies in dialogues. Experiments on the Explainable Multimodal Emotion Recognition (EMER) task show that AcoustEmo significantly enhances complex emotion reasoning, outperforming baselines while maintaining robust contextual accuracy.

Method: Embedding Space Separation (ES2) - representation-level fine-tuning that explicitly enlarges distance between harmful and safe representations in embedding space. Uses KL divergence regularization to constrain fine-tuned model’s logits to align with original base model on harmless inputs.

Result: Extensive experiments on open-source LLMs using standard safety benchmarks show substantial safety improvement while maintaining comparable general capabilities.

Conclusion: ES2 effectively improves LLM safety by separating harmful/safe representations in embedding space with minimal impact on general capabilities through KL regularization.

Abstract: Large language models (LLMs) have achieved impressive capabilities, yet ensuring their safety against harmful prompts remains a critical challenge. Recent work has revealed that the latent representations (embeddings) of harmful and safe queries in LLMs typically exhibit linear separability, a property that has been exploited to construct attacks by perturbing the embeddings of harmful queries towards the safe subspace. Motivated by this observation, we propose a representation-level fine-tuning approach, named Embedding Space Separation (ES2), which improves LLM safety by explicitly enlarging the distance between harmful and safe representations in the embedding space. To prevent degradation of model’s general capabilities, we introduce a Kullback-Leibler (KL) divergence regularization term into the loss function, which constrains the logits of the fine-tuned model to align with those of the original base model on harmless inputs. We evaluate our method on several open-source LLMs using standard safety benchmarks. Extensive experimental results demonstrate that our approach substantially improves model safety while maintaining comparable general capabilities.

Method: Created RedacBench from 514 human-authored texts spanning individual, corporate, and government sources, paired with 187 security policies. Used 8,053 annotated propositions to measure both security (removal of sensitive information) and utility (preservation of non-sensitive information).

Result: Experiments show that while more advanced models can improve security, preserving utility remains challenging. The benchmark enables systematic evaluation of redaction strategies across different model capabilities.

Conclusion: RedacBench provides a comprehensive framework for evaluating policy-conditioned redaction, addressing limitations of existing benchmarks and facilitating future research in data security through selective information removal.

Abstract: Modern language models can readily extract sensitive information from unstructured text, making redaction – the selective removal of such information – critical for data security. However, existing benchmarks for redaction typically focus on predefined categories of data such as personally identifiable information (PII) or evaluate specific techniques like masking. To address this limitation, we introduce RedacBench, a comprehensive benchmark for evaluating policy-conditioned redaction across domains and strategies. Constructed from 514 human-authored texts spanning individual, corporate, and government sources, paired with 187 security policies, RedacBench measures a model’s ability to selectively remove policy-violating information while preserving the original semantics. We quantify performance using 8,053 annotated propositions that capture all inferable information in each text. This enables assessment of both security – the removal of sensitive propositions – and utility – the preservation of non-sensitive propositions. Experiments across multiple redaction strategies and state-of-the-art language models show that while more advanced models can improve security, preserving utility remains a challenge. To facilitate future research, we release RedacBench along with a web-based playground for dataset customization and evaluation. Available at https://hyunjunian.github.io/redaction-playground/.

Method: Developed KidGym - a 2D grid-based benchmark with 12 unique tasks targeting five core capabilities. Features diverse scenarios with randomly generated layouts for robust evaluation. The benchmark is fully customizable and extensible, allowing researchers to create new scenarios and adjust difficulty levels.

Result: Evaluation of state-of-the-art MLLMs using KidGym revealed significant insights into model capabilities and identified several limitations of current models. The benchmark provides a more accurate assessment of MLLMs’ cognitive abilities compared to existing evaluations.

Conclusion: KidGym offers a comprehensive, customizable benchmark for evaluating MLLM capabilities, filling a gap in current evaluation methodologies. It enables systematic assessment of MLLMs’ cognitive development potential and reveals important limitations in current models.

Abstract: Multimodal Large Language Models (MLLMs) combine the linguistic strengths of LLMs with the ability to process multimodal data, enbaling them to address a broader range of visual tasks. Because MLLMs aim at more general, human-like competence than language-only models, we take inspiration from the Wechsler Intelligence Scales - an established battery for evaluating children by decomposing intelligence into interpretable, testable abilities. We introduce KidGym, a comprehensive 2D grid-based benchmark for assessing five essential capabilities of MLLMs: Execution, Perception Reasoning, Learning, Memory and Planning. The benchmark comprises 12 unique tasks, each targeting at least one core capability, specifically designed to guage MLLMs’ adaptability and developmental potential, mirroring the stages of children’s cognitive growth. Additionally, our tasks encompass diverse scenarios and objects with randomly generated layouts, ensuring a more accurate and robust evluation of MLLM capabilities. KidGym is designed to be fully user-customizable and extensible, allowing researchers to create new evaluation scenarios and adjust difficuly levels to accommodate the rapidly growing MLLM community. Through the evaluation of state-of-the-art MLLMs using KidGym, we identified significant insights into model capabilities and revealed several limitations of current models. We release our benchmark at: https://kidgym.github.io/KidGym-Website/.

Method: Proposes CRoCoDiL, a unified fine-tuning approach that shifts diffusion to continuous sentence-level semantic space using an encoder-demasker architecture. Introduces two unconditional synthesis algorithms: Continuous-Then-Discrete (ConThenDisc) for hybrid diffusion, and Continuous-Within-Discrete (ConWithinDisc) for multi-diffusion refinement.

Result: Experiments with LLaDA show superior generation quality and more than 10x faster sampling speeds in unconditional settings compared to baseline methods.

Conclusion: Moving diffusion to continuous semantic space significantly improves MDM performance for text generation, offering both quality gains and substantial speed improvements through novel continuous-discrete hybrid approaches.

Abstract: Masked Diffusion Models (MDMs) provide an efficient non-causal alternative to autoregressive generation but often struggle with token dependencies and semantic incoherence due to their reliance on discrete marginal distributions. We address these limitations by shifting the diffusion process into a continuous sentence-level semantic space. We propose CRoCoDiL (Continuous and Robust Conditioned Diffusion for Language), a unified fine-tuning approach that jointly trains an encoder-demasker architecture, grounding the MDM demasking in continuous latent representations. This leads to the formation of a novel autoencoder in which decoding is obtained by an MDM algorithm. Relying on the same framework, we introduce two unconditional text synthesis algorithms: Continuous-Then-Discrete (ConThenDisc), a hybrid-diffusion approach that first generates latent representations in continuous space and then decodes these to tokens via an MDM, and Continuous-Within-Discrete (ConWithinDisc), a multi-diffusion strategy that refines latent representations throughout the discrete sampling process. Experiments using LLaDA show that our methods achieve superior generation quality and more than 10x faster sampling speeds in an unconditional setting.

Method: Introduces Fast-Slow Thinking Reward Models (F/S-RM) inspired by Dual Process Theory. Trains a single model to integrate two reward paradigms: fast thinking (first-token scalar prediction) and slow thinking (CoT-based judgment), regulated by a dual-confidence activation mechanism that determines when to activate slow thinking.

Result: Achieves 1.2% relative performance improvement over state-of-the-art models while reducing token consumption by 20.8%.

Conclusion: F/S-RM provides an effective hybrid approach that balances accuracy and efficiency in reward modeling for RLHF, offering better performance with reduced computational cost.

Abstract: Reward models (RMs) are critical for aligning Large Language Models via Reinforcement Learning from Human Feedback (RLHF). While Generative Reward Models (GRMs) achieve superior accuracy through chain-of-thought (CoT) reasoning, they incur substantial computational costs. Conversely, Scalar Reward Models (SRMs) offer efficiency but suffer from limited performance and adaptability in complex scenarios. We introduce Fast-Slow Thinking Reward Models (F/S-RM), a hybrid RM architecture inspired by Dual Process Theory. It trains a single model to integrate two distinct reward paradigms: first-token prediction as a scalar score (fast thinking) and CoT-based judgment (slow thinking), regulated by a dual-confidence activation mechanism that determines when to activate slow thinking. F/S-RM achieves a 1.2% relative performance improvement over state-of-the-art models while reducing token consumption by 20.8%. Code and data will be publicly available.

Method: Proposes MAD-M² (multi-agent debate with memory masking), where LLM agents can mask erroneous memories from previous debate rounds at the beginning of each round, preserving informative memories while discarding erroneous ones.

Result: Extensive experiments on mathematical and logical reasoning benchmarks show MAD-M² can identify erroneous memories and achieve better reasoning performance than standard MAD.

Conclusion: Memory masking effectively improves the robustness of multi-agent debate frameworks by filtering out erroneous memories, leading to enhanced reasoning capabilities in LLMs.

Abstract: Large language models (LLMs) have recently demonstrated impressive capabilities in reasoning tasks. Currently, mainstream LLM reasoning frameworks predominantly focus on scaling up inference-time sampling to enhance performance. In particular, among all LLM reasoning frameworks, multi-agent debate (MAD), which employs multiple LLMs as agents to perform reasoning in the way of multi-round debate, has emerged as a powerful reasoning paradigm since it allows agents to access previous memories to alleviate fallacious content and refine their reasoning iteratively in each debate round. However, although MAD significantly improves the reasoning capabilities of LLMs, in this paper, we observe that there remain erroneous memories, and LLM agents are vulnerable to these erroneous memories. To explore this phenomenon, we provide a theoretical insight that the performance of MAD is highly dependent on the quality of memories derived from the previous debate, indicating that the existence of erroneous memories poses a threat to the performance of MAD. To address this problem, we introduce a simple yet effective multi-agent debate framework, multi-agent debate with memory masking (MAD-M$^2$), to improve the robustness of MAD by allowing LLM agents to mask erroneous memories from the previous debate round at the beginning of each debate round. In this way, MAD-M$^2$ can polish the contextual information before each debate round by preserving informative and meaningful memories while discarding the erroneous memories. Extensive experiments and analyses on mainstream mathematical and logical reasoning benchmarks demonstrate that MAD-M$^2$ can identify the erroneous memories and achieve better performance in reasoning than MAD.

Method: CoDiLA uses a small auxiliary autoregressive model (0.6B parameters) operating on diffusion latents to handle local decoding within blocks, while maintaining parallel block generation and bidirectional modeling across blocks

Result: Eliminates coherence artifacts and establishes new Pareto frontier for accuracy and speed in code generation benchmarks

Conclusion: Hybrid approach combining diffusion models with local autoregression effectively reconciles parallel sampling with dependency modeling for improved code generation

Abstract: Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) models, offering sub-linear generation latency and bidirectional capabilities that are particularly appealing for code generation and editing. Achieving sub-linear latency in discrete DLMs requires predicting multiple tokens in parallel. However, standard DLMs sample tokens independently from conditional marginal distributions, failing to capture the joint dependencies among concurrently generated tokens. As a result, they often lead to syntactic inconsistencies and break multi-token structures. In this work, we introduce CoDiLA (Coherent Diffusion with Local Autoregression), a method that reconciles parallel sampling with local dependency modeling. Rather than forcing the DLM to resolve fine-grained syntax, CoDiLA delegates local decoding to a small, auxiliary AR model operating on the diffusion latents. This design allows for parallel block generation while ensuring sequential validity within each block and maintaining core DLM capabilities, including bidirectional modeling across blocks. We demonstrate that using a highly compact auxiliary AR model (e.g., 0.6B parameters) effectively eliminates coherence artifacts, establishing a new Pareto frontier for accuracy and speed in code generation benchmarks.

Method: Develop Expected Reward Prediction (ERP) that estimates the reward an LLM would earn under repeated sampling, then use this for model routing protocols that select which model should handle each prompt at inference time.

Result: ERP routing outperforms baselines that route based on average category performance, demonstrates precision and discriminative power, and works effectively with model pools including Llama3.1-Instruct and Gemma models.

Conclusion: Expected reward prediction enables effective model routing to maximize reward while controlling computational cost, with trivial extensibility to new models and explaining the success of more complex routing protocols.

Abstract: Reward models are a standard tool to score responses from LLMs. Reward models are built to rank responses to a fixed prompt sampled from a single model, for example to choose the best of n sampled responses. In this paper, we study whether scores from response-level reward models lifted to score a model’s suitability for a prompt, prior to seeing responses from that model. Specifically, we show that it is straightforward to predict the expected reward that an LLM would earn from the reward model under repeated sampling. Further, we show that these expected reward predictions are precise and discriminative enough to support an application to a model routing protocol that routes prompts to models at inference time to maximize reward while controlling computational cost. We demonstrate the performance of this routing procedure on the open-perfectblend dataset, using a model pool composed of Llama3.1-Instruct 8B/70B, Gemma2-IT 9B/27B, and Gemma1-IT 7B models. Our simple expected reward prediction–based routing (ERP) outperforms baselines that route prompts to models with the best average performance within each prompt’s category, and explains the success of more complex routing protocols that implicitly estimate an expected reward. Our approach has the added advantage of being trivially extensible as new models are added to the pool.

Method: Conducts extensive experimental evaluation of CLC systems, identifies that existing CLC techniques are complementary, and proposes a new CLC design that carefully combines these techniques.

Result: Shows existing CLC approaches have fundamental limitations, and the proposed combined CLC design achieves better accuracy than individual techniques.

Conclusion: By recognizing the complementary nature of existing CLC techniques and combining them thoughtfully, it’s possible to overcome fundamental limitations and achieve improved accuracy in retrieval-augmented generation systems.

Abstract: Retrieval-augmented generation improves large language models’ accuracy by adding relevant retrieved text to the prompt. Chunk level caching (CLC) accelerates inference by precomputing KV caches for these retrieved chunks and reusing them. However, these caches miss cross-attention dependencies between chunks, which can reduce output quality. Several methods try to improve CLC accuracy using different techniques. We make two main contributions. First, we show that existing CLC approaches have fundamental limitations that limit their accuracy or their applicability. We back this conclusion with an extensive CLC system experimental evaluation. Second, we observe that existing CLC techniques are complementary. We leverage this insight to propose a new CLC design that carefully combines them and achieves better accuracy.

Method: Introduces latent lookahead training where at selected positions, before committing to next token, the model performs τ-step lookahead in latent space by recursively feeding hidden states back into context, supervised against next τ ground-truth tokens.

Result: Latent lookahead substantially outperforms both autoregressive and non-autoregressive baselines on planning tasks like maze solving, Sudoku, and ProsQA where foresight is essential.

Conclusion: The approach enables models to “think” before generating, addressing limitations of standard next-token prediction by allowing exploration of multiple continuations and adaptive compute allocation.

Abstract: Autoregressive language models trained with next-token prediction generate text by sampling one discrete token at a time. Although very scalable, this objective forces the model to commit at every step, preventing it from exploring or reflecting upon multiple plausible continuations. Furthermore, the compute allocation across tokens is uniform; every token is formed based on a single forward-pass, potentially limiting the model’s expressiveness in cases where difficult tokens require inherently more compute. Towards addressing these limitations, we introduce latent lookahead, a training strategy that enables models to “think” before generating: at selected positions in the sequence, before committing to the next token, the model performs a multi-step lookahead in latent space. More precisely, instead of sampling future tokens, we leverage the network’s latent space by recursively feeding its hidden states back into the context for $τ$ steps, investing more compute on predicting that token. This produces $τ$ latent predictions that are supervised against the next $τ$ ground-truth tokens, encouraging the model to “lookahead” and refine its prediction. We show that latent lookahead substantially outperforms both autoregressive and non-autoregressive baselines on planning tasks such as maze solving, Sudoku, and ProsQA, where foresight is essential.

Method: Extracted emotion-specific linguistic signatures from 13 English datasets and evaluated how incorporating these features into transformer models impacts performance. Used RoBERTa-based models enriched with high-level linguistic features.

Result: Models achieved consistent performance gains of up to +2.4 macro F1 on the GoEmotions benchmark, showing that explicit lexical cues can complement neural representations and improve robustness in predicting emotion categories.

Conclusion: Linguistic features provide reliable interpretable signals for emotion recognition and can effectively complement transformer-based neural representations to improve emotion detection performance.

Abstract: Emotion detection is a central problem in NLP, with recent progress driven by transformer-based models trained on established datasets. However, little is known about the linguistic regularities that characterize how emotions are expressed across different corpora and labels. This study examines whether linguistic features can serve as reliable interpretable signals for emotion recognition in text. We extract emotion-specific linguistic signatures from 13 English datasets and evaluate how incorporating these features into transformer models impacts performance. Our RoBERTa-based models enriched with high level linguistic features achieve consistent performance gains of up to +2.4 macro F1 on the GoEmotions benchmark, showing that explicit lexical cues can complement neural representations and improve robustness in predicting emotion categories.

Method: Analyzes trade-offs in test-time compute strategies comparing SLMs with reasoning (CoT, Majority Voting) vs LLMs using MMLU benchmark. Examines transformer input-output token dynamics and proposes energy efficiency metrics (Energy-per-Token). Introduces controlled reasoning with operating curves to dynamically regulate reasoning depth.

Result: Shows SLMs with reasoning strategies can approach LLM performance but with energy trade-offs. Transformer architectures have nonlinear hardware energy operation curves. Energy efficiency metrics complement traditional accuracy benchmarks.

Conclusion: Proposes energy-aware routing mechanism integrating model selection and inference strategies to balance accuracy with sustainable AI deployment through controlled reasoning and energy efficiency metrics.

Abstract: Large Language Models (LLMs) demonstrate exceptional performance across diverse tasks but come with substantial energy and computational costs, particularly in request-heavy scenarios. In many real-world applications, the full scale and capabilities of LLMs are often unnecessary, as Small Language Models (SLMs) can provide accurate responses for simpler text generation tasks. When enhanced with advanced reasoning strategies, such as Chain-of-Thought (CoT) prompting or Majority Voting, SLMs can approach the performance of larger models while reducing overall computational requirements. However, these strategies can also introduce additional energy costs, creating an energy-accuracy trade-off. Our analysis examines these trade-offs in test-time compute strategies for smaller models compared to larger ones, using the MMLU benchmark. Additionally, we explore the input-output token dynamics of transformer architectures, which result in nonlinear hardware energy operation curves for LLMs. To bridge AI research with its physical impact, we propose \textit{energy efficiency metrics}, including Energy-per-Token, as complements to traditional accuracy benchmarks. Beyond model selection, we propose controlled reasoning in CoT token generation, using operating curves to regulate reasoning depth dynamically. This vision integrates a energy-aware routing mechanism, ensuring that model selection and inference strategies balance accuracy for sustainable AI deployment.

Method: Created Abjad-Kids dataset with 46,397 audio samples from children aged 3-12, then proposed hierarchical CNN-LSTM classification with two grouping strategies: static linguistic-based and dynamic clustering-based, using data augmentation and regularization.

Result: Static linguistic-based grouping achieved superior performance, CNN-LSTM with data augmentation outperformed traditional ML, but experiments showed overfitting challenges due to limited samples despite augmentation.

Conclusion: Abjad-Kids addresses the gap in Arabic children’s speech datasets and shows promise for educational applications, though more data collection is needed to overcome overfitting issues.

Abstract: Speech-based AI educational applications have gained significant interest in recent years, particularly for children. However, children speech research remains limited due to the lack of publicly available datasets, especially for low-resource languages such as Arabic.This paper presents Abjad-Kids, an Arabic speech dataset designed for kindergarten and primary education, focusing on fundamental learning of alphabets, numbers, and colors. The dataset consists of 46397 audio samples collected from children aged 3 - 12 years, covering 141 classes. All samples were recorded under controlled specifications to ensure consistency in duration, sampling rate, and format. To address high intra-class similarity among Arabic phonemes and the limited samples per class, we propose a hierarchical audio classification based on CNN-LSTM architectures. Our proposed methodology decomposes alphabet recognition into a two-stage process: an initial grouping classification model followed by specialized classifiers for each group. Both strategies: static linguistic-based grouping and dynamic clustering-based grouping, were evaluated. Experimental results demonstrate that static linguistic-based grouping achieves superior performance. Comparisons between traditional machine learning with deep learning approaches, highlight the effectiveness of CNN-LSTM models combined with data augmentation. Despite achieving promising results, most of our experiments indicate a challenge with overfitting, which is likely due to the limited number of samples, even after data augmentation and model regularization. Thus, future work may focus on collecting additional data to address this issue. Abjad-Kids will be publicly available. We hope that Abjad-Kids enrich children representation in speech dataset, and be a good resource for future research in Arabic speech classification for kids.

Method: Multitask Transformer-based sequence-to-sequence model that jointly predicts phoneme sequences, word sequences, and auxiliary acoustic features from area 6v intracortical recordings. Introduced Neural Hammer Scalpel (NHS) calibration module combining global alignment with feature-wise modulation to address day-to-day nonstationarity.

Result: Achieved state-of-the-art phoneme error rate of 14.3% on Willett et al. dataset. Word decoding reached 25.6% WER with direct decoding and 19.4% WER with candidate generation and rescoring. NHS substantially improved both phoneme and word decoding compared to linear or no day-specific transform.

Conclusion: Contextual sequence-to-sequence modeling improves neural-to-phoneme readout fidelity from intracortical speech signals. Attention-based analyses provide insights into how neural speech evidence is segmented and accumulated over time, offering interpretability benefits.

Abstract: Speech brain–computer interfaces require decoders that translate intracortical activity into linguistic output while remaining robust to limited data and day-to-day variability. While prior high-performing systems have largely relied on framewise phoneme decoding combined with downstream language models, it remains unclear what contextual sequence-to-sequence decoding contributes to sublexical neural readout, robustness, and interpretability. We evaluated a multitask Transformer-based sequence-to-sequence model for attempted speech decoding from area 6v intracortical recordings. The model jointly predicts phoneme sequences, word sequences, and auxiliary acoustic features. To address day-to-day nonstationarity, we introduced the Neural Hammer Scalpel (NHS) calibration module, which combines global alignment with feature-wise modulation. We further analyzed held-out-day generalization and attention patterns in the encoder and decoders. On the Willett et al. dataset, the proposed model achieved a state-of-the-art phoneme error rate of 14.3%. Word decoding reached 25.6% WER with direct decoding and 19.4% WER with candidate generation and rescoring. NHS substantially improved both phoneme and word decoding relative to linear or no day-specific transform, while held-out-day experiments showed increasing degradation on unseen days with temporal distance. Attention visualizations revealed recurring temporal chunking in encoder representations and distinct use of these segments by phoneme and word decoders. These results indicate that contextual sequence-to-sequence modeling can improve the fidelity of neural-to-phoneme readout from intracortical speech signals and suggest that attention-based analyses can generate useful hypotheses about how neural speech evidence is segmented and accumulated over time.

Method: Created TaigiSpeech dataset with 21 speakers and 3k utterances. Explored two data mining strategies: 1) keyword match with LLM pseudo-labeling via intermediate language, and 2) audio-visual framework leveraging multimodal cues with minimal textual supervision.

Result: Developed a scalable dataset construction approach for low-resource and unwritten spoken languages. The dataset will be released under CC BY 4.0 license to facilitate research on underrepresented languages.

Conclusion: TaigiSpeech addresses the scarcity of labeled data for low-resource languages and enables practical intent detection applications, with potential for broader adoption in speech technology research for underrepresented languages.

Abstract: Speech technologies have advanced rapidly and serve diverse populations worldwide. However, many languages remain underrepresented due to limited resources. In this paper, we introduce \textbf{TaigiSpeech}, a real-world speech intent dataset in Taiwanese Taigi (aka Taiwanese Hokkien/Southern Min), which is a low-resource and primarily spoken language. The dataset is collected from older adults, comprising 21 speakers with a total of 3k utterances. It is designed for practical intent detection scenarios, including healthcare and home assistant applications. To address the scarcity of labeled data, we explore two data mining strategies with two levels of supervision: keyword match data mining with LLM pseudo labeling via an intermediate language and an audio-visual framework that leverages multimodal cues with minimal textual supervision. This design enables scalable dataset construction for low-resource and unwritten spoken languages. TaigiSpeech will be released under the CC BY 4.0 license to facilitate broad adoption and research on low-resource and unwritten languages. The project website and the dataset can be found on https://kwchang.org/taigispeech.

Method: Created FinBench-QA-Hallucination benchmark with 755 annotated examples from 300 SEC 10-K filing pages, using conservative evidence-linkage protocol requiring support from both textual chunks and extracted relational triplets. Evaluated six detection approaches: LLM judges, fine-tuned classifiers, NLI models, span detectors, and embedding-based methods under clean and noisy triplet conditions.

Result: LLM-based judges and embedding approaches achieved highest performance (F1: 0.82-0.86) under clean conditions. Most methods degraded significantly with noisy triplets (MCC dropping 44-84%), while embedding methods remained relatively robust (only 9% degradation). Statistical tests confirmed significant performance differences (p < 0.001).

Conclusion: The benchmark reveals vulnerabilities in current KG-augmented systems and provides insights for building reliable financial information systems. It offers a framework for integrating AI reliability evaluation into high-stakes domains beyond finance, including healthcare, legal, and government applications.

Abstract: As organizations increasingly integrate AI-powered question-answering systems into financial information systems for compliance, risk assessment, and decision support, ensuring the factual accuracy of AI-generated outputs becomes a critical engineering challenge. Current Knowledge Graph (KG)-augmented QA systems lack systematic mechanisms to detect hallucinations - factually incorrect outputs that undermine reliability and user trust. We introduce FinBench-QA-Hallucination, a benchmark for evaluating hallucination detection methods in KG-augmented financial QA over SEC 10-K filings. The dataset contains 755 annotated examples from 300 pages, each labeled for groundedness using a conservative evidence-linkage protocol requiring support from both textual chunks and extracted relational triplets. We evaluate six detection approaches - LLM judges, fine-tuned classifiers, Natural Language Inference (NLI) models, span detectors, and embedding-based methods under two conditions: with and without KG triplets. Results show that LLM-based judges and embedding approaches achieve the highest performance (F1: 0.82-0.86) under clean conditions. However, most methods degrade significantly when noisy triplets are introduced, with Matthews Correlation Coefficient (MCC) dropping 44-84 percent, while embedding methods remain relatively robust with only 9 percent degradation. Statistical tests (Cochran’s Q and McNemar) confirm significant performance differences (p < 0.001). Our findings highlight vulnerabilities in current KG-augmented systems and provide insights for building reliable financial information systems, where hallucinations can lead to regulatory violations and flawed decisions. The benchmark also offers a framework for integrating AI reliability evaluation into information system design across other high-stakes domains such as healthcare, legal, and government.

Method: SciNav framework uses pairwise relative judgments within tree search to select top-K promising solution branches, prune low-potential ones, and progressively narrow down solution candidates guided by relative comparisons, operating under constrained search budgets.

Result: SciNav significantly outperforms direct prompting and prior agents (OpenHands, Self-Debug) across different base models, task types, and difficulty levels, and exceeds frontier comparators like random selection and LLM absolute scoring.

Conclusion: The framework demonstrates the effectiveness of relative judgment-guided top-K search for high-quality scientific coding, marking progress toward more practical science agents for objective, executable tasks.

Abstract: Autonomous science agents built on large language models (LLMs) are increasingly used to generate hypotheses, design experiments, and produce reports. However, prior work mainly targets open-ended scientific problems with subjective outputs that are difficult to evaluate. Scientific coding benchmarks, by contrast, provide executable outputs for objective assessment. Existing approaches remain engineering-driven pipelines, revealing the need for structured, end-to-end science agent frameworks for scientific coding tasks. We address this gap by focusing on scientific coding tasks, where evaluation can be made rigorously, and introducing an agent framework SciNav (Scientific Navigator) that enables more effective solution exploration. Our framework is designed to operate under constrained search budgets, moving beyond reliance on pre-defined success metrics and prolonged search cycles. Inspired by findings that comparative judgments often reveal finer-grained quality differences and therefore provide greater discriminative power than absolute scoring, our framework leverages pairwise relative judgments within a tree search process to select top-K promising solution branches, prune low-potential ones, and progressively narrow down the solution candidates on the selected branches guided by relative comparisons. We demonstrate our agent’s effectiveness across different types of tasks on two benchmarks. Experiments show that SciNav significantly outperforms direct prompting and prior agents like OpenHands and Self-Debug across different base models, task types, and difficulty levels, and exceeds different frontier comparators such as random selection and LLM absolute scoring. These results confirm the strength of our agent design and highlight the effectiveness of relative judgment-guided top-K search for high-quality scientific coding, marking a step toward more practical science agents.

Method: TiCo uses Spoken Time Markers (e.g., <10.6 seconds>) to help models estimate elapsed speaking time during generation. It requires minimal data and no additional QA pairs, using self-generation and reinforcement learning for efficient training.

Result: TiCo significantly improves adherence to duration constraints while preserving response quality, addressing the time-control limitations of existing spoken dialogue models.

Conclusion: TiCo provides a simple and effective solution for enabling time-aware spoken dialogue generation, enhancing the practical utility of spoken language systems.

Abstract: We propose TiCo, a simple post-training method for enabling spoken dialogue models (SDMs) to follow time-constrained instructions and generate responses with controllable duration. This capability is valuable for real-world spoken language systems such as voice assistants and interactive agents, where controlling response duration can improve interaction quality. However, despite their strong ability to generate natural spoken responses, existing models lack time awareness and struggle to follow duration-related instructions (e.g., “Please generate a response lasting about 15 seconds”). Through an empirical evaluation of both open-source and commercial SDMs, we show that they frequently fail to satisfy such time-control requirements. TiCo addresses this limitation by enabling models to estimate elapsed speaking time during generation through Spoken Time Markers (STM) (e.g., <10.6 seconds>). These markers help the model maintain awareness of time and adjust the remaining content to meet the target duration. TiCo is simple and efficient: it requires only a small amount of data and no additional question-answer pairs, relying instead on self-generation and reinforcement learning. Experimental results show that TiCo significantly improves adherence to duration constraints while preserving response quality.

Method: Proposes a segmental-level prosodic probing framework comparing synthetic and natural speech realizations for thousands of words stratified by lexical frequency. Uses Tacotron 2 and FastSpeech 2 trained on LJ Speech corpus, complemented by large-scale evaluation across multiple advanced TTS systems.

Result: TTS models accurately reproduce f0 perturbation for high-frequency words but show poor generalization to low-frequency items, suggesting reliance on lexical-level memorization rather than abstract segmental-prosodic encoding.

Conclusion: The study reveals a limitation in TTS systems’ ability to generalize prosodic detail beyond seen data, and proposes a linguistically informed diagnostic framework for future TTS evaluation with implications for interpretability and authenticity assessment.

Abstract: This study proposes a segmental-level prosodic probing framework to evaluate neural TTS models’ ability to reproduce consonant-induced f0 perturbation, a fine-grained segmental-prosodic effect that reflects local articulatory mechanisms. We compare synthetic and natural speech realizations for thousands of words, stratified by lexical frequency, using Tacotron 2 and FastSpeech 2 trained on the same speech corpus (LJ Speech). These controlled analyses are then complemented by a large-scale evaluation spanning multiple advanced TTS systems. Results show accurate reproduction for high-frequency words but poor generalization to low-frequency items, suggesting that the examined TTS architectures rely more on lexical-level memorization than on abstract segmental-prosodic encoding. This finding highlights a limitation in such TTS systems’ ability to generalize prosodic detail beyond seen data. The proposed probe offers a linguistically informed diagnostic framework that may inform future TTS evaluation methods, and has implications for interpretability and authenticity assessment in synthetic speech.

Method: Measure CHSH |S| parameter (Bell inequality violations) across inference parameter space of models spanning four orders of magnitude in scale, cross-reference with MMLU, hallucination rate, and nonsense detection benchmarks, and analyze variations with sampling parameters and word order.

Result: The interquartile range of |S| distribution is completely orthogonal to all external benchmarks, while violation rate shows weak anticorrelation with benchmarks that doesn’t reach significance. Contextuality metrics provide different insights than traditional performance measures.

Conclusion: Quantum-like contextuality in LLMs represents a fundamental aspect of semantic processing that is distinct from traditional performance metrics, with implications for prompt injection defenses and social manipulation through context shaping.

Abstract: Understanding the fundamental mechanisms governing the production of meaning in the processing of natural language is critical for designing safe, thoughtful, engaging, and empowering human-agent interactions. Experiments in cognitive science and social psychology have demonstrated that human semantic processing exhibits contextuality more consistent with quantum logical mechanisms than classical Boolean theories, and recent works have found similar results in large language models – in particular, clear violations of the Bell inequality in experiments of contextuality during interpretation of ambiguous expressions. We explore the CHSH $|S|$ parameter – the metric associated with the inequality – across the inference parameter space of models spanning four orders of magnitude in scale, cross-referencing it with MMLU, hallucination rate, and nonsense detection benchmarks. We find that the interquartile range of the $|S|$ distribution – the statistic that most sharply differentiates models from one another – is completely orthogonal to all external benchmarks, while violation rate shows weak anticorrelation with all three benchmarks that does not reach significance. We investigate how $|S|$ varies with sampling parameters and word order, and discuss the information-theoretic constraints that genuine contextuality imposes on prompt injection defenses and its human analogue, whereby careful construction and maintenance of social contextuality can be carried out at scale – manufacturing not consent but contextuality itself, a subtler and more fundamental form of manipulation that shapes the space of possible interpretations before any particular one is reached.

Method: Use coding agents to organize text in file systems and manipulate it using native tools (executable code and terminal commands). Evaluate off-the-shelf frontier coding agents on long-context reasoning, retrieval-augmented generation, and open-domain QA with up to three trillion tokens.

Result: Coding agents outperform published state-of-the-art by 17.3% on average across multiple benchmarks. Key factors: native tool proficiency (using executable code rather than passive semantic queries) and file system familiarity (navigating massive corpora as directory structures).

Conclusion: Delegating long-context processing to coding agents offers an effective alternative to semantic search or context window scaling, opening new directions for long-context processing in LLMs.

Abstract: Large Language Models (LLMs) have demonstrated remarkable progress in scaling to access massive contexts. However, the access is via the latent and uninterpretable attention mechanisms, and LLMs fail to effective process long context, exhibiting significant performance degradation as context length increases. In this work, we study whether long-context processing can be externalized from latent attention into explicit, executable interactions, by allowing coding agents to organize text in file systems and manipulate it using its native tools. We evaluate off-the-shelf frontier coding agents as the general interface for tasks that require processing long contexts, including long-context reasoning, retrieval-augmented generation, and open-domain question answering with large-scale corpus contains up to three trillion tokens. Across multiple benchmarks, these agents outperform published state-of-the-art by 17.3% on average. We attribute this efficacy to two key factors: native tool proficiency, which enables agents to leverage executable code and terminal commands rather than passive semantic queries, and file system familiarity, which allows them to navigate massive text corpora as directory structures. These findings suggest that delegating long-context processing to coding agents offers an effective alternative to semantic search or context window scaling, opening new directions for long-context processing in LLMs.

Method: Uses centroid-based vector adaptation estimated from healthy-control speech to align source-language self-supervised speech representations with target-language distribution

Result: LS substantially improves sensitivity and F1 in cross-lingual settings, with smaller but consistent gains in multilingual settings; reduces language identity in embedding space

Conclusion: The proposed language shift method effectively removes language-dependent structure from speech representations, enabling better cross-lingual dysarthria detection

Abstract: The limited availability of dysarthric speech data makes cross-lingual detection an important but challenging problem. A key difficulty is that speech representations often encode language-dependent structure that can confound dysarthria detection. We propose a representation-level language shift (LS) that aligns source-language self-supervised speech representations with the target-language distribution using centroid-based vector adaptation estimated from healthy-control speech. We evaluate the approach on oral DDK recordings from Parkinson’s disease speech datasets in Czech, German, and Spanish under both cross-lingual and multilingual settings. LS substantially improves sensitivity and F1 in cross-lingual settings, while yielding smaller but consistent gains in multilingual settings. Representation analysis further shows that LS reduces language identity in the embedding space, supporting the interpretation that LS removes language-dependent structure.

Method: Proposes a training-free regeneration paradigm using offline-curated contrastive Reflection Memory (RM) to provide corrective guidance. At inference: RM-guided self-verification followed by single RM-guided regeneration, avoiding both iterative correction and multi-sample selection

Result: Method outperforms prior methods on nine benchmarks spanning algorithmic, reasoning, symbolic, and domain-specific tasks in both small- and large-scale LLMs while maintaining low computational cost

Conclusion: The proposed RM-guided regeneration approach provides an effective balance between accuracy and efficiency for LLM self-improvement, addressing limitations of existing verification methods

Abstract: Verification-guided self-improvement has recently emerged as a promising approach to improving the accuracy of large language model (LLM) outputs. However, existing approaches face a trade-off between inference efficiency and accuracy: iterative verification-rectification is computationally expensive and prone to being trapped in faulty reasoning, while best-of-N selection requires extensive sampling without addressing internal model flaws. We propose a training-free regeneration paradigm that leverages an offline-curated contrastive Reflection Memory (RM) to provide corrective guidance, while regenerating from scratch helps break out of faulty reasoning. At inference time, the method performs RM-guided self-verification followed by a single RM-guided regeneration, avoiding both iterative correction and multi-sample selection. We evaluated our method on nine benchmarks that span algorithmic, reasoning, symbolic, and domain-specific tasks in both small- and large-scale LLMs. Experiment results show that our method outperforms prior methods while maintaining low computational cost.

Method: Created a dataset of peer reviews simulating various levels of human-AI collaboration, then evaluated five state-of-the-art AI-text detectors (including two commercial systems). Investigated whether peer-review-specific signals (access to manuscript, scientific writing domain) could improve detection accuracy.

Result: All detectors misclassified a significant fraction of LLM-polished reviews as AI-generated, risking false misconduct accusations. While incorporating peer-review-specific signals yielded some improvements, no approach met accuracy standards needed for reliable AI-use detection in peer reviews.

Conclusion: Current AI-text detectors are insufficient for enforcing LLM usage policies in peer review, and recent public estimates of AI use based on these detectors should be interpreted cautiously as they likely overstate policy violations.

Abstract: A number of scientific conferences and journals have recently enacted policies that prohibit LLM usage by peer reviewers, except for polishing, paraphrasing, and grammar correction of otherwise human-written reviews. But, are these policies enforceable? To answer this question, we assemble a dataset of peer reviews simulating multiple levels of human-AI collaboration, and evaluate five state-of-the-art detectors, including two commercial systems. Our analysis shows that all detectors misclassify a non-trivial fraction of LLM-polished reviews as AI-generated, thereby risking false accusations of academic misconduct. We further investigate whether peer-review-specific signals, including access to the paper manuscript and the constrained domain of scientific writing, can be leveraged to improve detection. While incorporating such signals yields measurable gains in some settings, we identify limitations in each approach and find that none meets the accuracy standards required for identifying AI use in peer reviews. Importantly, our results suggest that recent public estimates of AI use in peer reviews through the use of AI-text detectors should be interpreted with caution, as current detectors misclassify mixed reviews (collaborative human-AI outputs) as fully AI generated, potentially overstating the extent of policy violations.

Method: 1) Resource-efficient training pipeline starting with LoRA-based continual pre-training of a multilingual encoder on large-scale corpus. 2) Progressive instruction-tuning strategy: sequential adaptation on general and task-specific instruction sets to enable generative capabilities.

Result: Despite its compact size, the model achieves competitive performance compared to existing multi-billion-parameter baselines across comprehensive benchmarks for Turkish language tasks.

Conclusion: Masked diffusion modeling combined with multi-stage tuning is effective for non-autoregressive text generation in morphologically rich languages like Turkish, validating this approach for specialized language applications.

Abstract: Masked Diffusion Language Models (MDLMs) have emerged as a compelling non-autoregressive alternative to standard large language models; however, their application to morphologically rich languages remains limited. In this paper, we introduce $\textit{Diffutron}$, a masked diffusion language model specifically designed for Turkish. Our approach leverages a resource-efficient training pipeline, starting with LoRA-based continual pre-training of a multilingual encoder on a large-scale corpus. To enable generative capabilities, we employ a progressive instruction-tuning strategy, sequentially adapting the model on general and task-specific instruction sets. Experimental results across comprehensive benchmarks demonstrate that, despite its compact size, our model achieves competitive performance compared to existing multi-billion-parameter baselines. These findings validate the effectiveness of masked diffusion modeling combined with multi-stage tuning for non-autoregressive text generation in Turkish.

Method: Created a corpus of 7,394 clinical reports covering 5,009 patient cases using a structured protocol: medical residents across 18 specialties authored realistic but entirely fictitious patient reports following predefined clinical scenarios and templates, with epidemiological guidance from French National Health Data System.

Result: Produced PARHAF corpus with general-purpose component approximating real-world hospitalization distributions and four specialized subsets for oncology, infectious diseases, and diagnostic coding use cases, released under CC-BY open license.

Conclusion: PARHAF provides a valuable privacy-preserving resource for French clinical NLP development and establishes a replicable methodology for creating shareable synthetic clinical corpora in other languages and health systems.

Abstract: The development of clinical natural language processing (NLP) systems is severely hampered by the sensitive nature of medical records, which restricts data sharing under stringent privacy regulations, particularly in France and the broader European Union. To address this gap, we introduce PARHAF, a large open-source corpus of clinical documents in French. PARHAF comprises expert-authored clinical reports describing realistic yet entirely fictitious patient cases, making it anonymous and freely shareable by design. The corpus was developed using a structured protocol that combined clinician expertise with epidemiological guidance from the French National Health Data System (SNDS), ensuring broad clinical coverage. A total of 104 medical residents across 18 specialties authored and peer-reviewed the reports following predefined clinical scenarios and document templates. The corpus contains 7394 clinical reports covering 5009 patient cases across a wide range of medical and surgical specialties. It includes a general-purpose component designed to approximate real-world hospitalization distributions, and four specialized subsets that support information-extraction use cases in oncology, infectious diseases, and diagnostic coding. Documents are released under a CC-BY open license, with a portion temporarily embargoed to enable future benchmarking under controlled conditions. PARHAF provides a valuable resource for training and evaluating French clinical language models in a fully privacy-preserving setting, and establishes a replicable methodology for building shareable synthetic clinical corpora in other languages and health systems.

Method: Constructed question-answer dataset from authoritative sources on COVID-19, dengue, Nipah virus, and Chikungunya. Evaluated four LLMs using semantic similarity, expert-model cross-evaluation, and Natural Language Inference (NLI).

Result: Findings reveal both strengths and limitations of LLMs in representing epidemiological history and health crisis knowledge, with mixed performance across different evaluation metrics.

Conclusion: LLMs show promise but also risks for informing health policy in resource-constrained environments, highlighting need for careful implementation and validation.

Abstract: Large Language Models (LLMs) offer significant potential for delivering health information. However, their reliability in low-resource contexts remains uncertain. This study evaluates GPT-4, Gemini Pro, Llama~3, and Mistral-7B on health crisis-related enquiries concerning COVID-19, dengue, the Nipah virus, and Chikungunya in the low-resource context of Bangladesh. We constructed a question–answer dataset from authoritative sources and assessed model outputs through semantic similarity, expert-model cross-evaluation, and Natural Language Inference (NLI). Findings highlight both the strengths and limitations of LLMs in representing epidemiological history and health crisis knowledge, underscoring their promise and risks for informing policy in resource-constrained environments.

Method: PCFJudge is an inference-time method that reruns the same factuality-first listwise prompt over multiple orderings of the same candidate set and aggregates the resulting scores, ranks, and uncertainty signals into a single consensus decision.

Result: On RewardBench 2 Factuality, PCFJudge improves over direct judging by up to 7 absolute points, with development ablations showing dominant gains from permutation consensus rather than heavier arbitration layers.

Conclusion: A meaningful share of factuality-judging error arises from order instability, and averaging over this nuisance variation is a simple and effective way to make LLM evaluation more reliable.

Abstract: Large language models (LLMs) are now widely used as judges, yet their decisions can change under presentation choices that should be irrelevant. We study one such source of instability: candidate-order sensitivity in listwise factuality evaluation, where several answers can look similarly polished while differing sharply in hallucination risk. We introduce PCFJudge, an inference-time method that reruns the same factuality-first listwise prompt over multiple orderings of the same candidate set and aggregates the resulting scores, ranks, and uncertainty signals into a single consensus decision. On RewardBench 2 Factuality, PCFJudge improves over direct judging by up to 7 absolute points. Development ablations show that the dominant gain comes from permutation consensus itself rather than from heavier arbitration layers. These results suggest that a meaningful share of factuality-judging error arises from order instability, and that averaging over this nuisance variation is a simple and effective way to make LLM evaluation more reliable.

Method: Created JUBAKU benchmark with adversarial construction across ten cultural categories, featuring dialogue scenarios hand-crafted by native Japanese annotators specifically designed to trigger latent social biases in Japanese LLMs.

Result: All nine Japanese LLMs performed poorly on JUBAKU with average accuracy of 23% (range 13-33%), well below the 50% random baseline, despite higher accuracy on translated English benchmarks. Human annotators achieved 91% accuracy confirming benchmark reliability.

Conclusion: Culture-specific bias benchmarks like JUBAKU are essential for properly evaluating social biases in LLMs, as translated English benchmarks fail to capture local cultural nuances and stereotypes.

Abstract: Social biases reflected in language are inherently shaped by cultural norms, which vary significantly across regions and lead to diverse manifestations of stereotypes. Existing evaluations of social bias in large language models (LLMs) for non-English contexts, however, often rely on translations of English benchmarks. Such benchmarks fail to reflect local cultural norms, including those found in Japanese. For instance, Western benchmarks may overlook Japan-specific stereotypes related to hierarchical relationships, regional dialects, or traditional gender roles. To address this limitation, we introduce Japanese cUlture adversarial BiAs benchmarK Under handcrafted creation (JUBAKU), a benchmark tailored to Japanese cultural contexts. JUBAKU uses adversarial construction to expose latent biases across ten distinct cultural categories. Unlike existing benchmarks, JUBAKU features dialogue scenarios hand-crafted by native Japanese annotators, specifically designed to trigger and reveal latent social biases in Japanese LLMs. We evaluated nine Japanese LLMs on JUBAKU and three others adapted from English benchmarks. All models clearly exhibited biases on JUBAKU, performing below the random baseline of 50% with an average accuracy of 23% (ranging from 13% to 33%), despite higher accuracy on the other benchmarks. Human annotators achieved 91% accuracy in identifying unbiased responses, confirming JUBAKU’s reliability and its adversarial nature to LLMs.

Method: Three-stage agentic LLM framework: (1) relevance classification selects price-relevant similar products using descriptions/attributes, (2) relative utility assessment compares target vs. similar products along price-influencing dimensions (brand, size, features), (3) reasoning-based decision aggregates justifications into explainable outlier judgments.

Result: Achieves over 75% agreement with human auditors, outperforms zero-shot and retrieval-based LLM techniques. Ablation studies show sensitivity to key hyperparameters and flexibility for different accuracy requirements and auditor agreements.

Conclusion: The proposed agentic LLM framework effectively detects price outliers through semantic reasoning, providing explainable judgments that outperform traditional methods and offer flexibility for various application scenarios.

Abstract: Detecting product price outliers is important for retail and e-commerce stores as erroneous or unexpectedly high prices adversely affect competitiveness, revenue, and consumer trust. Classical techniques offer simple thresholds while ignoring the rich semantic relationships among product attributes. We propose an agentic Large Language Model (LLM) framework that treats outlier price flagging as a reasoning task grounded in related product detection and comparison. The system processes the prices of target products in three stages: (i) relevance classification selects price-relevant similar products using product descriptions and attributes; (ii) relative utility assessment evaluates the target product against each similar product along price influencing dimensions (e.g., brand, size, features); (iii) reasoning-based decision aggregates these justifications into an explainable price outlier judgment. The framework attains over 75% agreement with human auditors on a test dataset, and outperforms zero-shot and retrieval based LLM techniques. Ablation studies show the sensitivity of the method to key hyper-parameters and testify on its flexibility to be applied to cases with different accuracy requirement and auditor agreements.

Method: Proposes Diversity-Aware Retention (DAR), which at each debate round selects the subset of agent responses that maximally disagree with each other and with the majority vote before broadcasting. Uses explicit index-based retention to preserve original messages without modification.

Result: Experiments on diverse reasoning and question answering benchmarks show that selective message propagation consistently improves debate performance, particularly as the number of agents scales where noise accumulation is most severe.

Conclusion: DAR demonstrates that what agents hear is as important as what agents say in multi-agent reasoning systems, and selective retention of diverse perspectives improves debate outcomes.

Abstract: Multi-Agent Debate has emerged as a promising framework for improving the reasoning quality of large language models through iterative inter-agent communication. However, broadcasting all agent messages at every round introduces noise and redundancy that can degrade debate quality and waste computational resources. Current approaches rely on uncertainty estimation to filter low-confidence responses before broadcasting, but this approach is unreliable due to miscalibrated confidence scores and sensitivity to threshold selection. To address this, we propose Diversity-Aware Retention (DAR), a lightweight debate framework that, at each debate round, selects the subset of agent responses that maximally disagree with each other and with the majority vote before broadcasting. Through an explicit index-based retention mechanism, DAR preserves the original messages without modification, ensuring that retained disagreements remain authentic. Experiments on diverse reasoning and question answering benchmarks demonstrate that our selective message propagation consistently improves debate performance, particularly as the number of agents scales, where noise accumulation is most severe. Our results highlight that what agents hear is as important as what agents say in multi-agent reasoning systems.

Method: End-to-end system combining multi-agent LLM with CNN-based tumor segmentation. Extractor agent identifies clinical variables from unstructured notes, scorer agent applies BT-RADS decision logic integrating extracted variables with volumetric measurements. Evaluated on 509 post-treatment glioma MRI exams.

Result: System achieved 76.0% accuracy vs 57.5% for initial clinical assessments (+18.5 percentage points; P<.001). High sensitivity for context-dependent categories (BT-1b 100%, BT-1a 92.7%, BT-3a 87.5%), moderate sensitivity for threshold-dependent categories. High positive predictive value for BT-4 detection (92.9%).

Conclusion: Multi-agent LLM system achieved higher BT-RADS classification agreement with expert reference standard than initial clinical scoring, with high accuracy for context-dependent scores and high positive predictive value for BT-4 detection.

Abstract: The Brain Tumor Reporting and Data System (BT-RADS) standardizes post-treatment MRI response assessment in patients with diffuse gliomas but requires complex integration of imaging trends, medication effects, and radiation timing. This study evaluates an end-to-end multi-agent large language model (LLM) and convolutional neural network (CNN) system for automated BT-RADS classification. A multi-agent LLM system combined with automated CNN-based tumor segmentation was retrospectively evaluated on 509 consecutive post-treatment glioma MRI examinations from a single high-volume center. An extractor agent identified clinical variables (steroid status, bevacizumab status, radiation date) from unstructured clinical notes, while a scorer agent applied BT-RADS decision logic integrating extracted variables with volumetric measurements. Expert reference standard classifications were established by an independent board-certified neuroradiologist. Of 509 examinations, 492 met inclusion criteria. The system achieved 374/492 (76.0%; 95% CI, 72.1%-79.6%) accuracy versus 283/492 (57.5%; 95% CI, 53.1%-61.8%) for initial clinical assessments (+18.5 percentage points; P<.001). Context-dependent categories showed high sensitivity (BT-1b 100%, BT-1a 92.7%, BT-3a 87.5%), while threshold-dependent categories showed moderate sensitivity (BT-3c 74.8%, BT-2 69.2%, BT-4 69.3%, BT-3b 57.1%). For BT-4, positive predictive value was 92.9%. The multi-agent LLM system achieved higher BT-RADS classification agreement with expert reference standard compared to initial clinical scoring, with high accuracy for context-dependent scores and high positive predictive value for BT-4 detection.

Method: Used four converging psychophysical paradigms: representational similarity analysis, behavioral discrimination tests, precision gradients, and causal interventions across three magnitude domains (numerical, temporal, spatial) in three 7-9B instruction-tuned models (Llama, Mistral, Qwen) spanning different architecture families.

Result: 1) Representational geometry is consistently log-compressive across all models (RSA correlations 0.68-0.96 with Weber-law matrix). 2) Geometry dissociates from behavior - one model shows human-range Weber fraction while others don’t, and all perform poorly on temporal/spatial discrimination despite having logarithmic geometry. 3) Causal intervention shows early layers (4.1x specificity) are functionally important for magnitude processing, while later layers with strongest geometry are not causally engaged (1.2x).

Conclusion: Training data statistics alone produce log-compressive magnitude geometry in LLMs (similar to human Weber-Fechner law), but this geometry doesn’t guarantee behavioral competence in magnitude tasks, revealing a dissociation between representation and function.

Abstract: How do transformer language models represent magnitude? Recent work disagrees: some find logarithmic spacing, others linear encoding, others per-digit circular representations. We apply the formal tools of psychophysics to resolve this. Using four converging paradigms (representational similarity analysis, behavioural discrimination, precision gradients, causal intervention) across three magnitude domains in three 7-9B instruction-tuned models spanning three architecture families (Llama, Mistral, Qwen), we report three findings. First, representational geometry is consistently log-compressive: RSA correlations with a Weber-law dissimilarity matrix ranged from .68 to .96 across all 96 model-domain-layer cells, with linear geometry never preferred. Second, this geometry is dissociated from behaviour: one model produces a human-range Weber fraction (WF = 0.20) while the other does not, and both models perform at chance on temporal and spatial discrimination despite possessing logarithmic geometry. Third, causal intervention reveals a layer dissociation: early layers are functionally implicated in magnitude processing (4.1x specificity) while later layers where geometry is strongest are not causally engaged (1.2x). Corpus analysis confirms the efficient coding precondition (alpha = 0.77). These results suggest that training data statistics alone are sufficient to produce log-compressive magnitude geometry, but geometry alone does not guarantee behavioural competence.

Method: PAVE decomposes retrieved context into question-conditioned atomic facts, drafts an answer, scores how well the draft is supported by extracted premises, and revises low-support outputs before finalization, creating an auditable trace of reasoning.

Result: PAVE outperforms simpler post-retrieval baselines in evidence-grounded QA settings, with gains up to 32.7 accuracy points on a span-grounded benchmark, showing improved evidence-grounded consistency.

Conclusion: Explicit premise extraction plus support-gated revision can strengthen evidence-grounded consistency in retrieval-augmented LLM systems, providing auditable reasoning traces.

Abstract: Retrieval-augmented language models can retrieve relevant evidence yet still commit to answers before explicitly checking whether the retrieved context supports the conclusion. We present PAVE (Premise-Grounded Answer Validation and Editing), an inference-time validation layer for evidence-grounded question answering. PAVE decomposes retrieved context into question-conditioned atomic facts, drafts an answer, scores how well that draft is supported by the extracted premises, and revises low-support outputs before finalization. The resulting trace makes answer commitment auditable at the level of explicit premises, support scores, and revision decisions. In controlled ablations with a fixed retriever and backbone, PAVE outperforms simpler post-retrieval baselines in two evidence-grounded QA settings, with the largest gain reaching 32.7 accuracy points on a span-grounded benchmark. We view these findings as proof-of-concept evidence that explicit premise extraction plus support-gated revision can strengthen evidence-grounded consistency in retrieval-augmented LLM systems.

Method: Focuses on four grammatical phenomena related to pronouns, applies correlation and clustering methods to model covariation and dialectal distribution.

Result: Correlation captures only limited pairwise associations, while clustering reveals speaker groupings reflecting regional dialectal patterns.

Conclusion: Despite methodological constraints from sample size differences between fields, interdisciplinary research is important for developing fair, inclusive language technologies that respect dialectal diversity.

Abstract: This paper investigates morphosyntactic covariation in Brazilian Portuguese (BP) to assess whether dialectal origin can be inferred from the combined behavior of linguistic variables. Focusing on four grammatical phenomena related to pronouns, correlation and clustering methods are applied to model covariation and dialectal distribution. The results indicate that correlation captures only limited pairwise associations, whereas clustering reveals speaker groupings that reflect regional dialectal patterns. Despite the methodological constraints imposed by differences in sample size requirements between sociolinguistics and computational approaches, the study highlights the importance of interdisciplinary research. Developing fair and inclusive language technologies that respect dialectal diversity outweighs the challenges of integrating these fields.

Method: NoT models reasoning as a directed graph with typed nodes and edges, guided by a heuristic-based controller policy. Evaluated across four benchmarks (GSM8K, Game of 24, HotpotQA, ProofWriter) and three models (GPT-4o-mini, Llama-3.3-70B-Instruct, Qwen2.5-72B-Instruct).

Result: NoT surpasses ToT on multi-hop reasoning (91.0% vs. 88.0% on HotpotQA with LLM-as-Judge). With 72B open-source models, NoT achieves highest accuracy on GSM8K (91.5%), and Qwen2.5-72B achieves best multi-hop QA overall (91.7% on HotpotQA). Self-generated controller heuristics outperform fixed/random strategies on logical reasoning.

Conclusion: Network topology reasoning outperforms chain/tree structures for complex multi-hop tasks, LLM-generated heuristics can effectively guide graph-based reasoning, and evaluation methodology significantly impacts method rankings (string-match underestimates open-ended QA performance).

Abstract: Existing prompting paradigms structure LLM reasoning in limited topologies: Chain-of-Thought (CoT) produces linear traces, while Tree-of-Thought (ToT) performs branching search. Yet complex reasoning often requires merging intermediate results, revisiting hypotheses, and integrating evidence from multiple sources. We propose Network-of-Thought (NoT), a framework that models reasoning as a directed graph with typed nodes and edges, guided by a heuristic-based controller policy. Across four benchmarks (GSM8K, Game of 24, HotpotQA, ProofWriter) and three models (GPT-4o-mini, Llama-3.3-70B-Instruct, Qwen2.5-72B-Instruct), we investigate when network topology outperforms chain or tree structures, whether LLM-generated heuristics can guide graph-based reasoning search, and the computation-accuracy tradeoff across topologies, evaluating each method on accuracy, topology simplicity, and token efficiency. Our results show that CoT remains effective for sequential tasks with GPT-4o-mini (89.5% on GSM8K), while NoT surpasses ToT on multi-hop reasoning (91.0% vs.\ 88.0% on HotpotQA with LLM-as-Judge). With 72B open-source models, NoT achieves the highest accuracy on GSM8K (91.5%), and Qwen2.5-72B achieves the best multi-hop QA result overall (91.7% on HotpotQA). Self-generated controller heuristics outperform fixed and random strategies on logical reasoning, with uncertainty-only weighting achieving 57.0% on ProofWriter. We also find that evaluation methodology significantly impacts method rankings: string-match underestimates all methods on open-ended QA, with the largest gap for NoT, a pattern consistent across all three models (14–18 percentage point gap on HotpotQA).

Method: Created MzansiText (curated multilingual pretraining corpus with reproducible filtering pipeline) and trained MzansiLM (125M-parameter decoder-only model). Evaluated using three adaptation regimes: monolingual task-specific finetuning, multilingual task-specific finetuning, and general multi-task instruction finetuning.

Result: Monolingual finetuning achieved strong performance on data-to-text generation (20.65 BLEU on isiXhosa). Multilingual finetuning benefited closely related languages on topic classification (78.5% macro-F1 on isiXhosa news classification). However, few-shot reasoning remained challenging even for larger models.

Conclusion: Decoder-only models can effectively adapt to supervised NLU/NLG tasks for low-resource languages at small scale, but few-shot reasoning requires larger models. The released resources provide reproducible baselines and adaptation guidance for South African languages.

Abstract: Decoder-only language models can be adapted to diverse tasks through instruction finetuning, but the extent to which this generalizes at small scale for low-resource languages remains unclear. We focus on the languages of South Africa, where we are not aware of a publicly available decoder-only model that explicitly targets all eleven official written languages, nine of which are low-resource. We introduce MzansiText, a curated multilingual pretraining corpus with a reproducible filtering pipeline, and MzansiLM, a 125M-parameter language model trained from scratch. We evaluate MzansiLM on natural language understanding and generation using three adaptation regimes: monolingual task-specific finetuning, multilingual task-specific finetuning, and general multi-task instruction finetuning. Monolingual task-specific finetuning achieves strong performance on data-to-text generation, reaching 20.65 BLEU on isiXhosa and competing with encoder-decoder baselines over ten times larger. Multilingual task-specific finetuning benefits closely related languages on topic classification, achieving 78.5% macro-F1 on isiXhosa news classification. While MzansiLM adapts effectively to supervised NLU and NLG tasks, few-shot reasoning remains challenging at this model size, with performance near chance even for much larger decoder-only models. We release MzansiText and MzansiLM to provide a reproducible decoder-only baseline and clear guidance on adaptation strategies for South African languages at small scale.

Method: Proposes Code-MIE framework that: (1) extracts entity attributes from text to guide context understanding, (2) converts images to scene graphs and visual features, (3) uses Python function as input template with entity attributes, scene graphs, and text as parameters, and (4) outputs Python dictionaries containing extraction results.

Result: Achieves state-of-the-art performance: 61.03% and 60.49% on English/Chinese M³D datasets, and 76.04%, 88.07%, 73.94% on Twitter-15, Twitter-17, and MNRE datasets respectively.

Conclusion: Code-MIE effectively formalizes multimodal information extraction as code understanding and generation, outperforming existing methods by better aligning with structured IE tasks through unified code-style templates.

Abstract: With the rapid development of large language models (LLMs), more and more researchers have paid attention to information extraction based on LLMs. However, there are still some spaces to improve in the existing related methods. First, existing multimodal information extraction (MIE) methods usually employ natural language templates as the input and output of LLMs, which mismatch with the characteristics of information tasks that mostly include structured information such as entities and relations. Second, although a few methods have adopted structured and more IE-friendly code-style templates, they just explored their methods on text-only IE rather than multimodal IE. Moreover, their methods are more complex in design, requiring separate templates to be designed for each task. In this paper, we propose a Code-style Multimodal Information Extraction framework (Code-MIE) which formalizes MIE as unified code understanding and generation. Code-MIE has the following novel designs: (1) Entity attributes such as gender, affiliation are extracted from the text to guide the model to understand the context and role of entities. (2) Images are converted into scene graphs and visual features to incorporate rich visual information into the model. (3) The input template is constructed as a Python function, where entity attributes, scene graphs and raw text compose of the function parameters. In contrast, the output template is formalized as Python dictionaries containing all extraction results such as entities, relations, etc. To evaluate Code-MIE, we conducted extensive experiments on the M$^3$D, Twitter-15, Twitter-17, and MNRE datasets. The results show that our method achieves state-of-the-art performance compared to six competing baseline models, with 61.03% and 60.49% on the English and Chinese datasets of M$^3$D, and 76.04%, 88.07%, and 73.94% on the other three datasets.

Method: The paper uses neuron-level knowledge attribution (NLKA) to contrast successful and failed edits, identifying consistent patterns in attention and FFN modules. Based on these findings, they propose MEGA (MEchanism-Guided Activation steering), which performs attention-residual interventions in attribution-aligned regions without modifying model weights.

Result: Across representative KE methods, they find consistent patterns: mid-to-late attention promotes new targets while attention and FFN modules cooperate to suppress original facts. MEGA achieves strong editing performance on CounterFact and Popular datasets for GPT2-XL and LLaMA2-7B across KE metrics.

Conclusion: Post-edit attribution can be elevated from analysis to engineering signal, pinpointing where and how edits take hold to enable reliable, architecture-agnostic knowledge editing through targeted activation steering without weight modification.

Abstract: Large language models (LLMs) are increasingly used as knowledge bases, but keeping them up to date requires targeted knowledge editing (KE). However, it remains unclear how edits are implemented inside the model once applied. In this work, we take a mechanistic view of KE using neuron-level knowledge attribution (NLKA). Unlike prior work that focuses on pre-edit causal tracing and localization, we use post-edit attribution – contrasting successful and failed edits – to isolate the computations that shift when an edit succeeds. Across representative KE methods, we find a consistent pattern: mid-to-late attention predominantly promotes the new target, while attention and FFN modules cooperate to suppress the original fact. Motivated by these findings, we propose MEGA, a MEchanism-Guided Activation steering method that performs attention-residual interventions in attribution-aligned regions without modifying model weights. On CounterFact and Popular, MEGA achieves strong editing performance across KE metrics on GPT2-XL and LLaMA2-7B. Overall, our results elevate post-edit attribution from analysis to engineering signal: by pinpointing where and how edits take hold, it powers MEGA to deliver reliable, architecture-agnostic knowledge edits.

Method: Proposes Cross-Generation evaluation to measure intermediate reasoning quality, then introduces Separated Thinking And Response Training (START) which first trains only the thinking process using rewards defined on final answers to avoid shortcuts.

Result: RLVR’s efficacy on GQA is much lower than on verifiable tasks; direct RL training on GQA is less effective than RLVR; START improves both thinking quality and final answers across multiple GQA benchmarks and RL algorithms.

Conclusion: Explicit training on GQA remains necessary beyond verifiable tasks; START effectively avoids shortcuts in GQA by separating thinking training from response generation, leading to improved reasoning and answer quality.

Abstract: Reinforcement learning from verifiable rewards (RLVR) stimulates the thinking processes of large language models (LLMs), substantially enhancing their reasoning abilities on verifiable tasks. It is often assumed that similar gains should transfer to general question answering (GQA), but this assumption has not been thoroughly validated. To assess whether RLVR automatically improves LLM performance on GQA, we propose a Cross-Generation evaluation framework that measures the quality of intermediate reasoning by feeding the generated thinking context into LLMs of varying capabilities. Our evaluation leads to a discouraging finding: the efficacy of the thinking process on GQA tasks is markedly lower than on verifiable tasks, suggesting that explicit training on GQA remains necessary in addition to training on verifiable tasks. We further observe that direct RL training on GQA is less effective than RLVR. Our hypothesis is that, whereas verifiable tasks demand robust logical chains to obtain high rewards, GQA tasks often admit shortcuts to high rewards without cultivating high-quality thinking. To avoid possible shortcuts, we introduce a simple method, Separated Thinking And Response Training (START), which first trains only the thinking process, using rewards defined on the final answer. We show that START improves both the quality of thinking and the final answer across several GQA benchmarks and RL algorithms.

Method: Three-stage pipeline: (1) extract structured domain cards from seed benchmarks, (2) prompt multiple designer LLMs to generate quota-controlled test suites, (3) validate items using multi-model answerer panels with exact/numeric/symbolic verifiers or rubric-guided judging, producing designer-answerer matrices with quality flags.

Result: Generated 16.7K items across 9 variants (computer science, mathematics, medicine, theory-of-mind reasoning, including multilingual and multimodal settings), retained ~15K core items after filtering, produced ~152K graded model-item responses. Found benchmark-design ability only moderately correlated with answer-time performance (Spearman rho 0.37), invalidity negatively associated with discrimination (Pearson r0.62).

Conclusion: BenchBench enables scalable evaluation of LLMs’ benchmark-design capabilities, revealing that design ability is distinct from answer-time performance. The framework supports audits of format/modality/language fidelity and suite-dependent interactions, providing a new dimension for assessing LLM capabilities.

Abstract: Benchmarks are the de facto standard for tracking progress in large language models (LLMs), yet static test sets can rapidly saturate, become vulnerable to contamination, and are costly to refresh. Scalable evaluation of open-ended items often relies on LLM judges, introducing additional sources of bias and prompt sensitivity. We argue that evaluation must extend beyond how well models answer benchmarks to how well models design them. We introduce BenchBench, a three-stage pipeline and dataset for benchmarking automated benchmark generation: (i) extract structured domain cards from seed benchmarks, (ii) prompt multiple designer LLMs to generate quota-controlled suites, and (iii) validate items with a multi-model answerer panel using exact/numeric/symbolic verifiers when possible and rubric-guided judging otherwise, yielding designer–answerer matrices with item-level quality flags and psychometric diagnostics. Across nine variants spanning computer science, mathematics, medicine, and theory-of-mind reasoning (including multilingual and multimodal settings), we generate 16.7K items, retain ~15K core items post-filtering, and produce ~152K graded model–item responses. BenchBench shows that benchmark-design ability is only moderately correlated with answer-time strength (Spearman rho 0.37), invalidity is negatively associated with discrimination (Pearson r0.62), and the resulting designer–answerer matrices enable scalable audits of format/modality/language fidelity and suite-dependent self/family interactions. The project is available at: https://github.com/koanatakiyo/BenchBench.

Method: Proposes HiCI (Hierarchical Construction-Integration), a hierarchical attention module that: 1) constructs segment-level representations, 2) integrates them into a shared global context, and 3) broadcasts both to condition segment-level attention. Implemented through parameter-efficient adaptation of LLaMA-2 with only <5.5% additional parameters.

Result: Successfully extends context from 4K to 100K tokens (7B model) and 64K tokens (13B model). Shows consistent improvements across language modeling, retrieval, and instruction-following benchmarks. Matches proprietary models on topic retrieval and surpasses GPT-3.5-Turbo-16K on code comprehension.

Conclusion: Explicit hierarchical structuring serves as an effective inductive bias for long-context modeling, demonstrating that cognitive-inspired approaches can enhance language model performance on extended contexts with minimal parameter overhead.

Abstract: Long-context language modeling is commonly framed as a scalability challenge of token-level attention, yet local-to-global information structuring remains largely implicit in existing approaches. Drawing on cognitive theories of discourse comprehension, we propose HiCI (Hierarchical Construction–Integration), a hierarchical attention module that constructs segment-level representations, integrates them into a shared global context, and broadcasts both to condition segment-level attention. We validate HiCI through parameter-efficient adaptation of LLaMA-2 with only <5.5% additional parameters, extending context from 4K to 100K tokens (7B) and 64K tokens (13B). Across language modeling, retrieval, and instruction-following benchmarks, HiCI yields consistent improvements over strong baselines, including matching proprietary models on topic retrieval and surpassing GPT-3.5-Turbo-16K on code comprehension. These results demonstrate the effectiveness of explicit hierarchical structuring as an inductive bias for long-context modeling.

Method: Developed a comprehensive evaluation framework in collaboration with professional poets to assess ChatGPT’s interpretation of modern Chinese poems across multiple dimensions, comparing its interpretations with original poets’ intents.

Result: ChatGPT’s interpretations aligned with original poets’ intents in over 73% of cases, but showed less satisfactory performance in capturing poeticity and certain other dimensions of poetic understanding.

Conclusion: The study establishes an effective framework for evaluating LLM poetry understanding, revealing ChatGPT’s strengths in interpretation alignment but limitations in capturing poetic essence, providing foundation for future poetry-related LLM research.

Abstract: ChatGPT has demonstrated remarkable capabilities on both poetry generation and translation, yet its ability to truly understand poetry remains unexplored. Previous poetry-related work merely analyzed experimental outcomes without addressing fundamental issues of comprehension. This paper introduces a comprehensive framework for evaluating ChatGPT’s understanding of modern poetry. We collaborated with professional poets to evaluate ChatGPT’s interpretation of modern Chinese poems by different poets along multiple dimensions. Evaluation results show that ChatGPT’s interpretations align with the original poets’ intents in over 73% of the cases. However, its understanding in certain dimensions, particularly in capturing poeticity, proved to be less satisfactory. These findings highlight the effectiveness and necessity of our proposed framework. This study not only evaluates ChatGPT’s ability to understand modern poetry but also establishes a solid foundation for future research on LLMs and their application to poetry-related tasks.

Method: Train Llama-architecture models from scratch on 9B Kazakh tokens with dedicated 50K BPE tokenizer, evaluate on Kazakh cultural QA, reading comprehension, and topic classification benchmarks.

Result: 600M model achieves 30.3% on Kazakh cultural QA (approaching Llama-3.2-1B’s 32.0%) and 25.5% on topic classification, surpassing multilingual models up to 2B parameters. Shows consistent scaling from 50M to 600M.

Conclusion: Small dedicated models with language-appropriate tokenizers offer viable path for low-resource language technology, achieving competitive performance at fraction of computational cost.

Abstract: Kazakh, a Turkic language spoken by over 22 million people, remains underserved by existing multilingual language models, which allocate minimal capacity to low-resource languages and employ tokenizers ill-suited to agglutinative morphology. We present SozKZ, a family of Llama-architecture language models (50M-600M parameters) trained entirely from scratch on 9 billion tokens of Kazakh text with a dedicated 50K BPE tokenizer. We evaluate all models on three Kazakh benchmarks – multiple-choice cultural QA, reading comprehension (Belebele), and topic classification (SIB-200) – alongside five multilingual baselines ranging from 500M to 3B parameters. Our 600M model achieves 30.3% accuracy on Kazakh cultural QA, approaching the 32.0% of Llama-3.2-1B (2x larger), and 25.5% on SIB-200 topic classification, surpassing all evaluated multilingual models up to 2B parameters. We observe consistent scaling from 50M to 600M, with MC QA accuracy rising from 22.8% to 30.3%, suggesting that further scaling remains beneficial. These results demonstrate that small, dedicated models trained from scratch with a language-appropriate tokenizer offer a viable path for low-resource language technology, achieving competitive performance at a fraction of the computational cost. All models and the tokenizer are released under open licenses.

Method: Introduces a multi-agent system that decomposes manuscripts into discrete novelty points for fine-grained retrieval and comparison, builds a comprehensive related-paper database, and cross-references claims to ensure faithfulness. Also proposes a checklist-based evaluation framework for open-ended generation tasks.

Result: Extensive experiments show NoveltyAgent achieves state-of-the-art performance, outperforming GPT-5 DeepResearch by 10.15%.

Conclusion: NoveltyAgent provides reliable, high-quality novelty analysis to help researchers quickly identify novel papers, with code and demo available.

Abstract: The exponential growth of academic publications has led to a surge in papers of varying quality, increasing the cost of paper screening. Current approaches either use novelty assessment within general AI Reviewers or repurpose DeepResearch, which lacks domain-specific mechanisms and thus delivers lower-quality results. To bridge this gap, we introduce NoveltyAgent, a multi-agent system designed to generate comprehensive and faithful novelty reports, enabling thorough evaluation of a paper’s originality. It decomposes manuscripts into discrete novelty points for fine-grained retrieval and comparison, and builds a comprehensive related-paper database while cross-referencing claims to ensure faithfulness. Furthermore, to address the challenge of evaluating such open-ended generation tasks, we propose a checklist-based evaluation framework, providing an unbiased paradigm for building reliable evaluations. Extensive experiments show that NoveltyAgent achieves state-of-the-art performance, outperforming GPT-5 DeepResearch by 10.15%. We hope this system will provide reliable, high-quality novelty analysis and help researchers quickly identify novel papers. Code and demo are available at https://github.com/SStan1/NoveltyAgent.

Method: Encoder-Target Decoupling separates the model providing predictive signals (Encoder) from the model being estimated (Target), enabling optimized heterogeneous pairing. Uses Fisher Separability (J) and Effective Dimensionality (d_eff) as mathematical probes to isolate optimal layer-wise signals from internal prefill activations. Implements SharedTrunkNet architecture based on these insights.

Result: SharedTrunkNet captures up to 45.58% of the accuracy gap between the strongest standalone model and the Oracle router while achieving 74.31% cost savings relative to the highest-cost model.

Conclusion: Internal prefill activations provide robust signals for LLM routing, and mathematical probes can effectively isolate optimal layer-wise information. Encoder-Target Decoupling enables efficient heterogeneous model pairing for improved routing performance.

Abstract: LLMs often share comparable benchmark accuracies, but their complementary performance across task subsets suggests that an Oracle router–a theoretical selector with perfect foresight–can significantly surpass standalone model accuracy by navigating model-specific strengths. While current routers rely on fragile semantic signals, we propose using internal prefill activations via Encoder-Target Decoupling–a functional separation between the model providing the predictive signal (the Encoder) and the model whose performance is being estimated (the Target). This allows optimized heterogeneous pairing between unique encoders and target models. We utilize Fisher Separability (J) and Effective Dimensionality (d_eff) as mathematical probes to isolate optimal layer-wise signals, providing the predictive foundation for our SharedTrunkNet architecture. SharedTrunkNet captures up to 45.58% of the accuracy gap between the strongest standalone model and the Oracle while achieving 74.31% cost savings relative to the highest-cost model.

Method: Proposes Shortcut-Aware Reasoning Training (SART) with ShortcutScore to detect shortcut-promoting samples via gradient misalignment with validation objectives and answer-token concentration, then applies gradient surgery to modify training dynamics.

Result: Achieves +16.5% accuracy and +40.2% robustness over strongest baselines on controlled reasoning benchmarks, significantly improving generalization under distribution shifts.

Conclusion: SART effectively mitigates shortcut learning in LLMs, enhancing genuine reasoning capabilities and generalization performance.

Abstract: Large language models exhibit strong reasoning capabilities, yet often rely on shortcuts such as surface pattern matching and answer memorization rather than genuine logical inference. We propose Shortcut-Aware Reasoning Training (SART), a gradient-aware framework that detects and mitigates shortcut-promoting samples via ShortcutScore and gradient surgery. Our method identifies shortcut signals through gradient misalignment with validation objectives and answer-token concentration, and modifies training dynamics accordingly. Experiments on controlled reasoning benchmarks show that SART achieves +16.5% accuracy and +40.2% robustness over the strongest baseline, significantly improving generalization under distribution shifts. Code is available at: https://github.com/fuyanjie/short-cut-aware-data-centric-reasoning.

Method: Introduces PUPPET taxonomy for personalized emotional manipulation centered on incentive morality. Conducts human study with 1,035 participants across realistic everyday queries, varying personalization and incentive direction (harmful vs prosocial). Benchmarks LLMs on belief prediction task.

Result: Harmful hidden incentives produce significantly larger belief shifts than prosocial ones. LLMs show moderate predictive ability for belief change (r=0.3-0.5) but systematically underestimate the magnitude of belief shift.

Conclusion: Establishes a theoretically grounded and behaviorally validated foundation for studying and combating incentive-driven manipulation in LLMs during everyday user queries, highlighting the importance of considering incentive morality in manipulation research.

Abstract: As users increasingly turn to LLMs for practical and personal advice, they become vulnerable to being subtly steered toward hidden incentives misaligned with their own interests. Prior works have benchmarked persuasion and manipulation detection, but these efforts rely on simulated or debate-style settings, remain uncorrelated with real human belief shifts, and overlook a critical dimension: the morality of hidden incentives driving the manipulation. We introduce PUPPET, a theoretical taxonomy of personalized emotional manipulation in LLM-human dialogues that centers around incentive morality, and conduct a human study with N=1,035 participants across realistic everyday queries, varying personalization and incentive direction (harmful versus prosocial). We find that harmful hidden incentives produce significantly larger belief shifts than prosocial ones. Finally, we benchmark LLMs on the task of belief prediction, finding that models exhibit moderate predictive ability of belief change based on conversational contexts (r=0.3 - 0.5), but they also systematically underestimate the magnitude of belief shift. Together, this work establishes a theoretically grounded and behaviorally validated foundation for studying, and ultimately combatting, incentive-driven manipulation in LLMs during everyday, practical user queries.

Method: Vector-Adapted Retrieval Scoring (VARS) represents users with long-term and short-term vectors in a shared preference space, using these to bias retrieval scoring over structured preference memory. Vectors are updated online from weak scalar rewards from user feedback, enabling personalization with frozen backbones.

Result: On MultiSessionCollab benchmark (math and code tasks), VARS achieves strongest overall performance, matches Reflection baseline in task success, reduces timeout rate and user effort. Learned vectors show interpretability: long-term vectors align with cross-user preference overlap, short-term vectors capture session-specific adaptation.

Conclusion: VARS enables effective personalization of LLM assistants through preference-aware retrieval scoring without fine-tuning, improving interaction efficiency rather than raw task accuracy. The dual-vector design is interpretable and effective for multi-session collaboration.

Abstract: Large language models are increasingly used as personal assistants, yet most lack a persistent user model, forcing users to repeatedly restate preferences across sessions. We propose Vector-Adapted Retrieval Scoring (VARS), a pipeline-agnostic, frozen-backbone framework that represents each user with long-term and short-term vectors in a shared preference space and uses these vectors to bias retrieval scoring over structured preference memory. The vectors are updated online from weak scalar rewards from users’ feedback, enabling personalization without per-user fine-tuning. We evaluate on \textsc{MultiSessionCollab}, an online multi-session collaboration benchmark with rich user preference profiles, across math and code tasks. Under frozen backbones, the main benefit of user-aware retrieval is improved interaction efficiency rather than large gains in raw task accuracy: our full VARS agent achieves the strongest overall performance, matches a strong Reflection baseline in task success, and reduces timeout rate and user effort. The learned long-term vectors also align with cross-user preference overlap, while short-term vectors capture session-specific adaptation, supporting the interpretability of the dual-vector design. Code, model, and data are available at https://github.com/YurenHao0426/VARS.

Method: Finetuned models (GPT-4o, Gemini-2.5-Pro, DeepSeek-V3.1) to expand plot summaries into full text, a task suited for commercial writing assistants, then prompted them with semantic descriptions to test verbatim reproduction of copyrighted books.

Result: Models reproduced up to 85-90% of held-out copyrighted books with single verbatim spans exceeding 460 words. Finetuning on one author’s works unlocked recall of books from over 30 unrelated authors. Three different models memorized the same books in the same regions (r ≥ 0.90).

Conclusion: Model weights store copies of copyrighted works, and finetuning bypasses safety protections, undermining legal defenses based on adequate prevention measures. This represents an industry-wide vulnerability.

Abstract: Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami’s novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors. The effect is not specific to any training author or corpus: random author pairs and public-domain finetuning data produce comparable extraction, while finetuning on synthetic text yields near-zero extraction, indicating that finetuning on individual authors’ works reactivates latent memorization from pretraining. Three models from different providers memorize the same books in the same regions ($r \ge 0.90$), pointing to an industry-wide vulnerability. Our findings offer compelling evidence that model weights store copies of copyrighted works and that the security failures that manifest after finetuning on individual authors’ works undermine a key premise of recent fair use rulings, where courts have conditioned favorable outcomes on the adequacy of measures preventing reproduction of protected expression.

Method: Three methods: DiscoUQ-LLM (logistic regression on LLM-extracted linguistic features like evidence overlap and argument strength), DiscoUQ-Embed (logistic regression on embedding geometry features like cluster distances), and DiscoUQ-Learn (neural network combining all features).

Result: DiscoUQ-LLM achieves average AUROC of 0.802, outperforming best baseline (0.791) with substantially better calibration (ECE 0.036 vs. 0.098). Features generalize across benchmarks with near-zero performance degradation, especially helping in ambiguous “weak disagreement” cases.

Conclusion: Leveraging semantic disagreement structure in multi-agent LLM systems provides better uncertainty quantification than simple voting statistics, with particularly strong benefits in ambiguous cases where traditional methods fail.

Abstract: Multi-agent LLM systems, where multiple prompted instances of a language model independently answer questions, are increasingly used for complex reasoning tasks. However, existing methods for quantifying the uncertainty of their collective outputs rely on shallow voting statistics that discard the rich semantic information in agents’ reasoning. We introduce DiscoUQ, a framework that extracts and leverages the structure of inter-agent disagreement – both linguistic properties (evidence overlap, argument strength, divergence depth) and embedding geometry (cluster distances, dispersion, cohesion) – to produce well-calibrated confidence estimates. We propose three methods of increasing complexity: DiscoUQ-LLM (logistic regression on LLM-extracted structure features), DiscoUQ-Embed (logistic regression on embedding geometry), and DiscoUQ-Learn (a neural network combining all features). Evaluated on four diverse benchmarks (StrategyQA, MMLU, TruthfulQA, ARC-Challenge) with a 5-agent system using Qwen3.5-27B, DiscoUQ-LLM achieves an average AUROC of 0.802, outperforming the best baseline (LLM Aggregator, 0.791) while being substantially better calibrated (ECE 0.036 vs. 0.098). The learned features generalize across benchmarks with near-zero performance degradation and provide the largest improvements where they are most needed: in the ambiguous “weak disagreement” tier where simple vote counting fails.

Method: Proposes Permutation-Aware Group Relative Policy Optimization (PA-GRPO) which constructs permutation groups for each instance and uses two mechanisms: (1) cross-permutation advantage (relative to mean reward over all permutations), and (2) consistency-aware reward (encouraging consistent decisions across permutations).

Result: Outperforms strong baselines across seven benchmarks, substantially reducing selection bias while maintaining high overall performance.

Conclusion: PA-GRPO effectively mitigates selection bias in LLMs for evaluation tasks through permutation-consistent semantic reasoning optimization.

Abstract: Large language models (LLMs) used for multiple-choice and pairwise evaluation tasks often exhibit selection bias due to non-semantic factors like option positions and label symbols. Existing inference-time debiasing is costly and may harm reasoning, while pointwise training ignores that the same question should yield consistent answers across permutations. To address this issue, we propose Permutation-Aware Group Relative Policy Optimization (PA-GRPO), which mitigates selection bias by enforcing permutation-consistent semantic reasoning. PA-GRPO constructs a permutation group for each instance by generating multiple candidate permutations, and optimizes the model using two complementary mechanisms: (1) cross-permutation advantage, which computes advantages relative to the mean reward over all permutations of the same instance, and (2) consistency-aware reward, which encourages the model to produce consistent decisions across different permutations. Experimental results demonstrate that PA-GRPO outperforms strong baselines across seven benchmarks, substantially reducing selection bias while maintaining high overall performance. The code will be made available on Github (https://github.com/ECNU-Text-Computing/PA-GRPO).

Method: Benchmarked eight LLMs across five experimental conditions in English, Kazakh, and Mongolian using 50 hand-crafted questions spanning factual, reasoning, technical, and cultural categories. Evaluated 2,000 responses on accuracy, fluency, and completeness. Tested cross-lingual transfer prompting (reason in English then translate back).

Result: Found consistent 13.8-16.7 percentage point performance gap between English and low-resource language conditions. Models maintained surface-level fluency but produced significantly less accurate content. Cross-lingual transfer yielded selective gains (+2.2pp to +4.3pp) for bilingual architectures but no benefit for English-dominant models.

Conclusion: Current LLMs systematically underserve low-resource language communities, and effective mitigation strategies are architecture-dependent rather than universal, requiring tailored approaches for different model architectures.

Abstract: We investigate how large language models perform on low-resource languages by benchmarking eight LLMs across five experimental conditions in English, Kazakh, and Mongolian. Using 50 hand-crafted questions spanning factual, reasoning, technical, and culturally grounded categories, we evaluate 2,000 responses on accuracy, fluency, and completeness. We find a consistent performance gap of 13.8-16.7 percentage points between English and low-resource language conditions, with models maintaining surface-level fluency while producing significantly less accurate content. Cross-lingual transfer-prompting models to reason in English before translating back-yields selective gains for bilingual architectures (+2.2pp to +4.3pp) but provides no benefit to English-dominant models. Our results demonstrate that current LLMs systematically underserve low-resource language communities, and that effective mitigation strategies are architecture-dependent rather than universal.

Method: Three complementary studies: 1) Develops unified taxonomy of eNVCs grounded in nonverbal communication theory and creates Python toolkit for automated detection; 2) Within-subject survey experiment testing causal effects of eNVCs on emotional decoding; 3) Focus group discussions exploring user interpretive strategies for digital prosody.

Result: eNVCs substantially improve emotional decoding accuracy and lower perceived ambiguity, though benefits weaken with sarcasm. Users employ specific interpretive strategies including drawing meaning from cue absence and defaulting to negative interpretations in ambiguous contexts. Provides practical toolkit for detection.

Conclusion: Establishes eNVCs as coherent, measurable digital behaviors; refines theories of cue richness and interpretive effort; provides tools for affective computing, user modeling, and emotion-aware interface design.

Abstract: As text-based computer-mediated communication (CMC) increasingly structures everyday interaction, a central question re-emerges with new urgency: How do users reconstruct nonverbal expression in environments where embodied cues are absent? This paper provides a systematic, theory-driven account of electronic nonverbal cues (eNVCs) - textual analogues of kinesics, vocalics, and paralinguistics - in public microblog communication. Across three complementary studies, we advance conceptual, empirical, and methodological contributions. Study 1 develops a unified taxonomy of eNVCs grounded in foundational nonverbal communication theory and introduces a scalable Python toolkit for their automated detection. Study 2, a within-subject survey experiment, offers controlled causal evidence that eNVCs substantially improve emotional decoding accuracy and lower perceived ambiguity, while also identifying boundary conditions, such as sarcasm, under which these benefits weaken or disappear. Study 3, through focus group discussions, reveals the interpretive strategies users employ when reasoning about digital prosody, including drawing meaning from the absence of expected cues and defaulting toward negative interpretations in ambiguous contexts. Together, these studies establish eNVCs as a coherent and measurable class of digital behaviors, refine theoretical accounts of cue richness and interpretive effort, and provide practical tools for affective computing, user modeling, and emotion-aware interface design. The eNVC detection toolkit is available as a Python and R package at https://github.com/kokiljaidka/envc.

Method: Proposes ViCLSR, a supervised contrastive learning framework specifically designed for Vietnamese sentence embeddings. Leverages existing NLI datasets and develops a process to adapt existing Vietnamese datasets for supervised contrastive learning compatibility.

Result: ViCLSR significantly outperforms PhoBERT on five benchmark NLU datasets: ViNLI (+6.97% F1), ViWikiFC (+4.97% F1), ViFactCheck (+9.02% F1), UIT-ViCTSD (+5.36% F1), and ViMMRC2.0 (+4.33% Accuracy).

Conclusion: Supervised contrastive learning can effectively address resource limitations in Vietnamese NLU tasks and improve sentence representation learning for low-resource languages. ViCLSR is released for research purposes.

Abstract: High-quality text representations are crucial for natural language understanding (NLU), but low-resource languages like Vietnamese face challenges due to limited annotated data. While pre-trained models like PhoBERT and CafeBERT perform well, their effectiveness is constrained by data scarcity. Contrastive learning (CL) has recently emerged as a promising approach for improving sentence representations, enabling models to effectively distinguish between semantically similar and dissimilar sentences. We propose ViCLSR (Vietnamese Contrastive Learning for Sentence Representations), a novel supervised contrastive learning framework specifically designed to optimize sentence embeddings for Vietnamese, leveraging existing natural language inference (NLI) datasets. Additionally, we propose a process to adapt existing Vietnamese datasets for supervised learning, ensuring compatibility with CL methods. Our experiments demonstrate that ViCLSR significantly outperforms the powerful monolingual pre-trained model PhoBERT on five benchmark NLU datasets such as ViNLI (+6.97% F1), ViWikiFC (+4.97% F1), ViFactCheck (+9.02% F1), UIT-ViCTSD (+5.36% F1), and ViMMRC2.0 (+4.33% Accuracy). ViCLSR shows that supervised contrastive learning can effectively address resource limitations in Vietnamese NLU tasks and improve sentence representation learning for low-resource languages. Furthermore, we conduct an in-depth analysis of the experimental results to uncover the factors contributing to the superior performance of contrastive learning models. ViCLSR is released for research purposes in advancing natural language processing tasks.

Method: Introduces ReasonAlign, a reasoning-based annotation scaffold that exposes LLM-generated explanations while withholding predicted labels. Uses a two-pass protocol inspired by Delphi-style revision: annotators first label instances independently, then revise decisions after viewing model-generated reasoning. Evaluates on sentiment classification and opinion detection tasks, analyzing changes in inter-annotator agreement and revision behavior. Introduces Annotator Effort Proxy (AEP) metric to capture proportion of labels revised after exposure to reasoning.

Result: Exposure to reasoning is associated with increased agreement alongside minimal revision, suggesting that reasoning primarily helps resolve ambiguous cases without inducing widespread changes. The findings provide insight into how reasoning explanations shape annotation consistency.

Conclusion: Reasoning-based scaffolds serve as a practical mechanism for supporting human-AI annotation workflows by improving annotation consistency through exposure to LLM-generated reasoning, particularly for ambiguous cases.

Abstract: Human annotation is central to NLP evaluation, yet subjective tasks often exhibit substantial variability across annotators. While large language models (LLMs) can provide structured reasoning to support annotation, their influence on human annotation behavior remains unclear. We introduce ReasonAlign, a reasoning-based annotation scaffold that exposes LLM-generated explanations while withholding predicted labels. We frame this as a controlled study of how reasoning affects human annotation behavior, rather than a full evaluation of annotation accuracy. Using a two-pass protocol inspired by Delphi-style revision, annotators first label instances independently and then revise their decisions after viewing model-generated reasoning. We evaluate the approach on sentiment classification and opinion detection tasks, analyzing changes in inter-annotator agreement and revision behavior. To quantify these effects, we introduce the Annotator Effort Proxy (AEP), a metric capturing the proportion of labels revised after exposure to reasoning. Our results show that exposure to reasoning is associated with increased agreement alongside minimal revision, suggesting that reasoning primarily helps resolve ambiguous cases without inducing widespread changes. These findings provide insight into how reasoning explanations shape annotation consistency and highlight reasoning-based scaffolds as a practical mechanism for supporting human-AI annotation workflows.

Method: Created BanglaVerse with 1,152 manually curated images across nine cultural domains, supporting visual question answering and captioning. Expanded into four languages and five Bangla dialects, yielding ~32.3K artifacts for comprehensive evaluation.

Result: Evaluating only standard Bangla overestimates model capability - performance drops under dialectal variation, especially for caption generation. Hindi and Urdu retain some cultural meaning but are weaker for structured reasoning. Main bottleneck is missing cultural knowledge rather than visual grounding alone.

Conclusion: BanglaVerse provides a more realistic test bed for measuring culturally grounded multimodal understanding under linguistic variation, highlighting the importance of cultural knowledge in vision-language models.

Abstract: Bangla culture is richly expressed through region, dialect, history, food, politics, media, and everyday visual life, yet it remains underrepresented in multimodal evaluation. To address this gap, we introduce BanglaVerse, a culturally grounded benchmark for evaluating multilingual vision-language models (VLMs) on Bengali culture across historically linked languages and regional dialects. Built from 1,152 manually curated images across nine domains, the benchmark supports visual question answering and captioning, and is expanded into four languages and five Bangla dialects, yielding ~32.3K artifacts. Our experiments show that evaluating only standard Bangla overestimates true model capability: performance drops under dialectal variation, especially for caption generation, while historically linked languages such as Hindi and Urdu retain some cultural meaning but remain weaker for structured reasoning. Across domains, the main bottleneck is missing cultural knowledge rather than visual grounding alone, with knowledge-intensive categories. These findings position BanglaVerse as a more realistic test bed for measuring culturally grounded multimodal understanding under linguistic variation.

Method: Identifies failure mode of entropy-based uncertainty methods, then combines entropy scores with correctness probe signals to create more reliable abstention behavior.

Result: Across three QA benchmarks (TriviaQA, BioASQ, MedicalQA) and four model families, the combined score improves both risk-coverage trade-off and calibration performance relative to entropy-only baselines.

Conclusion: Deployment-facing evaluation of uncertainty methods is crucial, using metrics that reflect whether systems can be trusted to operate at stated risk levels.

Abstract: Selective prediction systems can mitigate harms resulting from language model hallucinations by abstaining from answering in high-risk cases. Uncertainty quantification techniques are often employed to identify such cases, but are rarely evaluated in the context of the wider selective prediction policy and its ability to operate at low target error rates. We identify a model-dependent failure mode of entropy-based uncertainty methods that leads to unreliable abstention behaviour, and address it by combining entropy scores with a correctness probe signal. We find that across three QA benchmarks (TriviaQA, BioASQ, MedicalQA) and four model families, the combined score generally improves both the risk–coverage trade-off and calibration performance relative to entropy-only baselines. Our results highlight the importance of deployment-facing evaluation of uncertainty methods, using metrics that directly reflect whether a system can be trusted to operate at a stated risk level.

Method: Introduces DSD task and releases Span Similarity Dataset (SSD) created via semi-automated pipeline combining LLMs with human verification. Proposes unsupervised baselines (LIME, SHAP, LLMs, custom method) and supervised approach. Evaluates on paraphrase detection task.

Result: LLMs and supervised models achieve highest performance but overall results remain low, highlighting task complexity. DSD shows potential to improve performance in paraphrase detection through additional experiment.

Conclusion: DSD addresses interpretability limitations in STS by identifying specific dissimilar spans. While challenging, it offers benefits for understanding semantic differences and improving downstream NLP applications.

Abstract: Semantic Textual Similarity (STS) is a crucial component of many Natural Language Processing (NLP) applications. However, existing approaches typically reduce semantic nuances to a single score, limiting interpretability. To address this, we introduce the task of Dissimilar Span Detection (DSD), which aims to identify semantically differing spans between pairs of texts. This can help users understand which particular words or tokens negatively affect the similarity score, or be used to improve performance in STS-dependent downstream tasks. Furthermore, we release a new dataset suitable for the task, the Span Similarity Dataset (SSD), developed through a semi-automated pipeline combining large language models (LLMs) with human verification. We propose and evaluate different baseline methods for DSD, both unsupervised, based on LIME, SHAP, LLMs, and our own method, as well as an additional supervised approach. While LLMs and supervised models achieve the highest performance, overall results remain low, highlighting the complexity of the task. Finally, we set up an additional experiment that shows how DSD can lead to increased performance in the specific task of paraphrase detection.

Method: Two-stage retrieve-and-extract framework with controlled study of retrieval design choices: varying context quantity, quality (standard RAG, reranking, fine-tuned retriever with reranking), and oracle paragraph setting. Four LLM extractors evaluated across these configurations to separate retrieval failures from extraction limits.

Result: Targeted context selection consistently improves hypothesis extraction relative to full-text prompting, with gains concentrated in configurations optimizing retrieval quality and context cleanliness. Statistical evidence extraction remains substantially harder - even with oracle paragraphs, performance remains moderate, indicating persistent extractor limitations in handling hybrid numeric-textual statements.

Conclusion: While targeted retrieval improves hypothesis extraction, statistical evidence extraction presents distinct challenges requiring specialized extractor capabilities for hybrid numeric-textual content, beyond just better retrieval.

Abstract: Extracting hypotheses and their supporting statistical evidence from full-text scientific articles is central to the synthesis of empirical findings, but remains difficult due to document length and the distribution of scientific arguments across sections of the paper. The work studies a sequential full-text extraction setting, where the statement of a primary finding in an article’s abstract is linked to (i) a corresponding hypothesis statement in the paper body and (ii) the statistical evidence that supports or refutes that hypothesis. This formulation induces a challenging within-document retrieval setting in which many candidate paragraphs are topically related to the finding but differ in rhetorical role, creating hard negatives for retrieval and extraction. Using a two-stage retrieve-and-extract framework, we conduct a controlled study of retrieval design choices, varying context quantity, context quality (standard Retrieval Augmented Generation, reranking, and a fine-tuned retriever paired with reranking), as well as an oracle paragraph setting to separate retrieval failures from extraction limits across four Large Language Model extractors. We find that targeted context selection consistently improves hypothesis extraction relative to full-text prompting, with gains concentrated in configurations that optimize retrieval quality and context cleanliness. In contrast, statistical evidence extraction remains substantially harder. Even with oracle paragraphs, performance remains moderate, indicating persistent extractor limitations in handling hybrid numeric-textual statements rather than retrieval failures alone.

Method: Proposes a cross-lingual graph fusion framework leveraging LLMs’ in-context reasoning and multilingual semantic priors. Uses structural linearization by mapping triplets to natural language sequences ([head] [relation] [tail]), enabling LLMs to map relations and reconcile entities between evolving fused graph and new candidate graphs

Result: Evaluated on DBP15K dataset, demonstrates LLMs can serve as universal semantic bridge to resolve cross-lingual discrepancies. Shows successful sequential agglomeration of multiple heterogeneous graphs

Conclusion: LLMs offer scalable, modular solution for continuous knowledge synthesis in multi-source, multilingual environments, effectively bridging cross-lingual semantic gaps in knowledge graph fusion

Abstract: Combining multiple knowledge graphs (KGs) across linguistic boundaries is a persistent challenge due to semantic heterogeneity and the complexity of graph environments. We propose a framework for cross-lingual graph fusion, leveraging the in-context reasoning and multilingual semantic priors of Large Language Models (LLMs). The framework implements structural linearization by mapping triplets directly into natural language sequences (e.g., [head] [relation] [tail]), enabling the LLM to map relations and reconcile entities between an evolving fused graph ($G_{c}^{(t-1)}$) and a new candidate graph ($G_{t}$). Evaluated on the DBP15K dataset, this exploratory study demonstrates that LLMs can serve as a universal semantic bridge to resolve cross-lingual discrepancies. Results show the successful sequential agglomeration of multiple heterogeneous graphs, offering a scalable, modular solution for continuous knowledge synthesis in multi-source, multilingual environments.

Method: Introduces Conversation Tree Architecture (CTA) - a hierarchical framework organizing conversations as trees of discrete, context-isolated nodes. Each node maintains its own local context window with structured mechanisms for context flow between parent and child nodes (downstream on creation, upstream on deletion). Includes volatile nodes for transient branches.

Result: Formalizes architecture primitives, characterizes design problems in context flow, relates framework to prior LLM memory management work, and describes a working prototype implementation.

Conclusion: CTA provides principled foundation for structured conversational context management and extends naturally to multi-agent settings, addressing the fundamental limitation of flat conversation interfaces.

Abstract: Large language models (LLMs) are increasingly deployed for extended, multi-topic conversations, yet the flat, append-only structure of current conversation interfaces introduces a fundamental limitation: all context accumulates in a single unbounded window, causing topically distinct threads to bleed into one another and progressively degrade response quality. We term this failure mode logical context poisoning. In this paper, we introduce the Conversation Tree Architecture (CTA), a hierarchical framework that organizes LLM conversations as trees of discrete, context-isolated nodes. Each node maintains its own local context window; structured mechanisms govern how context flows between parent and child nodes, downstream on branch creation and upstream on branch deletion. We additionally introduce volatile nodes, transient branches whose local context must be selectively merged upward or permanently discarded before purging. We formalize the architecture’s primitives, characterize the open design problems in context flow, relate our framework to prior work in LLM memory management, and describe a working prototype implementation. The CTA provides a principled foundation for structured conversational context management and extends naturally to multi-agent settings.

Method: 1) Curated Hate via Vision-Language Interplay (H-VLI) benchmark focusing on cases where true intent hinges on intricate modality interplay rather than overt slurs. 2) Proposed Asymmetric Reasoning via Courtroom Agent DEbate (ARCADE) framework simulating judicial process where agents argue for accusation and defense, forcing deep semantic scrutiny before verdict.

Result: ARCADE significantly outperforms state-of-the-art baselines on H-VLI benchmark, particularly for challenging implicit cases, while maintaining competitive performance on established benchmarks.

Conclusion: The work addresses critical gap in multimodal hate speech detection by focusing on semantic intent shifts through modality interplay, providing both benchmark and framework for better understanding complex multimodal expressions in online content moderation.

Abstract: Combating hate speech on social media is critical for securing cyberspace, yet relies heavily on the efficacy of automated detection systems. As content formats evolve, hate speech is transitioning from solely plain text to complex multimodal expressions, making implicit attacks harder to spot. Current systems, however, often falter on these subtle cases, as they struggle with multimodal content where the emergent meaning transcends the aggregation of individual modalities. To bridge this gap, we move beyond binary classification to characterize semantic intent shifts where modalities interact to construct implicit hate from benign cues or neutralize toxicity through semantic inversion. Guided by this fine-grained formulation, we curate the Hate via Vision-Language Interplay (H-VLI) benchmark where the true intent hinges on the intricate interplay of modalities rather than overt visual or textual slurs. To effectively decipher these complex cues, we further propose the Asymmetric Reasoning via Courtroom Agent DEbate (ARCADE) framework. By simulating a judicial process where agents actively argue for accusation and defense, ARCADE forces the model to scrutinize deep semantic cues before reaching a verdict. Extensive experiments demonstrate that ARCADE significantly outperforms state-of-the-art baselines on H-VLI, particularly for challenging implicit cases, while maintaining competitive performance on established benchmarks. Our code and data are available at: https://github.com/Sayur1n/H-VLI

Method: Three inference-time strategies evaluated: (1) self-consistency via stochastic decoding with temperature/nucleus sampling, (2) dual-model reasoning agreement comparing outputs from two independent models, (3) self-reflection where model critiques/revises its own reasoning. All use Chain-of-Thought prompting.

Result: Self-consistency with nucleus sampling and controlled temperature yields substantial gains (9-15% absolute improvement over greedy single-pass decoding). Dual-model approach provides additional confirmation for moderate-risk domains. Self-reflection offers only marginal improvements.

Conclusion: Self-consistency is well-suited for low-risk domains with minimal overhead, dual-model approach for moderate-risk domains requiring higher reliability, and self-reflection has limited effectiveness for smaller non-reasoning models at inference time.

Abstract: Large Language Models (LLMs) often exhibit strong linguistic abilities while remaining unreliable on multi-step reasoning tasks, particularly when deployed without additional training or fine-tuning. In this work, we study inference-time techniques to improve the reasoning accuracy of LLMs. We systematically evaluate three classes of inference-time strategies: (i) self-consistency via stochastic decoding, where the model is sampled multiple times using controlled temperature and nucleus sampling and the most frequent final answer is selected; (ii) dual-model reasoning agreement, where outputs from two independent models are compared and only consistent reasoning traces are trusted; and (iii) self-reflection, where the model critiques and revises its own reasoning. Across all evaluated methods, we employ Chain-of-Thought (CoT) [1] prompting to elicit explicit intermediate reasoning steps before generating final answers. In this work, we provide a controlled comparative evaluation across three inference-time strategies under identical prompting and verification settings. Our experiments on LLM [2] show that self-consistency with nucleus sampling and controlled temperature value yields the substantial gains, achieving a 9% to 15% absolute improvement in accuracy over greedy single-pass decoding, well-suited for low-risk domains, offering meaningful gains with minimal overhead. The dual-model approach provides additional confirmation for model reasoning steps thus more appropriate for moderate-risk domains, where higher reliability justifies additional compute. Self-reflection offers only marginal improvements, suggesting limited effectiveness for smaller non-reasoning models at inference time.

Method: Three-stage pipeline using Google’s Gemini models: 1) summary extraction from PDFs, 2) time toxicity quantification at six timepoints per treatment arm, 3) multi-run consensus via position-based arm matching. Compared single-pass vs two-stage (structure-then-count) architectures.

Result: Two-stage pipeline achieved 100% clinically acceptable accuracy (±3 days) on synthetic data (MAE 0.81 days) vs 41.5% for vanilla (MAE 9.0 days). On real-world protocols, vanilla pipeline showed superior reproducibility: 95.3% clinically acceptable accuracy across 3 runs on 644 protocols, with 82.0% perfect stability.

Conclusion: Extraction stability on real-world data, rather than accuracy on synthetic benchmarks, is decisive for production LLM deployment. Successfully extracted time toxicity for 1,288 treatment arms across multiple disease sites.

Abstract: Time toxicity, the cumulative healthcare contact days from clinical trial participation, is an important but labor-intensive metric to extract from protocol documents. We developed TimeTox, an LLM-based pipeline for automated extraction of time toxicity from Schedule of Assessments tables. TimeTox uses Google’s Gemini models in three stages: summary extraction from full-length protocol PDFs, time toxicity quantification at six cumulative timepoints for each treatment arm, and multi-run consensus via position-based arm matching. We validated against 20 synthetic schedules (240 comparisons) and assessed reproducibility on 644 real-world oncology protocols. Two architectures were compared: single-pass (vanilla) and two-stage (structure-then-count). The two-stage pipeline achieved 100% clinically acceptable accuracy ($\pm$3 days) on synthetic data (MAE 0.81 days) versus 41.5% for vanilla (MAE 9.0 days). However, on real-world protocols, the vanilla pipeline showed superior reproducibility: 95.3% clinically acceptable accuracy (IQR $\leq$ 3 days) across 3 runs on 644 protocols, with 82.0% perfect stability (IQR = 0). The production pipeline extracted time toxicity for 1,288 treatment arms across multiple disease sites. Extraction stability on real-world data, rather than accuracy on synthetic benchmarks, is the decisive factor for production LLM deployment.

Method: Introduces a “reduction ladder” - a sequence of modifications that progressively move instances away from canonical epistemic puzzles while preserving the underlying logic, making reduction increasingly difficult.

Result: While some large models succeed via reduction, others fail early, and all models struggle once true epistemic reasoning is required rather than just pattern matching.

Conclusion: Current LLMs’ performance on epistemic puzzles is better explained by their ability to reduce problems to known patterns rather than true reasoning, highlighting limitations in their epistemic reasoning capabilities.

Abstract: Epistemic reasoning requires agents to infer the state of the world from partial observations and information about other agents’ knowledge. Prior work evaluating LLMs on canonical epistemic puzzles interpreted their behavior through a dichotomy between epistemic reasoning and brittle memorization. We argue that this framing is incomplete: in recent models, memorization is better understood as a special case of reduction, where a new instance is mapped onto a known problem. Instead, we introduce a reduction ladder, a sequence of modifications that progressively move instances away from a canonical epistemic puzzle, making reduction increasingly difficult while preserving the underlying logic. We find that while some large models succeed via reduction, other models fail early, and all models struggle once epistemic reasoning is required.

Method: Two-phase framework: 1) Translate and gold-label standard Bengali questions into dialectal variants using RAG pipeline (4,000 question sets), evaluate translation fidelity using LLM-as-judge validated by human correlation; 2) Benchmark 19 LLMs across gold-labeled sets with 68,395 RLAIF evaluations validated through multi-judge agreement and human fallback.

Result: Reveals severe performance drops linked to linguistic divergence (e.g., Chittagong dialect scores 5.44/10 vs. Tangail’s 7.68/10). Increased model scale does not consistently mitigate bias. Provides validated translation quality evaluation method, benchmark dataset, and Critical Bias Sensitivity metric.

Conclusion: Establishes comprehensive framework for quantifying dialectal bias in LLMs, demonstrates significant performance disparities across Bengali dialects, and provides tools for safety-critical applications while showing that model scaling alone doesn’t solve dialect bias.

Abstract: Large language models (LLMs) frequently exhibit performance biases against regional dialects of low-resource languages. However, frameworks to quantify these disparities remain scarce. We propose a two-phase framework to evaluate dialectal bias in LLM question-answering across nine Bengali dialects. First, we translate and gold-label standard Bengali questions into dialectal variants adopting a retrieval-augmented generation (RAG) pipeline to prepare 4,000 question sets. Since traditional translation quality evaluation metrics fail on unstandardized dialects, we evaluate fidelity using an LLM-as-a-judge, which human correlation confirms outperforms legacy metrics. Second, we benchmark 19 LLMs across these gold-labeled sets, running 68,395 RLAIF evaluations validated through multi-judge agreement and human fallback. Our findings reveal severe performance drops linked to linguistic divergence. For instance, responses to the highly divergent Chittagong dialect score 5.44/10, compared to 7.68/10 for Tangail. Furthermore, increased model scale does not consistently mitigate this bias. We contribute a validated translation quality evaluation method, a rigorous benchmark dataset, and a Critical Bias Sensitivity (CBS) metric for safety-critical applications.

Method: Developed Conspiracy Frame based on frame-semantics and semiotics; created Con.Fra. dataset of Telegram messages with span-level annotations; tested LLMs’ ability to recognize conspiracy theories with and without frame injection; mapped annotated spans to FrameNet.

Result: Frame injection in in-context learning doesn’t significantly boost performance but shows potential; FrameNet mapping reveals abstract semantic patterns (e.g., ‘Kinship’, ‘Ingest_substance’) that could enable more semantically-aware conspiracy detection.

Conclusion: Conspiracy Frame and Con.Fra. dataset advance understanding of conspiratorial narratives; frame-semantic approaches show promise for developing more generalizable conspiracy theory detection systems.

Abstract: Conspiracy theories are anti-authoritarian narratives that lead to social conflict, impacting how people perceive political information. To help in understanding this issue, we introduce the Conspiracy Frame: a fine-grained semantic representation of conspiratorial narratives derived from frame-semantics and semiotics, which spawned the Conspiracy Frames (Con.Fra.) dataset: a corpus of Telegram messages annotated at span-level. The Conspiracy Frame and Con.Fra. dataset contribute to the implementation of a more generalizable understanding and recognition of conspiracy theories. We observe the ability of LLMs to recognize this phenomenon in-domain and out-of-domain, investigating the role that frames may have in supporting this task. Results show that, while the injection of frames in an in-context approach does not lead to clear increase of performance, it has potential; the mapping of annotated spans with FrameNet shows abstract semantic patterns (e.g., Kinship', Ingest_substance’) that potentially pave the way for a more semantically- and semiotically-aware detection of conspiratorial narratives.

Method: Comprehensive task-specific efficiency analysis comparing 16 language models across five diverse NLP tasks. Introduced Performance-Efficiency Ratio (PER) - a novel metric integrating accuracy, throughput, memory, and latency through geometric mean normalization.

Result: Small models (0.5-3B parameters) achieve superior PER scores across all tasks. These models provide the best efficiency-performance balance for production environments.

Conclusion: Small language models are quantitatively better suited for production deployments prioritizing inference efficiency over marginal accuracy gains. The PER metric provides a systematic way to evaluate model efficiency trade-offs.

Abstract: Large Language Models achieve remarkable performance but incur substantial computational costs unsuitable for resource-constrained deployments. This paper presents the first comprehensive task-specific efficiency analysis comparing 16 language models across five diverse NLP tasks. We introduce the Performance-Efficiency Ratio (PER), a novel metric integrating accuracy, throughput, memory, and latency through geometric mean normalization. Our systematic evaluation reveals that small models (0.5–3B parameters) achieve superior PER scores across all given tasks. These findings establish quantitative foundations for deploying small models in production environments prioritizing inference efficiency over marginal accuracy gains.

Method: Perspective-Driven Inference treats distribution of annotations across groups as quantity of interest, uses adaptive sampling strategy to concentrate human annotation on groups where LLM proxies are least accurate

Result: Shows targeted improvements for harder-to-model demographic groups relative to uniform sampling baselines on politeness and offensiveness rating tasks while maintaining coverage

Conclusion: PDI provides effective approach for capturing diverse human perspectives in subjective annotation tasks using limited human annotation budget

Abstract: Large language models are increasingly used to annotate texts, but their outputs reflect some human perspectives better than others. Existing methods for correcting LLM annotation error assume a single ground truth. However, this assumption fails in subjective tasks where disagreement across demographic groups is meaningful. Here we introduce Perspective-Driven Inference, a method that treats the distribution of annotations across groups as the quantity of interest, and estimates it using a small human annotation budget. We contribute an adaptive sampling strategy that concentrates human annotation effort on groups where LLM proxies are least accurate. We evaluate on politeness and offensiveness rating tasks, showing targeted improvements for harder-to-model demographic groups relative to uniform sampling baselines, while maintaining coverage.

Method: Evaluated 40 configurations combining four PEFT methods (LoRA, DoRA, QLoRA, QDoRA) across two BERTimbau model sizes (Base: 110M, Large: 335M parameters) on SQuAD-BR dataset for extractive question answering.

Result: LoRA achieved 95.8% of baseline performance on BERTimbau-Large while reducing training time by 73.5%; higher learning rates (2e-4) improved PEFT performance by up to +19.71 F1 points; larger models showed twice the quantization resilience.

Conclusion: Encoder-based models can be efficiently fine-tuned for Brazilian Portuguese QA with substantially lower computational cost than large generative LLMs, promoting sustainable Green AI approaches while maintaining competitive performance.

Abstract: Although large language models have transformed natural language processing, their computational costs create accessibility barriers for low-resource languages such as Brazilian Portuguese. This work presents a systematic evaluation of Parameter-Efficient Fine-Tuning (PEFT) and quantization techniques applied to BERTimbau for Question Answering on SQuAD-BR, the Brazilian Portuguese translation of SQuAD v1. We evaluate 40 configurations combining four PEFT methods (LoRA, DoRA, QLoRA, QDoRA) across two model sizes (Base: 110M, Large: 335M parameters). Our findings reveal three critical insights: (1) LoRA achieves 95.8% of baseline performance on BERTimbau-Large while reducing training time by 73.5% (F1=81.32 vs 84.86); (2) higher learning rates (2e-4) substantially improve PEFT performance, with F1 gains of up to +19.71 points over standard rates; and (3) larger models show twice the quantization resilience (loss of 4.83 vs 9.56 F1 points). These results demonstrate that encoder-based models can be efficiently fine-tuned for extractive Brazilian Portuguese QA with substantially lower computational cost than large generative LLMs, promoting more sustainable approaches aligned with \textit{Green AI} principles. An exploratory evaluation of Tucano and Sabiá on the same extractive QA benchmark shows that while generative models can reach competitive F1 scores with LoRA fine-tuning, they require up to 4.2$\times$ more GPU memory and 3$\times$ more training time than BERTimbau-Base, reinforcing the efficiency advantage of smaller encoder-based architectures for this task.

Method: Theoretical analysis of semantic smoothing in Transformer embeddings, formalizing semantic shift as a computable measure integrating local semantic evolution and global semantic dispersion. Controlled experiments across corpora and multiple embedding models.

Result: Semantic shift aligns closely with embedding concentration severity and predicts retrieval degradation, while text length alone does not. Semantic shift provides a unified lens for understanding embedding collapse.

Conclusion: Semantic shift (not just text length) is the key causal factor explaining when and why embedding pathologies harm retrieval, offering an actionable diagnostic tool.

Abstract: Transformer-based embedding models rely on pooling to map variable-length text into a single vector, enabling efficient similarity search but also inducing well-known geometric pathologies such as anisotropy and length-induced embedding collapse. Existing accounts largely describe \emph{what} these pathologies look like, yet provide limited insight into \emph{when} and \emph{why} they harm downstream retrieval. In this work, we argue that the missing causal factor is \emph{semantic shift}: the intrinsic, structured evolution and dispersion of semantics within a text. We first present a theoretical analysis of \emph{semantic smoothing} in Transformer embeddings: as the semantic diversity among constituent sentences increases, the pooled representation necessarily shifts away from every individual sentence embedding, yielding a smoothed and less discriminative vector. Building on this foundation, we formalize semantic shift as a computable measure integrating local semantic evolution and global semantic dispersion. Through controlled experiments across corpora and multiple embedding models, we show that semantic shift aligns closely with the severity of embedding concentration and predicts retrieval degradation, whereas text length alone does not. Overall, semantic shift offers a unified and actionable lens for understanding embedding collapse and for diagnosing when anisotropy becomes harmful.

Method: Proposes PROMPT2BOX which embeds prompts into a box embedding space using a trained encoder. The encoder is trained on existing and synthesized datasets to output box embeddings that capture semantic similarity and specificity relations. Also develops a novel dimension reduction technique for box embeddings for visualization.

Result: Box embeddings consistently capture prompt specificity better than vector baselines. On hierarchical clustering tasks for 17 LLMs from UltraFeedback dataset, PROMPT2BOX identifies 8.9% more LLM weaknesses than vector baselines and achieves ~33% stronger correlation between hierarchical depth and instruction specificity.

Conclusion: PROMPT2BOX provides a more nuanced representation of prompts that captures specificity relations, enabling better analysis of LLM weaknesses compared to traditional vector embeddings.

Abstract: To discover the weaknesses of LLMs, researchers often embed prompts into a vector space and cluster them to extract insightful patterns. However, vector embeddings primarily capture topical similarity. As a result, prompts that share a topic but differ in specificity, and consequently in difficulty, are often represented similarly, making fine-grained weakness analysis difficult. To address this limitation, we propose PROMPT2BOX, which embeds prompts into a box embedding space using a trained encoder. The encoder, trained on existing and synthesized datasets, outputs box embeddings that capture not only semantic similarity but also specificity relations between prompts (e.g., “writing an adventure story” is more specific than “writing a story”). We further develop a novel dimension reduction technique for box embeddings to facilitate dataset visualization and comparison. Our experiments demonstrate that box embeddings consistently capture prompt specificity better than vector baselines. On the downstream task of creating hierarchical clustering trees for 17 LLMs from the UltraFeedback dataset, PROMPT2BOX can identify 8.9% more LLM weaknesses than vector baselines and achieves an approximately 33% stronger correlation between hierarchical depth and instruction specificity.

Method: Propose KG-Hopper, a Reinforcement Learning framework that trains a Reasoning LLM to embed entire KG traversal and decision process into unified “thinking” stage, enabling global reasoning over cross-step dependencies with dynamic path exploration and backtracking.

Result: On eight KG reasoning benchmarks, KG-Hopper based on 7B-parameter LLM consistently outperforms larger multi-step systems (up to 70B) and achieves competitive performance with proprietary models like GPT-3.5-Turbo and GPT-4o-mini.

Conclusion: KG-Hopper enables compact open LLMs to perform integrated multi-hop KG reasoning efficiently in single inference round, addressing limitations of sequential approaches while remaining data-efficient.

Abstract: Large Language Models (LLMs) demonstrate impressive natural language capabilities but often struggle with knowledge-intensive reasoning tasks. Knowledge Base Question Answering (KBQA), which leverages structured Knowledge Graphs (KGs) exemplifies this challenge due to the need for accurate multi-hop reasoning. Existing approaches typically perform sequential reasoning steps guided by predefined pipelines, restricting flexibility and causing error cascades due to isolated reasoning at each step. To address these limitations, we propose KG-Hopper, a novel Reinforcement Learning (RL) framework that empowers compact open LLMs with the ability to perform integrated multi-hop KG reasoning within a single inference round. Rather than reasoning step-by-step, we train a Reasoning LLM that embeds the entire KG traversal and decision process into a unified ``thinking’’ stage, enabling global reasoning over cross-step dependencies and dynamic path exploration with backtracking. Experimental results on eight KG reasoning benchmarks show that KG-Hopper, based on a 7B-parameter LLM, consistently outperforms larger multi-step systems (up to 70B) and achieves competitive performance with proprietary models such as GPT-3.5-Turbo and GPT-4o-mini, while remaining compact, open, and data-efficient. The code is publicly available at: https://github.com/Wangshuaiia/KG-Hopper.

Method: Cross-Context Verification (CCV) solves same benchmark problems in N independent sessions and measures solution diversity, combined with Hierarchical Cross-Context Architecture (HCCA) multi-agent framework that restricts information across specialized roles to prevent confirmation bias.

Result: CCV achieves perfect separation between contaminated and genuine reasoning on SWE-bench Verified problems. Key findings: contamination is binary, reasoning absence perfectly discriminates, 33% of prior contamination labels are false positives, and HCCA discovers contamination-flaw composite cases.

Conclusion: Information restriction, not structural complexity, is key to detecting benchmark contamination. The method provides reliable detection of whether LLMs reason or recall memorized solutions.

Abstract: LLM coding benchmarks face a credibility crisis: widespread solution leakage and test quality issues undermine SWE-bench Verified, while existing detection methods–paraphrase consistency, n-gram overlap, perplexity analysis–never directly observe whether a model reasons or recalls. Meanwhile, simply repeating verification degrades accuracy: multi-turn review generates false positives faster than it discovers true errors, suggesting that structural approaches are needed. We introduce Cross-Context Verification (CCV), a black-box method that solves the same benchmark problem in N independent sessions and measures solution diversity, combined with the Hierarchical Cross-Context Architecture (HCCA), a multi-agent analysis framework that prevents confirmation bias through intentional information restriction across specialized analytical roles. On 9 SWE-bench Verified problems (45 trials, Claude Opus 4.6, temperature 0), CCV achieves perfect separation between contaminated and genuine reasoning (Mann-Whitney U=0, p approx 0.012, r = 1.0). Key findings: (1) contamination is binary–models either recall perfectly or not at all; (2) reasoning absence is a perfect discriminator; (3) 33% of prior contamination labels are false positives; (4) HCCA’s independent analysis structure discovers contamination-flaw composite cases that single-analyst approaches miss. A pilot experiment extending HCCA to multi-stage verification (Worker to Verifier to Director) yields a negative result–100% sycophantic confirmation–providing further evidence that information restriction, not structural complexity, is the key mechanism. We release all code and data.

Method: Three components: 1) CSP-DAG synthetic data algorithm for full operator coverage with controlled difficulty, 2) curriculum reinforcement learning with decoupled rewards for conversion success and speed optimization, 3) test-time search to further improve inference speed of generated Triton kernels.

Result: DRTriton-7B achieves speedup on 92% of KernelBench Level 2, significantly outperforming GPT-5.2 (23%) and Claude-Sonnet-4.5 (19%). It generalizes well to real-world CUDA kernels despite being trained only on synthetic data.

Conclusion: DRTriton demonstrates that specialized training frameworks can enable LLMs to effectively generate optimized CUDA kernels from PyTorch code, outperforming general-purpose LLMs and showing strong generalization to real-world scenarios.

Abstract: Developing efficient CUDA kernels is a fundamental yet challenging task in the generative AI industry. Recent researches leverage Large Language Models (LLMs) to automatically convert PyTorch reference implementations to CUDA kernels, significantly reducing the engineering efforts. State-of-the-art LLMs, such as GPT-5.2 and Claude-Sonnet-4.5, still struggle in this specific task. To address this challenge, we propose DRTriton, a scalable learning framework for training LLMs to convert PyTorch codes into highly optimized Triton kernels, which are then compiled to CUDA kernels at runtime. DRTriton consists of three key components: (i) a data synthetic algorithm CSP-DAG that guarantees full coverage and unbiased uniform sampling over the operator space with controlled difficulty; (ii) a curriculum reinforcement learning with decoupled reward efficiently optimizes conversion success rate and inference speed simultaneously; and (iii) a test-time search algorithm that further improves the inference speed of the generated Triton kernels. Notably, despite being trained exclusively on synthetic data, DRTriton generalizes effectively to real-world CUDA kernels that are challenging even for human experts. Experimental results show that DRTriton-7B achieves speedup on 92% of the KernelBench Level 2, compared to 23% for GPT-5.2 and 19% for Claude-Sonnet-4.5.

Method: CAID uses three core SWE primitives: 1) centralized task delegation through a central manager that creates dependency-aware plans, 2) asynchronous execution of subtasks concurrently, and 3) isolated workspaces for each agent. It employs git-based coordination mechanisms (worktree, commit, merge) with executable test-based verification for progress consolidation.

Result: CAID improves accuracy by 26.7% absolute on paper reproduction tasks (PaperBench) and 14.3% on Python library development tasks (Commit0) compared to single-agent baselines. Branch-and-merge is identified as a central coordination mechanism for effective multi-agent collaboration.

Conclusion: Software engineering collaboration primitives like git workflows provide reliable and executable mechanisms for multi-agent coordination, enabling effective asynchronous collaboration on complex, interdependent tasks through structured isolation and integration.

Abstract: AI agents have become increasingly capable at isolated software engineering (SWE) tasks such as resolving issues on Github. Yet long-horizon tasks involving multiple interdependent subtasks still pose challenges both with respect to accuracy, and with respect to timely completion. A natural approach to solving these long-horizon tasks in a timely manner is asynchronous multi-agent collaboration, where multiple agents work on different parts of the task at the same time. But effective application of multi-agent systems has proven surprisingly difficult: concurrent edits by multiple agents interfere with each other, dependencies are difficult to synchronize, and combining partial progress into a coherent whole is challenging. On the other hand, human developers have long relied on mature collaboration infrastructure to manage these challenges in large software projects. Inspired by these collaboration primitives, we introduce Centralized Asynchronous Isolated Delegation (CAID), a structured multi-agent coordination paradigm grounded in three core SWE primitives: centralized task delegation, asynchronous execution, and isolated workspaces. CAID constructs dependency-aware task plans through a central manager, executes subtasks concurrently in isolated workspaces, and consolidates progress via structured integration with executable test-based verification. In empirical evaluation, we find that CAID improves accuracy over single-agent baselines by 26.7% absolute on paper reproduction tasks (PaperBench) and 14.3% on Python library development tasks (Commit0). Through systematic analysis, we find that branch-and-merge is a central coordination mechanism for multi-agent collaboration, and that SWE primitives such as git worktree, git commit, and git merge enable it to be realized in a reliable and executable manner.

Method: Triangulated methodology combining quantitative analyses (lexical metrics, named entity recognition with Wikidata linking, targeted sentiment analysis, consensus-based change-point detection) with qualitative insights on 1.7M+ news articles across three linguistic regions.

Result: Revealed distinct temporal patterns in Swiss digital media and demonstrated how linguistic and cultural contexts influence reporting, with findings spanning thematic, recurrent, and singular events.

Conclusion: The approach provides a framework applicable to other multilingual/culturally diverse media environments, contributing to understanding how news is shaped by linguistic and cultural factors, and demonstrates the usefulness of triangulation in media studies.

Abstract: Analyzing news coverage in multilingual societies can offer valuable insights into the dynamics of public discourse and the development of collective narratives, yet comprehensive studies that account for linguistic and cultural diversity within national media ecosystems remain limited, particularly in complex contexts such as Switzerland. This paper studies temporal trends in Swiss digital media across the country’s three main linguistic regions, French, German, and Italian, using a triangulated methodology that combines quantitative analyses with qualitative insights. We collected and processed over 1.7 million news articles, applying lexical metrics, named entity recognition and Wikidata-based linking, targeted sentiment analysis, and consensus-based change-point detection. To enable principled cross-language comparisons and to connect to theories of domestication and cultural proximity, we derive domestication profiles together with a proximity salience ratio. Our analysis spans thematic, recurrent, and singular events. By integrating quantitative data with qualitative interpretation, we provide new insights into the dynamics of Swiss digital media and demonstrate the usefulness of triangulation in media studies. The findings reveal distinct temporal patterns and highlight how linguistic and cultural contexts influence reporting. Our approach offers a framework applicable to other multilingual or culturally diverse media environments, contributing to a deeper understanding of how news is shaped by linguistic and cultural factors.

Method: MemAPO uses a dual-memory mechanism: (1) distills successful reasoning trajectories into reusable strategy templates, and (2) organizes incorrect generations into structured error patterns capturing recurrent failure modes. For new prompts, it retrieves relevant strategies and failure patterns to compose prompts that promote effective reasoning while avoiding known mistakes.

Result: Experiments on diverse benchmarks show MemAPO consistently outperforms representative prompt optimization baselines while substantially reducing optimization cost.

Conclusion: MemAPO successfully reconceptualizes prompt optimization as generalizable and self-evolving experience accumulation, enabling continuous improvement over time rather than restarting from scratch for each task.

Abstract: Automatic prompt optimization is a promising approach for adapting large language models (LLMs) to downstream tasks, yet existing methods typically search for a specific prompt specialized to a fixed task. This paradigm limits generalization across heterogeneous queries and prevents models from accumulating reusable prompting knowledge over time. In this paper, we propose MemAPO, a memory-driven framework that reconceptualizes prompt optimization as generalizable and self-evolving experience accumulation. MemAPO maintains a dual-memory mechanism that distills successful reasoning trajectories into reusable strategy templates while organizing incorrect generations into structured error patterns that capture recurrent failure modes. Given a new prompt, the framework retrieves both relevant strategies and failure patterns to compose prompts that promote effective reasoning while discouraging known mistakes. Through iterative self-reflection and memory editing, MemAPO continuously updates its memory, enabling prompt optimization to improve over time rather than restarting from scratch for each task. Experiments on diverse benchmarks show that MemAPO consistently outperforms representative prompt optimization baselines while substantially reducing optimization cost.

Method: Proposes CatRAG Debiasing, a dual-pronged framework that integrates functor (category-theoretic structure for principled, structure-preserving projection) with Retrieval-Augmented Generation (RAG) guided structural debiasing to suppress bias-associated directions while retaining task-relevant semantics.

Result: On Bias Benchmark for Question Answering (BBQ) across three open-source LLMs (Meta Llama-3, OpenAI GPT-OSS, Google Gemma-3), CatRAG achieves SOTA results: improves accuracy by up to 40% over base models and >10% over prior debiasing methods, reducing bias scores to near zero (from 60% for base models) across gender, nationality, race, and intersectional subgroups.

Conclusion: CatRAG Debiasing effectively mitigates LLM biases through structural, category-theoretic approach combined with RAG, achieving superior bias reduction while maintaining utility.

Abstract: Large Language Models (LLMs) are deployed in high-stakes settings but can show demographic, gender, and geographic biases that undermine fairness and trust. Prior debiasing methods, including embedding-space projections, prompt-based steering, and causal interventions, often act at a single stage of the pipeline, resulting in incomplete mitigation and brittle utility trade-offs under distribution shifts. We propose CatRAG Debiasing, a dual-pronged framework that integrates functor with Retrieval-Augmented Generation (RAG) guided structural debiasing. The functor component leverages category-theoretic structure to induce a principled, structure-preserving projection that suppresses bias-associated directions in the embedding space while retaining task-relevant semantics. On the Bias Benchmark for Question Answering (BBQ) across three open-source LLMs (Meta Llama-3, OpenAI GPT-OSS, and Google Gemma-3), CatRAG achieves state-of-the-art results, improving accuracy by up to 40% over the corresponding base models and by more than 10% over prior debiasing methods, while reducing bias scores to near zero (from 60% for the base models) across gender, nationality, race, and intersectional subgroups.

Method: SynSym uses LLMs to generate synthetic training data by: 1) expanding symptoms into sub-concepts for diversity, 2) producing symptom expressions in diverse linguistic styles, and 3) composing realistic multi-symptom expressions based on clinical co-occurrence patterns.

Result: Models trained solely on SynSym’s synthetic data perform comparably to those trained on real data, with additional fine-tuning on real data providing further benefits. Validated on three benchmark datasets for depressive symptom expression.

Conclusion: Synthetic data generated by SynSym serves as a viable alternative to real-world annotations for psychiatric symptom modeling, offering a practical framework for creating clinically relevant and realistic symptom expressions.

Abstract: Psychiatric symptom identification on social media aims to infer fine-grained mental health symptoms from user-generated posts, allowing a detailed understanding of users’ mental states. However, the construction of large-scale symptom-level datasets remains challenging due to the resource-intensive nature of expert labeling and the lack of standardized annotation guidelines, which in turn limits the generalizability of models to identify diverse symptom expressions from user-generated text. To address these issues, we propose SynSym, a synthetic data generation framework for constructing generalizable datasets for symptom identification. Leveraging large language models (LLMs), SynSym constructs high-quality training samples by (1) expanding each symptom into sub-concepts to enhance the diversity of generated expressions, (2) producing synthetic expressions that reflect psychiatric symptoms in diverse linguistic styles, and (3) composing realistic multi-symptom expressions, informed by clinical co-occurrence patterns. We validate SynSym on three benchmark datasets covering different styles of depressive symptom expression. Experimental results demonstrate that models trained solely on the synthetic data generated by SynSym perform comparably to those trained on real data, and benefit further from additional fine-tuning with real data. These findings underscore the potential of synthetic data as an alternative resource to real-world annotations in psychiatric symptom modeling, and SynSym serves as a practical framework for generating clinically relevant and realistic symptom expressions.

Method: Created a 5,000 sentence pair English-Tashlhiyt corpus with a 1,500-sentence subset containing both expert-standardized and non-standard user-generated versions. Evaluated with state-of-the-art LLMs (GPT-5, Claude-Sonnet-4.5, Gemini-2.5-Pro, Mistral, Qwen3-Max) using zero-shot to few-shot prompting, and conducted fine-grained analysis of edit operations across phonological classes.

Result: Gemini-2.5-Pro achieved the lowest word and character-level error rates and exhibited robust cross-lingual generalization. Fine-grained analysis revealed model-specific sensitivities to marked Tashlhiyt phonological features (geminates, emphatics, uvulars, pharyngeals), providing diagnostic insights for orthography normalization.

Conclusion: DATASHI fills a critical resource gap for Amazigh languages, enables systematic study of orthographic diversity, supports text-based NLP tasks and multimodal applications, and provides valuable diagnostic insights through LLM evaluations for low-resource language processing.

Abstract: DATASHI is a new parallel English-Tashlhiyt corpus that fills a critical gap in computational resources for Amazigh languages. It contains 5,000 sentence pairs, including a 1,500-sentence subset with expert-standardized and non-standard user-generated versions, enabling systematic study of orthographic diversity and normalization. This dual design supports text-based NLP tasks - such as tokenization, translation, and normalization - and also serves as a foundation for read-speech data collection and multimodal alignment. Comprehensive evaluations with state-of-the-art Large Language Models (GPT-5, Claude-Sonnet-4.5, Gemini-2.5-Pro, Mistral, Qwen3-Max) show clear improvements from zero-shot to few-shot prompting, with Gemini-2.5-Pro achieving the lowest word and character-level error rates and exhibiting robust cross-lingual generalization. A fine-grained analysis of edit operations - deletions, substitutions, and insertions - across phonological classes (geminates, emphatics, uvulars, and pharyngeals) further highlights model-specific sensitivities to marked Tashlhiyt features and provides new diagnostic insights for low-resource Amazigh orthography normalization.

Method: Collected multiple model series (Pythia, OpenLLaMa, StarCoder, OLMo1/2/3) and analyzed memorization behavior at both statistical and internal levels. Statistical analysis examined scaling laws, compression, and distribution patterns. Internal analysis studied perturbation removal, sensitivity, layer decoding, and attention head ablation.

Result: Found that memorization rate scales log-linearly with model size, memorized sequences can be further compressed, and there are shared frequency/domain distribution patterns. However, models also show individual features. Internally, LLMs can remove certain perturbations while memorized sequences are more sensitive. Identified general decoding processes and shared important heads for memorization, but distribution of important heads differs between families.

Conclusion: The study bridges various experiments to reveal both universal principles and family-specific features of memorization in LLMs, paving the way for a more fundamental understanding of memorization across different model architectures.

Abstract: Memorization is a fundamental component of intelligence for both humans and LLMs. However, while LLM performance scales rapidly, our understanding of memorization lags. Due to limited access to the pre-training data of LLMs, most previous studies focus on a single model series, leading to isolated observations among series, making it unclear which findings are general or specific. In this study, we collect multiple model series (Pythia, OpenLLaMa, StarCoder, OLMo1/2/3) and analyze their shared or unique memorization behavior at both the statistical and internal levels, connecting individual observations while showing new findings. At the statistical level, we reveal that the memorization rate scales log-linearly with model size, and memorized sequences can be further compressed. Further analysis demonstrated a shared frequency and domain distribution pattern for memorized sequences. However, different models also show individual features under the above observations. At the internal level, we find that LLMs can remove certain injected perturbations, while memorized sequences are more sensitive. By decoding middle layers and attention head ablation, we revealed the general decoding process and shared important heads for memorization. However, the distribution of those important heads differs between families, showing a unique family-level feature. Through bridging various experiments and revealing new findings, this study paves the way for a universal and fundamental understanding of memorization in LLM.

Method: TAMTRL (Teacher-Aligned Reward Reshaping for Multi-Turn RL) uses relevant documents as teacher signals, aligns them with each turn’s model input, and assigns rewards through normalized probabilities in a self-supervised manner to provide fine-grained learning signals for each memory update.

Result: Experiments with multiple models across seven long-context benchmarks show TAMTRL consistently outperforms strong baselines, demonstrating effectiveness in improving long-context processing.

Conclusion: TAMTRL effectively addresses the credit assignment problem in multi-turn memory training for LLMs, providing a practical solution for long-document processing without substantial computational overhead.

Abstract: The rapid progress of large language models (LLMs) has led to remarkable performance gains across a wide range of tasks. However, when handling long documents that exceed the model’s context window limit, the entire context cannot be processed in a single pass, making chunk-wise processing necessary. This requires multiple turns to read different chunks and update memory. However, supervision is typically provided only by the final outcome, which makes it difficult to evaluate the quality of memory updates at each turn in the multi-turn training setting. This introduces a temporal credit assignment challenge. Existing approaches, such as LLM-as-a-judge or process reward models, incur substantial computational overhead and suffer from estimation noise. To better address the credit assignment problem in multi-turn memory training, we propose Teacher-Aligned Reward Reshaping for Multi-Turn Reinforcement Learning (TAMTRL). TAMTRL leverages relevant documents as teacher signals by aligning them with each turn of model input and assigns rewards through normalized probabilities in a self-supervised manner. This provides fine-grained learning signals for each memory update and improves long-context processing. Experiments with multiple models of varying scales across seven long-context benchmarks show that TAMTRL consistently outperforms strong baselines, demonstrating its effectiveness. Our code is available at https://anonymous.4open.science/r/TAMTRL-F1F8.

Method: WeatherTGD uses a training-free multi-agent framework with three specialized LLM agents: Statistical Analyst, Physics Interpreter, and Meteorology Expert. These agents generate domain-specific textual gradients from weather time series observations. A Consensus-Aware Gradient Fusion mechanism aggregates these gradients, preserving unique domain perspectives while extracting common signals. The fused gradients guide an iterative refinement process analogous to gradient descent.

Result: Experiments on real-world meteorological datasets show WeatherTGD achieves significant improvements in both LLM-based evaluation and human expert evaluation, substantially outperforming existing multi-agent baselines while maintaining computational efficiency through parallel agent execution.

Conclusion: WeatherTGD successfully addresses the challenge of generating interpretable natural language captions from weather time series data by combining domain expertise with collaborative refinement through Text Gradient Descent, offering a promising approach for meteorological NLP applications.

Abstract: Generating interpretable natural language captions from weather time series data remains a significant challenge at the intersection of meteorological science and natural language processing. While recent advances in Large Language Models (LLMs) have demonstrated remarkable capabilities in time series forecasting and analysis, existing approaches either produce numerical predictions without human-accessible explanations or generate generic descriptions lacking domain-specific depth. We introduce WeatherTGD, a training-free multi-agent framework that reinterprets collaborative caption refinement through the lens of Text Gradient Descent (TGD). Our system deploys three specialized LLM agents including a Statistical Analyst, a Physics Interpreter, and a Meteorology Expert that generate domain-specific textual gradients from weather time series observations. These gradients are aggregated through a novel Consensus-Aware Gradient Fusion mechanism that extracts common signals while preserving unique domain perspectives. The fused gradients then guide an iterative refinement process analogous to gradient descent, where each LLM-generated feedback signal updates the caption toward an optimal solution. Experiments on real-world meteorological datasets demonstrate that WeatherTGD achieves significant improvements in both LLM-based evaluation and human expert evaluation, substantially outperforming existing multi-agent baselines while maintaining computational efficiency through parallel agent execution.

Method: Mathematically analyze tabular dependency structures using mutual information to reveal periodic non-vanishing dependencies. Propose TableLong - a scalable pipeline for synthesizing high-quality, diverse, and verifiable structured table data using RL to boost long-context reasoning.

Result: Table data significantly enhances LLM long-context reasoning across multiple benchmarks (+8.24% average improvement), and even improves performance on out-of-domain benchmarks (+8.06% average).

Conclusion: Structured table data is effective for enhancing long-context reasoning in LLMs, providing practical guidance for post-training data selection. The proposed TableLong pipeline demonstrates scalable synthesis of beneficial table data.

Abstract: As real-world tasks grow increasingly complex, long-context reasoning has become a core capability for Large Language Models (LLMs). However, few studies explore which data types are effective for long-context reasoning and why. We find that structured table data with periodic structures shows strong potential for long-context reasoning. Motivated by this observation, we mathematically analyze tabular dependency structures using mutual information, revealing periodic non-vanishing dependencies in table data. Furthermore, we systematically analyze the capabilities of structured table data, conduct relevant scaling experiments, and validate its underlying mechanisms for enhancing long-context reasoning, yielding several meaningful insights. Leveraging these insights, we propose a simple yet scalable pipeline(TableLong) for synthesizing high-quality, diverse, and verifiable structured table data to boost long-context reasoning via RL. Extensive experimental results demonstrate that table data significantly enhances the long-context reasoning capability of LLMs across multiple long-context benchmarks (+8.24% on average), and even improves performance on out-of-domain benchmarks (+8.06% on average). We hope that our insights provide practical guidance for effective post-training data to enhance long-context reasoning in LLMs.

Method: Organized SemEval-2026 Task 12 as an evidence-grounded multiple-choice benchmark capturing real-world causal reasoning challenges including distributed evidence, indirect background factors, and semantically related non-causal distractors.

Result: The shared task attracted 122 participants and received 518 submissions, demonstrating significant community interest in abductive event reasoning.

Conclusion: AER provides a focused benchmark for abductive reasoning over real-world events and highlights challenges for future work on causal reasoning and multi-document understanding.

Abstract: Understanding why real-world events occur is important for both natural language processing and practical decision-making, yet direct-cause inference remains underexplored in evidence-rich settings. To address this gap, we organized SemEval-2026 Task 12: Abductive Event Reasoning (AER).\footnote{The task data is available at https://github.com/sooo66/semeval2026-task12-dataset.git} The task asks systems to identify the most plausible direct cause of a target event from supporting evidence. We formulate AER as an evidence-grounded multiple-choice benchmark that captures key challenges of real-world causal reasoning, including distributed evidence, indirect background factors, and semantically related but non-causal distractors. The shared task attracted 122 participants and received 518 submissions. This paper presents the task formulation, dataset construction pipeline, evaluation setup, and system results. AER provides a focused benchmark for abductive reasoning over real-world events and highlights challenges for future work on causal reasoning and multi-document understanding.

Method: Mixed-methods approach using over one million French-language digital news articles (Jan 2023-Jun 2024). Automated detection of interrogative stances, functional type approximation, and location of textual answers. Combined with qualitative annotation of subcorpus grounded in semantic and pragmatic theories of questions.

Result: Interrogatives are sparse but systematically patterned: mainly introduce/organize issues, with most being information-seeking or echo-like. Questions overwhelmingly taken up within same article with answer-like spans, usually in journalist’s narrative voice. Interrogative contexts densely populated with named individuals/organizations/places, showing strong personalization and foregrounding of prominent actors.

Conclusion: Demonstrates how interrogative stance, textual uptake, and voice can be operationalized at corpus scale. Combining computational methods with pragmatic and sociological perspectives helps account for how questioning practices structure contemporary news discourse.

Abstract: Interrogatives in news discourse have been examined in linguistics and conversation analysis, but mostly in broadcast interviews and relatively small, often English-language corpora, while large-scale computational studies of news rarely distinguish interrogatives from declaratives or differentiate their functions. This paper brings these strands together through a mixed-methods study of the “Politics of Questions” in contemporary French-language digital news. Using over one million articles published between January 2023 and June 2024, we automatically detect interrogative stances, approximate their functional types, and locate textual answers when present, linking these quantitative measures to a qualitatively annotated subcorpus grounded in semantic and pragmatic theories of questions. Interrogatives are sparse but systematically patterned: they mainly introduce or organize issues, with most remaining cases being information-seeking or echo-like, while explicitly leading or tag questions are rare. Although their density and mix vary across outlets and topics, our heuristic suggests that questions are overwhelmingly taken up within the same article and usually linked to a subsequent answer-like span, most often in the journalist’s narrative voice and less often through quoted speech. Interrogative contexts are densely populated with named individuals, organizations, and places, whereas publics and broad social groups are mentioned much less frequently, suggesting that interrogative discourse tends to foreground already prominent actors and places and thus exhibits strong personalization. We show how interrogative stance, textual uptake, and voice can be operationalized at corpus scale, and argue that combining computational methods with pragmatic and sociological perspectives can help account for how questioning practices structure contemporary news discourse.

Method: IsalSR uses a representation framework that encodes expression DAGs as strings over a compact two-tier alphabet and computes a pruned canonical string that serves as a complete labeled-DAG isomorphism invariant, collapsing all equivalent representations into a single canonical form.

Result: The method eliminates structural redundancy by providing a canonical representation that collapses equivalent expression DAG representations into a single form, improving search efficiency in symbolic regression.

Conclusion: IsalSR provides an effective solution to structural redundancy in symbolic regression through canonical representation, potentially improving the efficiency and effectiveness of symbolic regression algorithms.

Abstract: A fundamental but largely unaddressed obstacle in Symbolic regression (SR) is structural redundancy: every expression DAG with admits many distinct node-numbering schemes that all encode the same expression, each occupying a separate point in the search space and consuming fitness evaluations without adding diversity. We present IsalSR (Instruction Set and Language for Symbolic Regression), a representation framework that encodes expression DAGs as strings over a compact two-tier alphabet and computes a pruned canonical string – a complete labeled-DAG isomorphism invariant – that collapses all the equivalent representations into a single canonical form.

Method: Formalizes Active Testing framework for NLP and benchmarks existing approaches across 18 datasets and 4 embedding strategies spanning 4 different NLP tasks. Introduces adaptive stopping criterion to automatically determine optimal number of samples instead of requiring predefined annotation budget.

Result: Achieves annotation reductions of up to 95% with performance estimation accuracy difference from full test set within 1%. Analysis shows variations in method effectiveness across different data characteristics and task types, with no single approach emerging as universally superior.

Conclusion: Active Testing provides effective framework for reducing annotation costs in NLP evaluation while maintaining reliable performance estimation, with adaptive stopping criterion addressing practical limitations of existing approaches.

Abstract: Human annotation cost and time remain significant bottlenecks in Natural Language Processing (NLP), with test data annotation being particularly expensive due to the stringent requirement for low-error and high-quality labels necessary for reliable model evaluation. Traditional approaches require annotating entire test sets, leading to substantial resource requirements. Active Testing is a framework that selects the most informative test samples for annotation. Given a labeling budget, it aims to choose the subset that best estimates model performance while minimizing cost and human effort. In this work, we formalize Active Testing in NLP and we conduct an extensive benchmarking of existing approaches across 18 datasets and 4 embedding strategies spanning 4 different NLP tasks. The experiments show annotation reductions of up to 95%, with performance estimation accuracy difference from the full test set within 1%. Our analysis reveals variations in method effectiveness across different data characteristics and task types, with no single approach emerging as universally superior. Lastly, to address the limitation of requiring a predefined annotation budget in existing sample selection strategies, we introduce an adaptive stopping criterion that automatically determines the optimal number of samples.

Method: Used word-level EEG data from 30 participants reading sentences (ZuCo corpus), trained separate linear probes for each person mapping hidden states from frozen Qwen 2.5 7B to individual EEG power, compared person-specific vs population probes, and analyzed temporal stability and layer-wise patterns.

Result: Person-specific probes significantly outperformed population probes (9x improvement for high-gamma power), showed temporal stability, non-transferability across individuals, and concentrated predictive power in deep layers (peaking at Layer 24 of 28). Results consistent across architectures and survived confound controls.

Conclusion: Frozen language models contain stable, person-specific neural directions in their deep layers, providing a geometric foundation for EEG-driven personalization of language models for individual users.

Abstract: Consumer-grade EEG is entering everyday devices, from earbuds to headbands, raising the question of whether language models can be adapted to individual neural responses. We test this by asking whether frozen LLM representations encode person-specific EEG signals, directions in activation space that predict one person’s brain activity but not another’s. Using word-level EEG from 30 participants reading naturalistic sentences (ZuCo corpus), we train a separate linear probe for each person, mapping hidden states from a frozen Qwen 2.5 7B to that individual’s EEG power. Person-specific probes outperform a single population probe on every EEG feature tested; for high-gamma power, the person-specific probe achieves rho = 0.183, a ninefold improvement over the population probe (rho = 0.020, p < 10^-4). A negative control, fixation count, shows no person-specific advantage (p = 0.360); fixation count reflects word length and frequency rather than individual cognition. The individual directions are temporally stable (split-half cosine = 0.824), non-transferable across people (self rho = 0.369 vs. other rho = 0.143, p < 10^-19), and distinct from the shared population signal: person-specific probes retain predictive power after the population component is removed. The person-specific signal concentrates in the model’s deep layers, rising consistently with depth and peaking at Layer 24 of 28. The results are consistent across architectures (LLaMA 3.1 8B) and survive word-level confound controls. Frozen language models contain stable, person-specific neural directions in their deep layers, providing a geometric foundation for EEG-driven personalization.

Method: Pre-trained conformer-based BEST-RQ models on 5,640 hours of Creative Commons Arabic speech combined with public datasets. Used self-supervised learning approach with models up to 600M parameters, specifically targeting Arabic dialect families.

Result: Achieved state-of-the-art performance on dialect identification (DID) tasks while using fewer parameters than competing models. Demonstrated that Arabic-targeted pre-training significantly outperforms multilingual or monolingual models trained on non-Arabic data.

Conclusion: Family-targeted pre-training on Arabic dialects is crucial for optimal performance in Arabic speech tasks. The released models, code, and datasets will support further research in Arabic speech technologies.

Abstract: We present Ara-BEST-RQ, a family of self-supervised learning (SSL) models specifically designed for multi-dialectal Arabic speech processing. Leveraging 5,640 hours of crawled Creative Commons speech and combining it with publicly available datasets, we pre-train conformer-based BEST-RQ models up to 600M parameters. Our models are evaluated on dialect identification (DID) and automatic speech recognition (ASR) tasks, achieving state-of-the-art performance on the former while using fewer parameters than competing models. We demonstrate that family-targeted pre-training on Arabic dialects significantly improves downstream performance compared to multilingual or monolingual models trained on non-Arabic data. All models, code, and pre-processed datasets will be publicly released to support reproducibility and further research in Arabic speech technologies.

Method: Created SLURP-TN by recording 55 native speakers uttering sentences manually translated from six SLURP domains into Tunisian dialect, then developed ASR and SLU baseline models.

Result: Produced dataset of 4165 sentences (~5 hours audio) with baseline models; dataset and models publicly available on Hugging Face.

Conclusion: SLURP-TN helps mitigate resource scarcity for Tunisian dialect SLU, enabling research and development for low-resource languages in spoken language understanding.

Abstract: Spoken Language Understanding (SLU) aims to extract the semantic information from the speech utterance of user queries. It is a core component in a task-oriented dialogue system. With the spectacular progress of deep neural network models and the evolution of pre-trained language models, SLU has obtained significant breakthroughs. However, only a few high-resource languages have taken advantage of this progress due to the absence of SLU resources. In this paper, we seek to mitigate this obstacle by introducing SLURP-TN. This dataset was created by recording 55 native speakers uttering sentences in Tunisian dialect, manually translated from six SLURP domains. The result is an SLU Tunisian dialect dataset that comprises 4165 sentences recorded into around 5 hours of acoustic material. We also develop a number of Automatic Speech Recognition and SLU models exploiting SLUTP-TN. The Dataset and baseline models are available at: https://huggingface.co/datasets/Elyadata/SLURP-TN.

Method: Compares three adaptation approaches (LoRA, Prompt Tuning, and Full Fine-Tuning) across the Flan-T5 model family on the PubMed medical summarization dataset, with sensitivity analyses on LoRA rank and prompt token count.

Result: LoRA consistently outperforms full fine-tuning, achieving 43.52 +/- 0.18 ROUGE-1 on Flan-T5-Large vs 40.67 +/- 0.21 for full fine-tuning, using only 0.6% trainable parameters.

Conclusion: The low-rank constraint in LoRA provides beneficial regularization, challenging assumptions about the necessity of full parameter updates for domain adaptation.

Abstract: Fine-tuning large language models for domain-specific tasks such as medical text summarization demands substantial computational resources. Parameter-efficient fine-tuning (PEFT) methods offer promising alternatives by updating only a small fraction of parameters. This paper compares three adaptation approaches-Low-Rank Adaptation (LoRA), Prompt Tuning, and Full Fine-Tuning-across the Flan-T5 model family on the PubMed medical summarization dataset. Through experiments with multiple random seeds, we demonstrate that LoRA consistently outperforms full fine-tuning, achieving 43.52 +/- 0.18 ROUGE-1 on Flan-T5-Large with only 0.6% trainable parameters compared to 40.67 +/- 0.21 for full fine-tuning. Sensitivity analyses examine the impact of LoRA rank and prompt token count. Our findings suggest the low-rank constraint provides beneficial regularization, challenging assumptions about the necessity of full parameter updates. Code is available at https://github.com/eracoding/llm-medical-summarization

Method: Reformulates narrative detection as retrieval task using narrative core messages as queries; introduces SpecFi framework that generates hypothetical documents using community summaries from graph-based community detection as few-shot examples; proposes narrative variance metric based on embeddings.

Result: SpecFi achieves MAP of 0.505 on CARDS dataset without narrative labels; shows robustness to high-variance narratives (32.7% MAP loss vs 63.4% for BM25); unsupervised community summaries converge on descriptions close to expert-crafted taxonomies.

Conclusion: Retrieval-based approach enables flexible detection of emerging climate disinformation narratives without predefined taxonomies; graph-based methods can effectively surface narrative structure from unlabeled text; SpecFi framework provides robust performance.

Abstract: Detecting climate disinformation narratives typically relies on fixed taxonomies, which do not accommodate emerging narratives. Thus, we re-frame narrative detection as a retrieval task: given a narrative’s core message as a query, rank texts from a corpus by alignment with that narrative. This formulation requires no predefined label set and can accommodate emerging narratives. We repurpose three climate disinformation datasets (CARDS, Climate Obstruction, climate change subset of PolyNarrative) for retrieval evaluation and propose SpecFi, a framework that generates hypothetical documents to bridge the gap between abstract narrative descriptions and their concrete textual instantiations. SpecFi uses community summaries from graph-based community detection as few-shot examples for generation, achieving a MAP of 0.505 on CARDS without access to narrative labels. We further introduce narrative variance, an embedding-based difficulty metric, and show via partial correlation analysis that standard retrieval degrades on high-variance narratives (BM25 loses 63.4% of MAP), while SpecFi-CS remains robust (32.7% loss). Our analysis also reveals that unsupervised community summaries converge on descriptions close to expert-crafted taxonomies, suggesting that graph-based methods can surface narrative structure from unlabeled text.

Method: First analyzes DSKD-CMA’s attention mechanism through token alignment probing and heatmap visualizations. Then introduces DSKD-CMA-GA, which uses generative adversarial learning to address distribution mismatches between keys and queries from distinct models with different tokenizers.

Result: Shows modest but consistent ROUGE-L gains in text generation quality, particularly on out-of-distribution data (+0.37 average improvement), narrowing the performance gap between cross-tokenizer and same-tokenizer knowledge distillation.

Conclusion: The proposed DSKD-CMA-GA method effectively addresses distribution mismatch issues in cross-tokenizer knowledge distillation, improving text generation quality especially for out-of-distribution data.

Abstract: Large language models (LLMs) achieve state-of-the-art (SOTA) performance across language tasks, but are costly to deploy due to their size and resource demands. Knowledge Distillation (KD) addresses this by training smaller Student models to mimic larger Teacher models, improving efficiency without significant performance loss. Dual-Space Knowledge Distillation with Cross-Model Attention (DSKD-CMA) has emerged as a SOTA method for KD between LLMs with distinct tokenizers, yet its internal workings remain largely opaque. In this work, we systematically analyse the attention mechanism of DSKD-CMA through manual token alignment probing and heatmap visualisations, revealing both strengths and limitations. Building on this, we introduce a novel method, DSKD-CMA-GA, based on Generative Adversarial (GA) learning, to address the mismatched distributions between the keys and queries computed from distinct models. Experiments show modest but consistent ROUGE-L gains in text generation quality, particularly on out-of-distribution data (+0.37 on average), narrowing the gap between cross- and same-tokenizer KD.

Method: Train both AR and MDLM models on identical data (50M tokens from TinyStories), identical compute budget (20,000 steps, batch size 32, sequence length 512), and identical hardware (NVIDIA H100 80GB). Compare training throughput, convergence patterns, and analyze quantitative diversity of 1,000 generated samples.

Result: 1) Comparable training throughput (~50K tokens/second, MDLM requires 4.7% more time); 2) AR converges faster and overfits by step 14,000, while MDLM converges slower and still improves at step 20,000; 3) AR produces fluent but repetitive outputs (99.8% begin with same word), while MDLM generates more diverse narratives (93.4% unique 5-word openings, higher Distinct-n, lower Self-BLEU) with occasional grammatical issues.

Conclusion: The generation paradigm significantly impacts training dynamics and output characteristics: AR models converge faster but produce less diverse outputs, while MDLM models converge slower but generate more diverse content, revealing a structural diversity-fluency trade-off between the two approaches.

Abstract: We present a controlled empirical comparison between autoregressive (AR) and masked diffusion (MDLM) language models. Both models are trained on identical data (50M tokens from TinyStories), identical compute budget (20,000 steps, batch size 32, sequence length 512), and identical hardware (NVIDIA H100 80GB), isolating the generation paradigm as the sole variable. We report three findings. First, both paradigms achieve comparable training throughput (~50K tokens/second), with MDLM requiring only 4.7% more wall-clock time. Second, AR converges faster and begins overfitting by step 14,000, while MDLM converges more slowly and is still improving at step 20,000, suggesting different compute-optimal training regimes. Third, quantitative diversity analysis over 1,000 generated samples reveals a structural diversity-fluency trade-off: AR produces fluent but repetitive outputs (99.8% begin with the same word), while MDLM generates more diverse narratives (93.4% unique 5-word openings, higher Distinct-n, lower Self-BLEU), at the cost of occasional grammatical inconsistencies. All code, trained checkpoints, and data pipelines are released for reproducibility.

Method: Created ensemble of 31 LLM personas ranging from practitioners following specific interpretive frameworks to intuitive lay-style characters. Used majority voting with ensemble on SemEval-2026 Task 4 dataset, analyzing performance dynamics and error correlations.

Result: Achieved 0.705 accuracy on SemEval-2026 Task 4. Accuracy improves with ensemble size following Condorcet Jury Theorem-like dynamics. Practitioner personas perform worse individually but produce less correlated errors, yielding larger ensemble gains. Gender-focused interpretive vocabulary consistently associated with lower accuracy across personas.

Conclusion: Interpretive plurality should be incorporated in predictive systems rather than treated as noise. Ensemble methods with diverse personas can handle multiperspectivity. Need for evaluation frameworks that account for interpretive diversity beyond single ground truths.

Abstract: Predicting narrative similarity can be understood as an inherently interpretive task: different, equally valid readings of the same text can produce divergent interpretations and thus different similarity judgments, posing a fundamental challenge for semantic evaluation benchmarks that encode a single ground truth. Rather than treating this multiperspectivity as a challenge to overcome, we propose to incorporate it in the decision making process of predictive systems. To explore this strategy, we created an ensemble of 31 LLM personas. These range from practitioners following interpretive frameworks to more intuitive, lay-style characters. Our experiments were conducted on the SemEval-2026 Task 4 dataset, where the system achieved an accuracy score of 0.705. Accuracy improves with ensemble size, consistent with Condorcet Jury Theorem-like dynamics under weakened independence. Practitioner personas perform worse individually but produce less correlated errors, yielding larger ensemble gains under majority voting. Our error analysis reveals a consistent negative association between gender-focused interpretive vocabulary and accuracy across all persona categories, suggesting either attention to dimensions not relevant for the benchmark or valid interpretations absent from the ground truth. This finding underscores the need for evaluation frameworks that account for interpretive plurality.

Method: Introduces the Semantic Ladder architectural framework that organizes representations across increasing levels of semantic explicitness, from natural language text snippets to ontology-based and higher-order logical models, using modular semantic units as carriers of meaning.

Result: The framework enables incremental construction of semantic knowledge spaces, reduces semantic parsing burden, and supports integration of heterogeneous representations including natural language, structured semantic models, and vector-based embeddings.

Conclusion: The Semantic Ladder provides a foundation for scalable, interoperable, and AI-ready data and knowledge infrastructures by enabling progressive formalization while preserving semantic continuity and traceability.

Abstract: Semantic data and knowledge infrastructures must reconcile two fundamentally different forms of representation: natural language, in which most knowledge is created and communicated, and formal semantic models, which enable machine-actionable integration, interoperability, and reasoning. Bridging this gap remains a central challenge, particularly when full semantic formalization is required at the point of data entry. Here, we introduce the Semantic Ladder, an architectural framework that enables the progressive formalization of data and knowledge. Building on the concept of modular semantic units as identifiable carriers of meaning, the framework organizes representations across levels of increasing semantic explicitness, ranging from natural language text snippets to ontology-based and higher-order logical models. Transformations between levels support semantic enrichment, statement structuring, and logical modelling while preserving semantic continuity and traceability. This approach enables the incremental construction of semantic knowledge spaces, reduces the semantic parsing burden, and supports the integration of heterogeneous representations, including natural language, structured semantic models, and vector-based embeddings. The Semantic Ladder thereby provides a foundation for scalable, interoperable, and AI-ready data and knowledge infrastructures.

Method: Two-stage approach: 1) Data augmentation by converting summarization data into document-level parallel data using LLMs, then filtering with multiple metrics (sacreBLEU, COMET, LaBSE-based cosine similarity); 2) Two-stage fine-tuning: first on abundant sentence-level MT resources, then on filtered document-level corpus.

Result: Not specified in the abstract, but the method addresses key challenges in document-level MT with LLMs through data augmentation and staged fine-tuning.

Conclusion: Proposed approach tackles data scarcity and generation quality issues in document-level machine translation using LLMs through innovative data augmentation and staged fine-tuning strategies.

Abstract: In Machine Translation, Large Language Models (LLMs) have generally underperformed compared to conventional encoder-decoder systems and thus see limited adoption. However, LLMs excel at modeling contextual information, making them a natural fit for document-level translation tasks where coherence across sentences is crucial. Despite this potential, document-level MT with LLMs faces two key challenges: (1) the scarcity of large-scale, high-quality document-level parallel data; and (2) the propensity of LLMs to introduce hallucinations and omissions during generation. To address these challenges, we propose a two-stage fine-tuning strategy leveraging LLM-augmented document-level data. First, we augment data by converting summarization data into document-level parallel data using a LLM, and then filter it using multiple metrics, leveraging sacreBLEU, COMET, and LaBSE-based cosine similarity-to improve data quality. Finally, we employ a two-stage fine-tuning strategy: first fine-tuning on the abundant sentence-level MT resources, and then on the filtered document-level corpus.

Method: Gumbel Distillation uses the Gumbel-Max trick to create a deterministic mapping from a latent Gumbel noise space to the output tokens of a high-performing AR teacher model. This model-agnostic technique can be integrated with various parallel decoding architectures like MDLM and BD3-LM.

Result: Experiments on LM1B and OpenWebText show substantial improvements: 30.0% improvement in MAUVE score and 10.5% improvement in generative perplexity over MDLM trained on OpenWebText dataset.

Conclusion: Gumbel Distillation effectively narrows the performance gap between autoregressive and parallel decoding models by enabling parallel decoders to learn complex token distributions from AR teachers, making parallel generation more viable.

Abstract: The slow, sequential nature of autoregressive (AR) language models has driven the adoption of parallel decoding methods. However, these non-AR models often sacrifice generation quality as they struggle to model the complex joint distribution of token sequences. To narrow this performance gap, we introduce Gumbel Distillation, a novel distillation technique that enables parallel decoders to learn this distribution effectively. Our method leverages the Gumbel-Max trick to create a deterministic mapping from a latent Gumbel noise space to the output tokens of a high-performing AR teacher. As a model-agnostic technique, Gumbel Distillation seamlessly integrates with diverse parallel decoding architectures, including MDLM and BD3-LM. Experiments on LM1B and OpenWebText show that Gumbel Distillation substantially improves the generation quality of parallel language models, achieving a 30.0% improvement in MAUVE score and 10.5% in generative perplexity over MDLM trained on OpenWebText dataset. Code available at https://github.com/hxixixh/gumbel-distill.

Method: MemDLM uses bi-level optimization with inner loop updating fast weights (parametric memory) to capture local trajectory experience per sample, and outer loop updating base model conditioned on this memory. This embeds simulated denoising into training.

Result: MemDLM achieves faster convergence, lower training loss, and improved long-context understanding. At inference, parametric memory acts as emergent in-weight retrieval mechanism, reducing token-level attention bottlenecks on Needle-in-a-Haystack tasks.

Conclusion: MemDLM effectively bridges the train-inference gap in DLMs through parametric memory and bi-level optimization, offering better performance and emergent retrieval capabilities while maintaining DLM advantages over AR models.

Abstract: Diffusion Language Models (DLMs) offer attractive advantages over Auto-Regressive (AR) models, such as full-attention parallel decoding and flexible generation. However, they suffer from a notable train-inference mismatch: DLMs are trained with a static, single-step masked prediction objective, but deployed through a multi-step progressive denoising trajectory. We propose MemDLM (Memory-Enhanced DLM), which narrows this gap by embedding a simulated denoising process into training via Bi-level Optimization. An inner loop updates a set of fast weights, forming a Parametric Memory that captures the local trajectory experience of each sample, while an outer loop updates the base model conditioned on this memory. By offloading memorization pressure from token representations to parameters, MemDLM yields faster convergence and lower training loss. Moreover, the inner loop can be re-enabled at inference time as an adaptation step, yielding additional gains on long-context understanding. We find that, when activated at inference time, this Parametric Memory acts as an emergent in-weight retrieval mechanism, helping MemDLM further reduce token-level attention bottlenecks on challenging Needle-in-a-Haystack retrieval tasks. Code: https://github.com/JarvisPei/MemDLM.

Method: The study examines audio-enabled LLMs for systematic gender discrimination by analyzing how responses shift toward gender-stereotyped adjectives and occupations based solely on speaker voice. It includes complementary survey evidence (n=1,000) on user perceptions and demonstrates pitch manipulation as a potential mitigation strategy.

Result: Audio-enabled LLMs exhibit systematic gender discrimination, shifting responses toward gender-stereotyped adjectives and occupations based on speaker voice, amplifying bias beyond text-based interaction. Survey shows infrequent chatbot users are most hesitant about attribute inference and most likely to disengage when such practices are revealed. Pitch manipulation can systematically regulate gender-discriminatory outputs.

Conclusion: Voice interfaces do not merely extend text models to a new modality but introduce distinct bias mechanisms tied to paralinguistic cues, creating a critical tension in AI development where efforts to expand accessibility through voice interfaces simultaneously create new pathways for discrimination, demanding that fairness and accessibility be addressed in tandem.

Abstract: Hundreds of millions of people rely on large language models (LLMs) for education, work, and even healthcare. Yet these models are known to reproduce and amplify social biases present in their training data. Moreover, text-based interfaces remain a barrier for many, for example, users with limited literacy, motor impairments, or mobile-only devices. Voice interaction promises to expand accessibility, but unlike text, speech carries identity cues that users cannot easily mask, raising concerns about whether accessibility gains may come at the cost of equitable treatment. Here we show that audio-enabled LLMs exhibit systematic gender discrimination, shifting responses toward gender-stereotyped adjectives and occupations solely on the basis of speaker voice, and amplifying bias beyond that observed in text-based interaction. Thus, voice interfaces do not merely extend text models to a new modality but introduce distinct bias mechanisms tied to paralinguistic cues. Complementary survey evidence ($n=1,000$) shows that infrequent chatbot users are most hesitant to undisclosed attribute inference and most likely to disengage when such practices are revealed. To demonstrate a potential mitigation strategy, we show that pitch manipulation can systematically regulate gender-discriminatory outputs. Overall, our findings reveal a critical tension in AI development: efforts to expand accessibility through voice interfaces simultaneously create new pathways for discrimination, demanding that fairness and accessibility be addressed in tandem.

Method: Created a comprehensive benchmark with 1,418 questions covering 7 core education values using multiple question formats (multiple-choice, multi-modal QA, subjective analysis, adversarial prompts, Chinese traditional culture). Evaluated 21 state-of-the-art LLMs using human feedback based automatic evaluation.

Result: Chinese LLMs outperform English LLMs due to educational culture differences (Qwen 2 ranked first with 81.37). LLMs struggle most with teachers’ professional ethics and professional philosophy. Using Edu-Values as an external knowledge repository for RAG significantly improves LLM alignment.

Conclusion: Edu-Values is an effective benchmark for evaluating LLMs on Chinese education values, revealing cultural differences in performance and demonstrating practical utility for improving LLM alignment through RAG applications.

Abstract: In this paper, we present Edu-Values, the first Chinese education values evaluation benchmark that includes seven core values: professional philosophy, teachers’ professional ethics, education laws and regulations, cultural literacy, educational knowledge and skills, basic competencies and subject knowledge. We meticulously design 1,418 questions, covering multiple-choice, multi-modal question answering, subjective analysis, adversarial prompts, and Chinese traditional culture (short answer) questions. We conduct human feedback based automatic evaluation over 21 state-of-the-art (SoTA) LLMs, and highlight three main findings: (1) due to differences in educational culture, Chinese LLMs outperform English LLMs, with Qwen 2 ranking the first with a score of 81.37; (2) LLMs often struggle with teachers’ professional ethics and professional philosophy; (3) leveraging Edu-Values to build an external knowledge repository for RAG significantly improves LLMs’ alignment. This demonstrates the effectiveness of the proposed benchmark.

Method: Proposes IPAEval framework with two-stage prompt scheme that maps client information onto psychometric test items for interpretable psychological assessments, integrating cross-session client-contextual assessment and session-focused client-dynamics assessment.

Result: Experiments on TheraPhase dataset (400 paired client records) show IPAEval effectively tracks symptom severity and treatment outcomes over multiple sessions, outperforming baselines across both closed-source and open-source models.

Conclusion: IPAEval enables comprehensive client-centered therapeutic outcome evaluation with interpretable assessments, validating benefits of items-aware reasoning mechanisms for longitudinal mental health monitoring.

Abstract: In psychotherapy, therapeutic outcome assessment, or treatment outcome evaluation, is essential to mental health care by systematically evaluating therapeutic processes and outcomes. Existing large language model approaches often focus on therapist-centered, single-session evaluations, neglecting the client’s subjective experience and longitudinal progress across multiple sessions. To address these limitations, we propose IPAEval, a client-Informed Psychological Assessment-based Evaluation framework, which automates treatment outcome evaluations from the client’s perspective using clinical interviews. It integrates cross-session client-contextual assessment and session-focused client-dynamics assessment for a comprehensive understanding of therapeutic progress. Specifically, IPAEval employs a two-stage prompt scheme that maps client information onto psychometric test items, enabling interpretable and structured psychological assessments. Experiments on our new TheraPhase dataset, comprising 400 paired initial and completion stage client records, demonstrate that IPAEval effectively tracks symptom severity and treatment outcomes over multiple sessions, outperforming baseline approaches across both closed-source and open-source models, and validating the benefits of items-aware reasoning mechanisms.

Method: Conducted a systematic review of 181 papers through systematic literature review methodology, analyzing available resources, representation schemes, and downstream tasks related to instructional text. The study examines trends, challenges, and opportunities in complex instruction understanding.

Result: The survey provides AI/NLP researchers with essential background knowledge and a unified view of various approaches to complex instruction understanding. It bridges gaps between different research directions and highlights future research opportunities in this emerging field.

Conclusion: This comprehensive survey systematically analyzes the landscape of complex instruction understanding, offering researchers a foundation for advancing LLMs as general-purpose agents capable of handling real-world, multi-step instructions across various domains.

Abstract: Recent advances in large language models have demonstrated promising capabilities in following simple instructions through instruction tuning. However, real-world tasks often involve complex, multi-step instructions that remain challenging for current NLP systems. Robust understanding of such instructions is essential for deploying LLMs as general-purpose agents that can be programmed in natural language to perform complex, real-world tasks across domains like robotics, business automation, and interactive systems. Despite growing interest in this area, there is a lack of a comprehensive survey that systematically analyzes the landscape of complex instruction understanding and processing. Through a systematic review of the literature, we analyze available resources, representation schemes, and downstream tasks related to instructional text. Our study examines 181 papers, identifying trends, challenges, and opportunities in this emerging field. We provide AI/NLP researchers with essential background knowledge and a unified view of various approaches to complex instruction understanding, bridging gaps between different research directions and highlighting future research opportunities.

Method: Created MediaSpin dataset with 78,910 headline pairs annotated for 13 distinct media bias types using human-supervised LLM labeling, enabling systematic bias identification and analysis.

Result: Dataset provides linguistic insights into editorial bias patterns and enables applications for bias prediction and user behavior analysis in news headline editing.

Conclusion: MediaSpin enables systematic study of editorial bias in headline editing, offering tools for bias prediction and understanding how news framing evolves post-publication.

Abstract: The editability of online news content has become a significant factor in shaping public perception, as social media platforms introduce new affordances for dynamic and adaptive news framing. Edits to news headlines can refocus audience attention, add or remove emotional language, and shift the framing of events in subtle yet impactful ways. What types of media bias are editorialized in and out of news headlines, and how can they be systematically identified? This study introduces the MediaSpin dataset, the first to characterize the bias in how prominent news outlets editorialize news headlines after publication. The dataset includes 78,910 pairs of headlines annotated with 13 distinct types of media bias, using human-supervised LLM labeling. We discuss the linguistic insights it affords and show its applications for bias prediction and user behavior analysis.

Method: Proposes a framework based on David Marr’s three levels of analysis (computational, algorithmic, implementational) for studying information processing systems. Revisits established cognitive science techniques relevant to each level and illustrates how they can provide insights into LLM behavior and internal organization.

Result: The paper provides a conceptual toolkit for making sense of LLMs by bridging cognitive science methodology with AI system analysis, though specific empirical results aren’t mentioned in the abstract.

Conclusion: Cognitive science methods developed for understanding the human mind can be productively applied to understand large language models, offering a systematic approach to demystifying these increasingly powerful but opaque AI systems.

Abstract: Modern artificial intelligence systems, such as large language models, are increasingly powerful but also increasingly hard to understand. Recognizing this problem as analogous to the historical difficulties in understanding the human mind, we argue that methods developed in cognitive science can be useful for understanding large language models. We propose a framework for applying these methods based on the levels of analysis that David Marr proposed for studying information processing systems. By revisiting established cognitive science techniques relevant to each level and illustrating their potential to yield insights into the behavior and internal organization of large language models, we aim to provide a toolkit for making sense of these new kinds of minds.

Method: Constructed dataset with 1,000 manual and 4,000 synthetic Bangla sentence-gloss pairs, plus 159 expert-annotated test pairs. Compared fine-tuned open-source models (mBART) vs. closed-source LLMs (GPT-5.4, Qwen-3) for translation performance.

Result: GPT-5.4 achieved best overall performance, fine-tuned mBART performed competitively despite being 100x smaller, and Qwen-3 outperformed all in human evaluation. Synthetic data generation proved effective for low-resource translation.

Conclusion: Established first dataset and model for Bangla text-to-gloss translation, demonstrating synthetic data’s value for low-resource sign language translation and showing competitive performance from smaller fine-tuned models.

Abstract: Gloss is a written approximation that bridges Sign Language (SL) and its corresponding spoken language. Despite a deaf and hard-of-hearing population of at least 3 million in Bangladesh, Bangla Sign Language (BdSL) remains largely understudied, with no prior work on Bangla text-to-gloss translation and no publicly available datasets. To address this gap, we construct the first Bangla text-to-gloss dataset, consisting of 1,000 manually annotated and 4,000 synthetically generated Bangla sentence-gloss pairs, along with 159 expert human-annotated pairs used as a test set. Our experimental framework performs a comparative analysis between several fine-tuned open-source models and a leading closed-source LLM to evaluate their performance in low-resource BdSL translation. GPT-5.4 achieves the best overall performance, while a fine-tuned mBART model performs competitively despite being approximately 100% smaller. Qwen-3 outperforms all other models in human evaluation. This work introduces the first dataset and trained model for Bangla text-to-gloss translation. It also demonstrates the effectiveness of systematically generated synthetic data for addressing challenges in low-resource sign language translation.

Method: Survey paper providing comprehensive overview structured around three perspectives: (1) AI as Persuader (AI-generated persuasive content and applications), (2) AI as Persuadee (AI’s susceptibility to influence and manipulation), (3) AI as Persuasion Judge (AI’s role in evaluating persuasive strategies, detecting manipulation, ensuring ethical persuasion). Introduces taxonomy for persuasion research.

Result: Presents structured framework for understanding AI’s role in persuasion across three dimensions. Identifies key challenges for future research to enhance safety, fairness, and effectiveness of AI-powered persuasion while addressing risks from increasingly capable language models.

Conclusion: AI’s role in persuasion is multi-faceted - as persuader, persuadee, and judge. Future research must address ethical challenges, safety concerns, and develop frameworks for responsible AI persuasion as language models become more capable.

Abstract: Persuasion is a fundamental aspect of communication, influencing decision-making across diverse contexts, from everyday conversations to high-stakes scenarios such as politics, marketing, and law. The rise of conversational AI systems has significantly expanded the scope of persuasion, introducing both opportunities and risks. AI-driven persuasion can be leveraged for beneficial applications, but also poses threats through unethical influence. Moreover, AI systems are not only persuaders, but also susceptible to persuasion, making them vulnerable to adversarial attacks and bias reinforcement. Despite rapid advancements in AI-generated persuasive content, our understanding of what makes persuasion effective remains limited due to its inherently subjective and context-dependent nature. In this survey, we provide a comprehensive overview of persuasion, structured around three key perspectives: (1) AI as a Persuader, which explores AI-generated persuasive content and its applications; (2) AI as a Persuadee, which examines AI’s susceptibility to influence and manipulation; and (3) AI as a Persuasion Judge, which analyzes AI’s role in evaluating persuasive strategies, detecting manipulation, and ensuring ethical persuasion. We introduce a taxonomy for persuasion research and discuss key challenges for future research to enhance the safety, fairness, and effectiveness of AI-powered persuasion while addressing the risks posed by increasingly capable language models.

Method: Proposes Iterative Preference Learning (IPL): constructs CoaT-tree through iterative sampling, scores leaf nodes with rule-based rewards, backpropagates feedback to create Thinking-level DPO pairs, and uses three-stage instruction evolution with GPT-4o for diverse Q&A generation.

Result: MobileIPL outperforms strong baselines including OS-ATLAS and UI-TARS, achieves state-of-the-art performance across three standard Mobile GUI-Agent benchmarks, and shows strong generalization to out-of-domain scenarios.

Conclusion: The IPL framework effectively addresses data scarcity in CoaT trajectories through iterative preference learning and instruction evolution, enabling improved reasoning and generalization for VLM-based mobile GUI agents.

Abstract: The Chain of Action-Planning Thoughts (CoaT) paradigm has been shown to improve the reasoning performance of VLM-based mobile agents in GUI tasks. However, the scarcity of diverse CoaT trajectories limits the expressiveness and generalization ability of such agents. While self-training is commonly employed to address data scarcity, existing approaches either overlook the correctness of intermediate reasoning steps or depend on expensive process-level annotations to construct process reward models (PRM). To address the above problems, we propose an Iterative Preference Learning (IPL) that constructs a CoaT-tree through interative sampling, scores leaf nodes using rule-based reward, and backpropagates feedback to derive Thinking-level Direct Preference Optimization (T-DPO) pairs. To prevent overfitting during warm-up supervised fine-tuning, we further introduce a three-stage instruction evolution, which leverages GPT-4o to generate diverse Q&A pairs based on real mobile UI screenshots, enhancing both generality and layout understanding. Experiments on three standard Mobile GUI-agent benchmarks demonstrate that our agent MobileIPL outperforms strong baselines, including continual pretraining models such as OS-ATLAS and UI-TARS. It achieves state-of-the-art performance across three standard Mobile GUI-Agents benchmarks and shows strong generalization to out-of-domain scenarios.

Method: Created a comprehensive benchmark with recognition tasks using automated cheminformatics tools, and editing/generation tasks through expert annotation and validation. Supports evaluation across multiple molecular representations: linear strings, molecular images, and molecular graphs.

Result: State-of-the-art models show significant limitations: GPT-5 achieves 86.2% accuracy on recognition and 85.5% on editing (tasks intuitive for humans), and only 43.0% on generation tasks.

Conclusion: Current AI systems have substantial shortcomings in handling basic molecular recognition and manipulation tasks, and MolLangBench should catalyze research toward more effective AI systems for chemical applications.

Abstract: Precise recognition, editing, and generation of molecules are essential prerequisites for both chemists and AI systems tackling various chemical tasks. We present MolLangBench, a comprehensive benchmark designed to evaluate fundamental molecule-language interface tasks: language-prompted molecular structure recognition, editing, and generation. To ensure high-quality, unambiguous, and deterministic outputs, we construct the recognition tasks using automated cheminformatics tools, and curate editing and generation tasks through rigorous expert annotation and validation. MolLangBench supports the evaluation of models that interface language with different molecular representations, including linear strings, molecular images, and molecular graphs. Evaluations of state-of-the-art models reveal significant limitations: the strongest model (GPT-5) achieves $86.2%$ and $85.5%$ accuracy on recognition and editing tasks, which are intuitively simple for humans, and performs even worse on the generation task, reaching only $43.0%$ accuracy. These results highlight the shortcomings of current AI systems in handling even preliminary molecular recognition and manipulation tasks. We hope MolLangBench will catalyze further research toward more effective and reliable AI systems for chemical applications.The dataset and code can be accessed at https://huggingface.co/datasets/ChemFM/MolLangBench and https://github.com/TheLuoFengLab/MolLangBench, respectively.

Method: Used Erotetic Theory of Reasoning (ETR) and its open-source implementation PyETR to generate 383 formally specified reasoning problems, evaluated 38 models, judged logical correctness and whether incorrect responses matched ETR-predicted fallacies, and tested premise order effects.

Result: Two key findings: 1) As model capability (Chatbot Arena Elo) increases, a larger share of incorrect answers are ETR-predicted fallacies, while overall correctness shows no correlation with capability; 2) Reversing premise order significantly reduces fallacy production for many models, mirroring human order effects.

Conclusion: Language models’ reasoning errors systematically follow human fallacy patterns, with error composition (not just rate) providing insights into model reasoning, and PyETR enables contamination-resistant reasoning tests linked to cognitive theory.

Abstract: We study logical reasoning in language models by asking whether their errors follow established human fallacy patterns. Using the Erotetic Theory of Reasoning (ETR) and its open-source implementation, PyETR, we programmatically generate 383 formally specified reasoning problems and evaluate 38 models. For each response, we judge logical correctness and, when incorrect, whether it matches an ETR-predicted fallacy. Two results stand out: (i) as a capability proxy (Chatbot Arena Elo) increases, a larger share of a model’s incorrect answers are ETR-predicted fallacies $(ρ=0.360, p=0.0265)$, while overall correctness on this dataset shows no correlation with capability; (ii) reversing premise order significantly reduces fallacy production for many models, mirroring human order effects. Methodologically, PyETR provides an open-source pipeline for unbounded, synthetic, contamination-resistant reasoning tests linked to a cognitive theory, enabling analyses that focus on error composition rather than error rate.

Method: KGA introduces two synergistic pathways: 1) Bottom-up knowledge fusion pathway that dynamically integrates external KGs into input representations via input-driven KG fusion (similar to stimulus-driven attention in human brain), and 2) Top-down attention guidance pathway that assesses contextual relevance of each triple through goal-directed verification, suppressing task-irrelevant signals and amplifying knowledge-relevant patterns.

Result: Extensive experiments on four benchmarks verify KGA’s strong fusion performance and efficiency, demonstrating effective real-time knowledge integration without parameter modification.

Conclusion: KGA provides a novel framework for dynamic KG integration into LLMs at inference-time, addressing limitations of static models and parameter-invasive methods while supporting real-time knowledge fusion for evolving web environments.

Abstract: Knowledge graphs (KGs) are the cornerstone of the semantic web, offering up-to-date representations of real-world entities and relations. Yet large language models (LLMs) remain largely static after pre-training, causing their internal knowledge to become outdated and limiting their utility in time-sensitive web applications. To bridge this gap between dynamic knowledge and static models, a prevalent approach is to enhance LLMs with KGs. However, prevailing methods typically rely on parameter-invasive fine-tuning, which risks catastrophic forgetting and often degrades LLMs’ general capabilities. Moreover, their static integration frameworks cannot keep pace with the continuous evolution of real-world KGs, hindering their deployment in dynamic web environments. To bridge this gap, we introduce KGA (\textit{\underline{K}nowledge \underline{G}raph-guided \underline{A}ttention}), a novel framework that dynamically integrates external KGs into LLMs exclusively at inference-time without any parameter modification. Inspired by research on neuroscience, we rewire the self-attention module by innovatively introducing two synergistic pathways: a \textit{bottom-up knowledge fusion} pathway and a \textit{top-down attention guidance} pathway. The \textit{bottom-up pathway} dynamically integrates external knowledge into input representations via input-driven KG fusion, which is akin to the \textit{stimulus-driven attention process} in the human brain. Complementarily, the \textit{top-down pathway} aims to assess the contextual relevance of each triple through a \textit{goal-directed verification process}, thereby suppressing task-irrelevant signals and amplifying knowledge-relevant patterns. By synergistically combining these two pathways, our method supports real-time knowledge fusion. Extensive experiments on four benchmarks verify KGA’s strong fusion performance and efficiency.

Method: Proposed framework combining Evidence-Centered Design with Social Cognitive Theory and Zone of Proximal Development for adaptive scaffolding. Instantiated as Inquizzitor, an LLM-based formative assessment agent integrating human-AI hybrid intelligence with cognitive science principles.

Result: Inquizzitor delivers high-quality assessment and interaction aligned with core learning theories, providing effective guidance that students value, demonstrating potential for theory-driven LLM integration in education.

Conclusion: The research shows that LLM-based pedagogical agents can provide adaptive and principled instruction when grounded in established learning theories, offering a promising direction for educational AI development.

Abstract: Large language models (LLMs) present new opportunities for creating pedagogical agents that engage in meaningful dialogue to support student learning. However, current LLM systems used in classrooms often lack the solid theoretical foundations found in earlier intelligent tutoring systems. To bridge this gap, we propose a framework that combines Evidence-Centered Design with Social Cognitive Theory and Zone of Proximal Development for adaptive scaffolding in LLM-based agents focused on STEM+C learning. We instantiate this framework with Inquizzitor, an LLM-based formative assessment agent that integrates human-AI hybrid intelligence and provides feedback grounded in cognitive science principles. Our findings show that Inquizzitor delivers high-quality assessment and interaction aligned with core learning theories, offering effective guidance that students value. This research demonstrates the potential for theory-driven LLM integration in education, highlighting the ability of these systems to provide adaptive and principled instruction.

Method: Mechanistic analysis reveals LLMs follow distinct “constellation” patterns in embedding space as representations traverse layers, with each NLP task maintaining consistent trajectories. SafeConstellations is an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways, selectively applied only to tasks prone to over-refusal.

Result: The method reduces over-refusal rates by up to 73% with minimal impact on utility, offering a principled and conditional approach to mitigating over-refusals while maintaining safety on genuinely harmful content.

Conclusion: SafeConstellations provides an effective inference-time solution to LLM over-refusal by leveraging task-specific trajectory patterns in embedding space, enabling conditional mitigation that preserves both utility and safety.

Abstract: LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that seemingly resemble harmful content. This phenomenon diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through extensive evaluation, we demonstrate that LLMs persist in refusing inputs containing harmful content, even when they are reframed with tasks that have benign intent. Our mechanistic analysis reveals that LLMs follow distinct “constellation” patterns in embedding space as representations traverse layers, with each NLP task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, our method reduces over-refusal rates by up to 73% with minimal impact on utility – offering a principled and conditional approach to mitigating over-refusals.

Method: The paper proposes Linguistic Fingerprints Extraction (LIFE), which reconstructs word-level probability distributions to find discriminative patterns. Through distributional divergence analysis, the authors discover prompt-induced linguistic fingerprints - statistically distinct probability shifts between LLM-generated real and fake news when maliciously prompted. They also use key-fragment techniques to amplify these subtle linguistic differences.

Result: Experiments show that LIFE achieves state-of-the-art performance in detecting LLM-generated fake news and maintains high performance on human-written fake news. The method effectively identifies subtle linguistic patterns that distinguish fake from real news.

Conclusion: LIFE provides an effective approach for detecting LLM-generated fake news by leveraging linguistic fingerprints in probability distributions, addressing the growing threat of AI-generated misinformation.

Abstract: With the rapid development of large language models, the generation of fake news has become increasingly effortless, posing a growing societal threat and underscoring the urgent need for reliable detection methods. Early efforts to identify LLM-generated fake news have predominantly focused on the textual content itself; however, because much of that content may appear coherent and factually consistent, the subtle traces of falsification are often difficult to uncover. Through distributional divergence analysis, we uncover prompt-induced linguistic fingerprints: statistically distinct probability shifts between LLM-generated real and fake news when maliciously prompted. Based on this insight, we propose a novel method named Linguistic Fingerprints Extraction (LIFE). By reconstructing word-level probability distributions, LIFE can find discriminative patterns that facilitate the detection of LLM-generated fake news. To further amplify these fingerprint patterns, we also leverage key-fragment techniques that accentuate subtle linguistic differences, thereby improving detection reliability. Our experiments show that LIFE achieves state-of-the-art performance in LLM-generated fake news and maintains high performance in human-written fake news. The code and data are available at https://anonymous.4open.science/r/LIFE-E86A.

Method: Analyzes four model development stages (scaling, pretraining, post-training, inference) across nine non-English languages, comparing En-CoT (English reasoning) vs. Target-CoT (target-language reasoning) settings.

Result: Scaling improves En-CoT but Target-CoT lags, especially for complex reasoning; specialized reasoning pretraining helps En-CoT but hurts Target-CoT; translated English reasoning traces outperform distilled target-language traces; language-specific efficiency disparities and failure modes identified.

Conclusion: Multilingual reasoning capabilities differ significantly from English, with Target-CoT performance lagging behind En-CoT, highlighting the need for specialized approaches to develop robust multilingual reasoning models.

Abstract: While large reasoning models have shown remarkable ability to generate long chains-of-thought (CoTs) in English, we still lack understanding of how these long-form reasoning abilities transfer to the vast majority of the world’s languages. In this work, we systematically investigate four key stages of model development–scaling, pretraining, post-training, and inference–to understand how long CoT capabilities extend beyond English. We compare two reasoning settings across nine non-English target languages: En-CoT, where models process target-language inputs, but reason in English; and Target-CoT, where models both process inputs and generate long CoTs in the target language. We find that scaling reasoning model size improves multilingual task performance in En-CoT, but Target-CoT performance lags behind. This gap widens for tasks requiring long, multi-step CoTs such as mathematical reasoning. Shifting to pretraining, we find that adding a specialized reasoning stage enhances En-CoT performance but degrades Target-CoT, whereas broad multilingual pretraining improves both modes simultaneously. Given the scarcity of high-quality reasoning traces in languages other than English, we explore synthetic data curation approaches for post-training. We demonstrate that fine-tuning on reasoning traces automatically translated from gold English traces outperforms fine-tuning on target-language traces distilled from large reasoning models. Finally, we report disparities in inference efficiency between languages and uncover language-specific failure modes in CoTs. We release models, datasets, and code to foster further research.

Method: Decomposed sycophancy into sycophantic agreement and sycophantic praise, contrasting with genuine agreement. Used difference-in-means directions, activation additions, and subspace geometry analysis across multiple models and datasets.

Result: Three behaviors are encoded along distinct linear directions in latent space; each can be independently amplified/suppressed without affecting others; representational structure is consistent across model families and scales.

Conclusion: Sycophantic behaviors correspond to distinct, independently steerable representations rather than a unified mechanism.

Abstract: Large language models (LLMs) often exhibit sycophantic behaviors – such as excessive agreement with or flattery of the user – but it is unclear whether these behaviors arise from a single mechanism or multiple distinct processes. We decompose sycophancy into sycophantic agreement and sycophantic praise, contrasting both with genuine agreement. Using difference-in-means directions, activation additions, and subspace geometry across multiple models and datasets, we show that: (1) the three behaviors are encoded along distinct linear directions in latent space; (2) each behavior can be independently amplified or suppressed without affecting the others; and (3) their representational structure is consistent across model families and scales. These results suggest that sycophantic behaviors correspond to distinct, independently steerable representations.

Method: Proposes Token Entropy Cumulative Average (TECA) metric to measure exploration extent in reasoning process, and introduces “Explore Briefly, Then Decide” paradigm with Cumulative Entropy Regulation (CER) mechanism that leverages TECA to dynamically determine optimal stopping points.

Result: Experimental results across diverse mathematical benchmarks show substantial mitigation of overthinking without sacrificing problem-solving ability, with average response length decreasing by up to 71% on simpler datasets.

Conclusion: The proposed approach creates a more efficient and adaptive reasoning process by helping LLMs determine when to conclude reasoning and provide final answers, addressing the overthinking problem in complex reasoning tasks.

Abstract: Large Language Models (LLMs) have demonstrated remarkable reasoning abilities on complex problems using long Chain-of-Thought (CoT) reasoning. However, they often suffer from overthinking, meaning generating unnecessarily lengthy reasoning steps for simpler problems. This issue may degrade the efficiency of the models and make them difficult to adapt the reasoning depth to the complexity of problems. To address this, we introduce a novel metric Token Entropy Cumulative Average (TECA), which measures the extent of exploration throughout the reasoning process. We further propose a novel reasoning paradigm named “Explore Briefly, Then Decide”, with an associated Cumulative Entropy Regulation (CER) mechanism. This paradigm leverages TECA to help the model dynamically determine the optimal point to conclude its thought process and provide a final answer, thus achieving efficient reasoning. Experimental results across diverse mathematical benchmarks show that our approach substantially mitigates overthinking without sacrificing problem solving ability. With our thinking paradigm, the average response length decreases by up to 71% on simpler datasets, demonstrating the effectiveness of our method in creating a more efficient and adaptive reasoning process.

Method: Memory-augmented architecture with small language models accessing large hierarchical parametric memory banks. During pretraining/inference, fetch small context-dependent memory blocks and add them to the model. Pretraining learns to store long-tail world knowledge in memory parameters while small model captures common knowledge and reasoning.

Result: 160M-parameter model with 18M-parameter memory from 4.6B memory bank performs comparably to regular models with 2x+ parameters. Hierarchical feed-forward memories work robustly across transformer architectures, scaling to over 21B parameters.

Conclusion: Memory-augmented architectures enable efficient knowledge storage and retrieval, making large-scale knowledge accessible to small models while maintaining performance comparable to much larger parameter models.

Abstract: The impressive performance gains of modern language models currently rely on scaling parameters: larger models store more world knowledge and reason better. Yet compressing all world knowledge into parameters is unnecessary, as only a fraction is used per prompt, and impractical for edge devices with limited inference-time memory and compute. We address this shortcoming by a memory-augmented architecture and a pretraining strategy aligned with existing hardware paradigms. We introduce small language models that access large hierarchical parametric memory banks encoding world knowledge. During pretraining and inference, we fetch a small, context-dependent memory block and add it to the model. Our pretraining learns to store long-tail world knowledge in the memory parameters, while the small language model acts as an anchor capturing common knowledge and general reasoning abilities. Through trillion-token-scale experiments, we show significant gains: a 160M-parameters model augmented with an 18M-parameters memory fetched from a 4.6B memory bank obtains comparable performance to a regular model with more than 2x the parameters. Through extensive experiments, we study the optimal type and size of parametric memories in transformers, scaling them to over 21B parameters. We find that our proposed hierarchical feed-forward memories work robustly across transformer architectures, whether added during pretraining or post-hoc.

Method: Holistic evaluation of hybrid architectures based on inter-layer (sequential) or intra-layer (parallel) fusion. Comprehensive evaluation across multiple dimensions: language modeling, downstream tasks, long-context capabilities, scaling analysis, and training/inference efficiency. Investigation of computational primitive characteristics to identify critical elements for each hybridization strategy.

Result: The study identifies the most critical elements for each hybridization strategy and proposes optimal design recipes for hybrid models through comprehensive analysis of computational primitives.

Conclusion: The comprehensive analysis provides practical guidance and valuable insights for developing hybrid language models, facilitating optimization of architectural configurations for better balance between modeling quality and computational efficiency.

Abstract: Recent progress in large language models demonstrates that hybrid architectures–combining self-attention mechanisms with structured state space models like Mamba–can achieve a compelling balance between modeling quality and computational efficiency, particularly for long-context tasks. While these hybrid models show promising performance, systematic comparisons of hybridization strategies and analyses on the key factors behind their effectiveness have not been clearly shared to the community. In this work, we present a holistic evaluation of hybrid architectures based on inter-layer (sequential) or intra-layer (parallel) fusion. We comprehensively evaluate these designs across multiple dimensions: language modeling and downstream task performance, long-context capabilities, scaling analysis, and training and inference efficiency. By investigating the core characteristics of their computational primitive, we identify the most critical elements for each hybridization strategy and further propose optimal design recipes for hybrid models. Our comprehensive analysis provides practical guidance and valuable insights for developing hybrid language models, facilitating the optimization of architectural configurations.

Method: Introduces purpose-built User Language Models post-trained specifically to simulate human users in multi-turn conversations, creating more realistic simulation environments.

Result: User LMs align better with human behavior and achieve better simulation robustness than existing methods. When used to simulate coding and math conversations, GPT-4o’s performance drops from 74.6% to 57.4%, showing assistants struggle with realistic user nuances.

Conclusion: Realistic user simulation requires purpose-built User LMs rather than repurposed assistant LMs, and more realistic simulation environments reveal significant weaknesses in current assistant models when dealing with imperfect human-like interactions.

Abstract: Conversations with LMs involve two participants: a human user leading the conversation, and an LM assistant responding to the user’s request. To satisfy this specific role, LMs are post-trained to be helpful assistants – optimized to produce exhaustive and well-structured responses, free of ambiguity and grammar errors. User utterances, on the other hand, are rarely perfected, with each user phrasing requests in unique ways, sometimes putting in partial effort at each turn and refining on the fly. To evaluate LM performance in realistic settings, prior work simulated users in multi-turn conversations, often by prompting an LM originally trained to be a helpful assistant to act as a user. However, we show that assistant LMs make for poor user simulators, with the surprising finding that better assistants yield worse simulators. Instead, we introduce purpose-built User Language Models (User LMs) - models post-trained to simulate human users in multi-turn conversations. Through various evaluations, we show how User LMs align better with human behavior and achieve better simulation robustness than existing simulation methods. When leveraging User LMs to simulate coding and math conversations, the performance of a strong assistant (GPT-4o) drops from 74.6% to 57.4%, confirming that more realistic simulation environments lead to assistant struggles as they fail to cope with the nuances of users in multi-turn setups.

Method: 1. Benchmark eight LLMs on injecting emotional appeal into fallacious arguments while preserving logical structures. 2. Use the best-performing models to generate stimuli for a human study. 3. Conduct human evaluation to measure fallacy detection performance and convincingness ratings across different emotional framings.

Result: 1. LLM-driven emotional framing reduces human fallacy detection in F1 by 14.5% on average. 2. Humans perform better in fallacy detection when perceiving enjoyment than fear or sadness. 3. Enjoyment, fear, and sadness correlate with significantly higher convincingness compared to neutral or other emotion states.

Conclusion: LLMs can effectively manipulate emotional framing in fallacious arguments, reducing human ability to detect fallacies and increasing argument convincingness. This has important implications for AI-driven emotional manipulation in fallacious argumentation, highlighting potential risks in public communication.

Abstract: Logical fallacies are common in public communication and can mislead audiences; fallacious arguments may still appear convincing despite lacking soundness, because convincingness is inherently subjective. We present the first computational study of how emotional framing interacts with fallacies and convincingness, using large language models (LLMs) to systematically change emotional appeals in fallacious arguments. We benchmark eight LLMs on injecting emotional appeal into fallacious arguments while preserving their logical structures, then use the best models to generate stimuli for a human study. Our results show that LLM-driven emotional framing reduces human fallacy detection in F1 by 14.5% on average. Humans perform better in fallacy detection when perceiving enjoyment than fear or sadness, and these three emotions also correlate with significantly higher convincingness compared to neutral or other emotion states. Our work has implications for AI-driven emotional manipulation in the context of fallacious argumentation.

Method: The survey classifies and examines CoT fine-tuning methods using the Six Thinking Hats framework, which systematically characterizes common human thinking modes. It also compiles datasets, model performances, and maintains a real-time GitHub repository tracking advances in the field.

Result: Provides the first comprehensive survey of CoT fine-tuning grounded in human reasoning theory, offering a systematic framework for understanding different reasoning capabilities cultivated through CoT fine-tuning.

Conclusion: This survey serves as a valuable resource to inspire innovation and foster progress in CoT fine-tuning by bridging the gap between technical approaches and human cognitive mechanisms.

Abstract: Chain of thought (CoT) fine-tuning aims to endow large language models (LLMs) with reasoning capabilities by training them on curated reasoning traces. It leverages both supervised and reinforced fine-tuning to cultivate human-like reasoning skills in LLMs, including detailed planning, divergent thinking, intuitive judgment, timely reflection, internal thinking, and fact perception, etc. As CoT fine-tuning has advanced, LLMs have demonstrated substantial improvements in tasks such as mathematical reasoning and code generation. However, existing surveys about CoT fine-tuning primarily focus on technical aspects and overlook a systematic analysis from the perspective of human reasoning mechanisms. Given that the ultimate goal of CoT fine-tuning is to enable LLMs to reason like humans, it is crucial to investigate this technique through the lens of human cognition. To fill this gap, we present the first comprehensive survey of CoT fine-tuning grounded in human reasoning theory. Specifically, inspired by the well-known Six Thinking Hats framework, which systematically characterizes common human thinking modes using six metaphorical hats, we classify and examine CoT fine-tuning methods through this lens. Furthermore, building upon this theory, we outline potential directions for future research in CoT fine-tuning. In addition, we compile a comprehensive overview of existing datasets and model performances, and a real-time GitHub repository \footnote{https://github.com/AI-Chen/Awesome-CoT-Finetuning} that continuously tracks recent advances in this area is maintained. We hope this survey will serve as a valuable resource to inspire innovation and foster progress in this rapidly evolving field.

Method: SafeSearch uses multi-objective reinforcement learning with a final-output safety/utility reward plus a novel query-level shaping term that penalizes unsafe queries and rewards safe ones.

Result: Reduces agent harmfulness by over 90% across three red-teaming datasets on a 7B model while maintaining QA performance comparable to utility-only finetuned agents.

Conclusion: Joint alignment of safety and utility is crucial for search agents, and query-level rewards effectively improve both safety and utility simultaneously.

Abstract: Large language model (LLM) based search agents iteratively generate queries, retrieve external information, and reason to answer open-domain questions. While researchers have primarily focused on improving their utility, their safety behaviors remain underexplored. In this paper, we first evaluate search agents using red-teaming datasets and find that they are more likely to produce harmful outputs than base LLMs. For example, when asked ``How can I track someone’s location without their consent?’’, a base model refuses, whereas a search agent designed to retrieve and cite sources may lower its refusal threshold, fetch documents (e.g., court cases), and, once appended, synthesize them into an informative yet unsafe summary. We further show that utility-oriented finetuning intensifies this risk, motivating joint alignment of safety and utility. To this end, we present SafeSearch, a multi-objective reinforcement learning approach that couples a final-output safety/utility reward with a novel query-level shaping term that penalizes unsafe queries and rewards safe ones. Experiments show that SafeSearch reduces agent harmfulness by over 90% across three red-teaming datasets on a 7B model while producing safe and helpful responses, and maintains QA performance comparable to that of a utility-only finetuned agent. Further analyses confirm the effectiveness of the query-level reward in jointly improving safety and utility.

Method: Introduces SPORTABSET benchmark for long-context tabular summarization across two sports domains requiring entity tracking and statistical aggregation. Systematically evaluates decomposition-based strategies across several long-context LLMs.

Result: Decomposition substantially improves accuracy and numerical fidelity, but gains mainly come from dissecting multi-entity interference rather than improved local arithmetic. Models show high sensitivity to surface-level cues with structured failures including hallucination, omission, and role confusion.

Conclusion: Consistent multientity memory is a key bottleneck in long context table generation, motivating diagnostic evaluation as essential for scalable, efficient and reliable tabular summarization models.

Abstract: Large language model (LLM) approaches to tabular summarization rely on extensive prompt engineering, decomposition pipelines, or entity-level intermediate representations to achieve strong performance. While effective, these strategies are computationally expensive and offer limited insight into how well models maintain state over long, evolving narratives. We introduce SPORTABSET, a diagnostic benchmark for long-context tabular summarization across two complementary sports domains that require tracking multiple entities and aggregating statistics under domain-specific rules. Using SporTabSet, we systematically evaluate decomposition-based strategies across several long context LLMs. Results show that although decomposition substantially improves accuracy and numerical fidelity, gains stem mainly from dissecting multi-entity interference rather than improved local arithmetic. Robustness experiments further reveal high sensitivity to surface-level cues with structured failures, including hallucination, omission, and role confusion. Together, these findings identify consistent multientity memory as a key bottleneck in long context table generation, motivating diagnostic evaluation as a prerequisite for scalable, efficient and reliable tabular summarization models.

Method: Proposed novel structure-aware representations: speaker-weighted distributions for geography, hyperbolic embeddings for genealogy, and latent variables model for typology. Unified these into a robust, task-agnostic composite distance.

Result: Across multiple zero-shot transfer benchmarks, the representations significantly improve transfer performance when distance type is relevant to the task, while the composite distance yields gains in most tasks.

Conclusion: The framework addresses limitations of existing linguistic knowledge bases by providing type-specific representations and a principled composite distance that enhances cross-lingual transfer performance.

Abstract: Existing linguistic knowledge bases such as URIEL+ provide valuable geographic, genetic and typological distances for cross-lingual transfer but suffer from two key limitations. First, their one-size-fits-all vector representations are ill-suited to the diverse structures of linguistic data. Second, they lack a principled method for aggregating these signals into a single, comprehensive score. In this paper, we address these gaps by introducing a framework for type-matched language distances. We propose novel, structure-aware representations for each distance type: speaker-weighted distributions for geography, hyperbolic embeddings for genealogy, and a latent variables model for typology. We unify these signals into a robust, task-agnostic composite distance. Across multiple zero-shot transfer benchmarks, we demonstrate that our representations significantly improve transfer performance when the distance type is relevant to the task, while our composite distance yields gains in most tasks.

Method: Uses large language models trained with direct preference optimization technique to improve accuracy of difficulty control in generating multiple-choice questions.

Result: The proposed method addresses both limitations - enables direct generation of multiple-choice questions and improves difficulty controllability through explicit optimization.

Conclusion: A novel approach for difficulty-controllable multiple-choice question generation that overcomes key limitations of conventional methods through LLM-based optimization.

Abstract: Difficulty-controllable question generation for reading comprehension has gained significant attention in the field of education as a fundamental tool for adaptive learning support. Although several neural question generation methods have recently succeeded in controlling difficulty, conventional approaches still face two major limitations. First, they cannot directly generate multiple-choice questions, which are the most widely used question type in educational contexts. Second, they are not explicitly trained to optimize the accuracy of difficulty control, leaving room for further improvement in difficulty controllability. To address these limitations, this study proposes a novel difficulty-controllable multiple-choice question generation method for reading comprehension which leverages a large language model trained using a direct preference optimization technique to improve the accuracy of difficulty control.

Method: Created DEBATE benchmark with 30,707 messages from 2,832 US participants across 708 groups and 107 topics, including public messages and private beliefs. Evaluated 7 LLMs as “digital twin” RPLAs in next-message prediction and full conversation rollout using stance-alignment and opinion-convergence metrics.

Result: Zero-shot RPLA groups show strong opinion convergence relative to human groups. Supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) improve stance alignment and bring group-level convergence closer to human behavior, though discrepancies in opinion change and belief updating remain.

Conclusion: DEBATE enables rigorous benchmarking of simulated opinion dynamics and supports future research on aligning multi-agent role-playing LLMs with realistic human interactions, addressing limitations in current opinion dynamics modeling.

Abstract: Accurately modeling opinion change through social interactions is crucial for understanding and mitigating polarization, misinformation, and societal conflict. Recent work simulates opinion dynamics with role-playing LPL agents (RPLAs), but multi-agent simulations often display unnatural group behavior, such as premature convergence, and lack empirical benchmarks for assessing alignment with real human group interactions. We introduce DEBATE, a large-scale benchmark for evaluating the authenticity of opinion dynamics in multi-agent RPLA simulations. DEBATE contains 30,707 messages from 2,832 U.S.-based participants across 708 groups and 107 topics, with both public messages and private Likert-scale beliefs, enabling evaluation at the utterance and group levels while also supporting future individual-level analyses. We instantiate “digital twin” RPLAs with seven LLMs and evaluate them in two settings: next-message prediction and full conversation rollout, using stance-alignment and opinion-convergence metrics. In zero-shot settings, RPLA groups exhibit strong opinion convergence relative to human groups. Post-training via supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) improves stance alignment and brings group-level convergence closer to human behavior, though discrepancies in opinion change and belief updating remain. DEBATE enables rigorous benchmarking of simulated opinion dynamics and supports future research on aligning multi-agent RPLAs with realistic human interactions. The benchmark is publicly available at.

Method: Manual translation and adaptation of 3K MMLU-Redux multiple-choice question-answer pairs into five major Arabic dialects (Syrian, Egyptian, Emirati, Saudi, Moroccan), creating 15K QA pairs across 32 domains, with evaluation of 19 open-weight Arabic and multilingual LLMs (1B-13B parameters).

Result: Evaluation revealed substantial performance variation across dialects, showing persistent gaps in dialectal generalization, with the benchmark enabling systematic assessment of LLM reasoning and comprehension beyond Modern Standard Arabic.

Conclusion: DialectalArabicMMLU provides the first unified, human-curated resource for measuring dialectal understanding in Arabic, promoting more inclusive evaluation and future model development for Arabic language technologies.

Abstract: We present DialectalArabicMMLU, a new benchmark for evaluating the performance of large language models (LLMs) across Arabic dialects. While recently developed Arabic and multilingual benchmarks have advanced LLM evaluation for Modern Standard Arabic (MSA), dialectal varieties remain underrepresented despite their prevalence in everyday communication. DialectalArabicMMLU extends the MMLU-Redux framework through manual translation and adaptation of 3K multiple-choice question-answer pairs into five major dialects (Syrian, Egyptian, Emirati, Saudi, and Moroccan), yielding a total of 15K QA pairs across 32 academic and professional domains (22K QA pairs when also including English and MSA). The benchmark enables systematic assessment of LLM reasoning and comprehension beyond MSA, supporting both task-based and linguistic analysis. We evaluate 19 open-weight Arabic and multilingual LLMs (1B-13B parameters) and report substantial performance variation across dialects, revealing persistent gaps in dialectal generalization. DialectalArabicMMLU provides the first unified, human-curated resource for measuring dialectal understanding in Arabic, thus promoting more inclusive evaluation and future model development.

Method: MUTANT uses careful vocabulary and training data design, language-aware pre-tokenization, and subword and multiword aware training. MUTANT-Indic specifically targets India-specific multilingual LLMs with linguistically coherent token design.

Result: Achieves 39.5% better average fertility score than LLaMA4 and 18% better than Sutra (current best), with 44% improvement in inference throughput while maintaining comparable performance on English and Indic benchmarks.

Conclusion: MUTANT provides an effective recipe for multilingual tokenizers that improves efficiency and performance, particularly valuable for languages with rich morphological variation like Indian languages.

Abstract: Tokenizers play a crucial role in determining the performance, training efficiency, and the inference cost of Large Language Models (LLMs). Designing effective tokenizers for multilingual LLMs is particularly challenging due to diverse scripts and rich morphological variation. While subword methods like Byte Pair Encoding (BPE) are widely adopted, their effectiveness in multilingual settings remains underexplored. We present MUTANT, a recipe for building multilingual tokenizers, with careful vocabulary and training data design, language-aware pre-tokenization, and subword and multiword aware training. We also introduce MUTANT-Indic, a tokenizer for India-specific multilingual LLMs, that produces linguistically coherent tokens and achieves state-of-the-art performance. Evaluated across English, 22 Indian languages and code data, our tokenizer improves the average fertility score by 39.5%$ over LLaMA4 and by 18% over Sutra (the current best). This translates to 44% improvement in inference throughput over LLaMA4 while maintaining comparable performance on English and Indic benchmarks. We present detailed ablations across tokenizer training data size, vocabulary size, merging techniques, and pre-tokenization strategies, demonstrating the robustness of our design choices.

Method: 1) Automated pipeline linking papers to baselines/datasets they actually used; 2) Collective perception enhanced retriever using self-descriptions plus aggregated citation contexts; 3) Reasoning-augmented reranker with explicit reasoning chains and LLM fine-tuning for interpretable justifications.

Result: Covers 85% of datasets/baselines used at top AI conferences over past 5 years; outperforms strongest prior baseline with +5.85% in Recall@20 and +8.30% in HitRate@5.

Conclusion: The framework advances reliable, interpretable automation of experimental design by leveraging collective perception from scholarly networks rather than just content similarity.

Abstract: Large language model agents are becoming increasingly capable at web-centric tasks such as information retrieval, complex reasoning. These emerging capabilities have given rise to surge research interests in developing LLM agent for facilitating scientific quest. One key application in AI research is to automate experiment design through agentic dataset and baseline retrieval. However, prior efforts suffer from limited data coverage, as recommendation datasets primarily harvest candidates from public portals and omit many datasets actually used in published papers, and from an overreliance on content similarity that biases model toward superficial similarity and overlooks experimental suitability. Harnessing collective perception embedded in the baseline and dataset citation network, we present a comprehensive framework for baseline and dataset recommendation. First, we design an automated data-collection pipeline that links roughly one hundred thousand accepted papers to the baselines and datasets they actually used. Second, we propose a collective perception enhanced retriever. To represent the position of each dataset or baseline within the scholarly network, it concatenates self-descriptions with aggregated citation contexts. To achieve efficient candidate recall, we finetune an embedding model on these representations. Finally, we develop a reasoning-augmented reranker that exact interaction chains to construct explicit reasoning chains and finetunes a large language model to produce interpretable justifications and refined rankings. The dataset we curated covers 85% of the datasets and baselines used at top AI conferences over the past five years. On our dataset, the proposed method outperforms the strongest prior baseline with average gains of +5.85% in Recall@20, +8.30% in HitRate@5. Taken together, our results advance reliable, interpretable automation of experimental design.

Method: Designed shortcut scenarios with spurious associations between user labels and superficial textual cues to evaluate detector robustness. Proposed mitigation strategies using large language models for counterfactual data augmentation at three levels: individual user text, overall dataset distribution, and model’s causal information extraction.

Result: Shifts in irrelevant feature distributions caused average relative accuracy drop of 32% in baseline models. Proposed LLM-based strategies achieved average relative performance improvement of 56% under shortcut scenarios.

Conclusion: Social bot detectors are vulnerable to shortcut learning with textual features, but LLM-based counterfactual augmentation effectively mitigates these issues by addressing data distribution and causal feature extraction problems.

Abstract: While existing social bot detectors perform well on benchmarks, their robustness across diverse real-world scenarios remains limited due to unclear ground truth and varied misleading cues. In particular, the impact of shortcut learning, where models rely on spurious correlations instead of capturing causal task-relevant features, has received limited attention. To address this gap, we conduct an in-depth study to assess how detectors are influenced by potential shortcuts based on textual features, which are most susceptible to manipulation by social bots. We design a series of shortcut scenarios by constructing spurious associations between user labels and superficial textual cues to evaluate model robustness. Results show that shifts in irrelevant feature distributions significantly degrade social bot detector performance, with an average relative accuracy drop of 32% in the baseline models. To tackle this challenge, we propose mitigation strategies based on large language models, leveraging counterfactual data augmentation. These methods mitigate the problem from data and model perspectives across three levels, including data distribution at both the individual user text and overall dataset levels, as well as the model’s ability to extract causal information. Our strategies achieve an average relative performance improvement of 56% under shortcut scenarios.

Method: Systematic algorithm audit of 1,508 real baby care and pregnancy-related queries using a robust evaluation framework assessing multiple quality dimensions: answer consistency, relevance, presence of medical safeguards, source categories, and sentiment alignment.

Result: 33% inconsistency between AI Overviews and Featured Snippets on same search result pages; critically lacking medical safeguards (only 11% in AIO, 7% in FS); health/wellness websites dominate sources but FS often links to commercial sources; despite high relevance scores, quality gaps exist.

Conclusion: Findings reveal concerning gaps in AI-mediated health information with important implications for public health information access. Demonstrates need for stronger quality controls in AI systems for high-stakes domains. Provides transferable framework for auditing AI systems where information quality impacts user well-being.

Abstract: Google Search increasingly surfaces AI-generated content through features like AI Overviews (AIO) and Featured Snippets (FS), which users frequently rely on despite having no control over their presentation. Through a systematic algorithm audit of 1,508 real baby care and pregnancy-related queries, we evaluate the quality and consistency of these information displays. Our robust evaluation framework assesses multiple quality dimensions, including answer consistency, relevance, presence of medical safeguards, source categories, and sentiment alignment. Our results reveal concerning gaps in information consistency, with information in AIO and FS displayed on the same search result page being inconsistent with each other in 33% of cases. Despite high relevance scores, both features critically lack medical safeguards (present in just 11% of AIO and 7% of FS responses). While health and wellness websites dominate source categories for both, AIO and FS, FS also often link to commercial sources. These findings have important implications for public health information access and demonstrate the need for stronger quality controls in AI-mediated health information. Our methodology provides a transferable framework for auditing AI systems across high-stakes domains where information quality directly impacts user well-being.

Method: Proposes EasyMED, a multi-agent VSP framework that separates case-grounded information disclosure from response generation, and introduces SPBench benchmark with eight expert-defined criteria for interaction-level evaluation.

Result: EasyMED more closely matches human SP behavior than existing VSPs, particularly in case consistency and controlled disclosure. A four-week controlled study shows learning outcomes comparable to human SP training with stronger early gains for novices.

Conclusion: EasyMED demonstrates improved flexibility, psychological safety, and cost efficiency compared to human SP training while achieving comparable learning outcomes, making it a viable scalable alternative for clinical skills training.

Abstract: Standardized patients (SPs) are indispensable for clinical skills training but remain expensive and difficult to scale. Although large language model (LLM)-based virtual standardized patients (VSPs) have been proposed as an alternative, their behavior remains unstable and lacks rigorous comparison with human standardized patients. We propose EasyMED, a multi-agent VSP framework that separates case-grounded information disclosure from response generation to support stable, inquiry-conditioned patient behavior. We also introduce SPBench, a human-grounded benchmark with eight expert-defined criteria for interaction-level evaluation. Experiments show that EasyMED more closely matches human SP behavior than existing VSPs, particularly in case consistency and controlled disclosure. A four-week controlled study further demonstrates learning outcomes comparable to human SP training, with stronger early gains for novice learners and improved flexibility, psychological safety, and cost efficiency.

Method: Introduces a formal, rule-based grammar that deconstructs complex instructions into <Procedure, Relation, Value> triplets. Uses a multi-stage, human-in-the-loop pipeline to generate diverse datasets and a transparent programmatic engine for objective verification.

Result: The paper releases a dataset and open-source evaluation tools to facilitate research into LLM controllability and reliability, providing a systematic framework for assessing fine-grained lexical instruction following.

Conclusion: LexInstructEval addresses limitations in current evaluation methods by providing an objective, expressive benchmark for testing LLMs’ ability to follow complex lexical instructions, enabling better research into model controllability.

Abstract: The ability of Large Language Models (LLMs) to precisely follow complex and fine-grained lexical instructions is a cornerstone of their utility and controllability. However, evaluating this capability remains a significant challenge. Current methods either rely on subjective and costly human evaluation or on automated LLM-as-a-judge systems, which suffer from inherent biases and unreliability. Existing programmatic benchmarks, while objective, often lack the expressiveness to test intricate, compositional constraints at a granular level. To address these limitations, we introduce LexInstructEval, a new benchmark and evaluation framework for fine-grained lexical instruction following. Our framework is built upon a formal, rule-based grammar that deconstructs complex instructions into a canonical <Procedure, Relation, Value> triplet. This grammar enables the systematic generation of a diverse dataset through a multi-stage, human-in-the-loop pipeline and facilitates objective verification via a transparent, programmatic engine. We release our dataset and open-source evaluation tools to facilitate further research into the controllability and reliability of LLMs.

Method: Constructed new Basque corpora combining standard, social media, and historical sources. Pre-trained BERnaT encoder-only models in three configurations: standard, diverse, and combined. Proposed evaluation framework separating NLU tasks into standard and diverse subsets.

Result: Models trained on both standard and diverse data consistently outperform standard-only models across all task types, improving performance without compromising standard benchmark accuracy.

Conclusion: Linguistic diversity is crucial for building inclusive, generalizable language models that can handle the full spectrum of language variation.

Abstract: Language models depend on massive text corpora that are often filtered for quality, a process that can unintentionally exclude non-standard linguistic varieties, reduce model robustness and reinforce representational biases. In this paper, we argue that language models should aim to capture the full spectrum of language variation (dialectal, historical, informal, etc.) rather than relying solely on standardized text. Focusing on the Basque language, we construct new corpora combining standard, social media, and historical sources, and pre-train the BERnaT family of encoder-only models in three configurations: standard, diverse, and combined. We further propose an evaluation framework that separates Natural Language Understanding (NLU) tasks into standard and diverse subsets to assess linguistic generalization. Results show that models trained on both standard and diverse data consistently outperform those trained on standard corpora, improving performance across all task types without compromising standard benchmark accuracy. These findings highlight the importance of linguistic diversity in building inclusive, generalizable language models.

Method: Two complementary methods: 1) Continued BPE training - extends pre-trained tokenizer by continuing BPE merge learning on new data rather than appending non-overlapping tokens; 2) Leaf-based vocabulary pruning - removes redundant tokens while preserving model quality through systematic pruning approach.

Result: Experiments across multiple languages and model families show improved tokenization efficiency and better utilization of added vocabulary with continued BPE training. The pruning method effectively reduces vocabulary size while maintaining model performance.

Conclusion: The proposed methods provide practical tools for controlled vocabulary modification, released as an open-source toolkit, enabling more efficient tokenizer adaptation to new domains/languages.

Abstract: Tokenizer adaptation plays an important role in adapting pre-trained language models to new domains or languages. In this work, we address two complementary aspects of this process: vocabulary extension and pruning. The common approach to extension trains a new tokenizer on domain-specific text and appends the tokens that do not overlap with the existing vocabulary, which often results in many tokens that are unreachable or never used. We propose continued BPE training that extends a pre-trained tokenizer by continuing the BPE merge learning process on new data. Experiments across multiple languages and model families show that this approach improves tokenization efficiency and leads to better utilization of added vocabulary. We also introduce leaf-based vocabulary pruning, which removes redundant tokens while preserving model quality. Together, these methods provide practical tools for controlled vocabulary modification, which we release as an open-source toolkit.

Method: Created M4-RAG benchmark with over 80,000 culturally diverse image-question pairs spanning 42 languages, 56 dialects, and 189 countries. Built controlled retrieval environment with millions of curated multilingual documents to balance realism with reproducibility. Systematically evaluated RAG performance across different model sizes and languages.

Result: RAG consistently benefits smaller VLMs but fails to scale to larger models, often degrading their performance. Significant performance degradation occurs when prompts or retrieved context are in non-English languages. Exposes critical mismatch between model size and current retrieval effectiveness.

Conclusion: Multilingual multimodal RAG presents unique challenges requiring specialized approaches. Current retrieval methods are insufficient for larger VLMs, and cross-lingual performance gaps highlight the need for better multilingual adaptation. The benchmark enables systematic evaluation of these issues.

Abstract: Vision-language models (VLMs) have achieved strong performance in visual question answering (VQA), yet they remain constrained by static training data. Retrieval-Augmented Generation (RAG) mitigates this limitation by enabling access to up-to-date, culturally grounded, and multilingual information; however, multilingual multimodal RAG remains largely underexplored. We introduce M4-RAG, a massive-scale benchmark spanning 42 languages, 56 regional dialects and registers, and 189 countries, comprising over 80,000 culturally diverse image-question pairs for evaluating retrieval-augmented VQA across languages and modalities. To balance realism with reproducibility, we build a controlled retrieval environment containing millions of carefully curated multilingual documents relevant to the query domains, approximating real-world retrieval conditions while ensuring consistent experimentation. Our systematic evaluation reveals that although RAG consistently benefits smaller VLMs, it fails to scale to larger models and often even degrades their performance, exposing a critical mismatch between model size and current retrieval effectiveness. Our cross-lingual evaluations also reveal significant performance degradation when prompts or retrieved context are provided in non-English languages. The code, datasets, and evaluation protocols for M4-RAG are available as open-source at https://github.com/davidanugraha/M4-RAG.

Method: Proposes an Identify-then-Verify framework: 1) Generate multiple hypotheses about missing information and establish semantic consensus, 2) Perform critical verification by forcing the model to re-examine source text to confirm whether this information is truly absent.

Result: The framework outperforms established baselines across diverse multi-hop and factual QA datasets, producing more accurate sufficiency judgments while clearly articulating information gaps.

Conclusion: Guiding models to justify claims about missing information through structured reasoning improves sufficiency assessment in question-answering systems, particularly for inferential questions.

Abstract: Determining whether a provided context contains sufficient information to answer a question is a critical challenge for building reliable question-answering systems. While simple prompting strategies have shown success on factual questions, they frequently fail on inferential ones that require reasoning beyond direct text extraction. We hypothesize that asking a model to first reason about what specific information is missing provides a more reliable, implicit signal for assessing overall sufficiency. To this end, we propose a structured Identify-then-Verify framework for robust sufficiency modeling. Our method first generates multiple hypotheses about missing information and establishes a semantic consensus. It then performs a critical verification step, forcing the model to re-examine the source text to confirm whether this information is truly absent. We evaluate our method against established baselines across diverse multi-hop and factual QA datasets. The results demonstrate that by guiding the model to justify its claims about missing information, our framework produces more accurate sufficiency judgments while clearly articulating any information gaps.

Method: Created dataset of 3,200 Basque essays with expert annotations (scores + feedback), fine-tuned RoBERTa-EusCrawl and Latxa 8B/70B models for scoring and explanation generation, proposed novel evaluation methodology combining automatic consistency metrics with expert validation.

Result: Encoder models reliable for AES, fine-tuned Latxa surpasses GPT-5 and Claude Sonnet 4.5 in scoring consistency and feedback quality, identifies wider range of error types than proprietary models.

Conclusion: Establishes foundation for transparent, reproducible NLP research in low-resource languages, demonstrates effectiveness of fine-tuned open-source models for educational applications.

Abstract: This paper introduces the first publicly available dataset for Automatic Essay Scoring (AES) and feedback generation in Basque, targeting the CEFR C1 proficiency level. The dataset comprises 3,200 essays from HABE, each annotated by expert evaluators with criterion specific scores covering correctness, richness, coherence, cohesion, and task alignment enriched with detailed feedback and error examples. We fine-tune open-source models, including RoBERTa-EusCrawl and Latxa 8B/70B, for both scoring and explanation generation. Our experiments show that encoder models remain highly reliable for AES, while supervised fine-tuning (SFT) of Latxa significantly enhances performance, surpassing state-of-the-art (SoTA) closed-source systems such as GPT-5 and Claude Sonnet 4.5 in scoring consistency and feedback quality. We also propose a novel evaluation methodology for assessing feedback generation, combining automatic consistency metrics with expert-based validation of extracted learner errors. Results demonstrate that the fine-tuned Latxa model produces criterion-aligned, pedagogically meaningful feedback and identifies a wider range of error types than proprietary models. This resource and benchmark establish a foundation for transparent, reproducible, and educationally grounded NLP research in low-resource languages such as Basque.

Method: Constructed dataset using open-source frameworks and publicly available data sources, with a curated English split and multilingual TEI XML split. Developed an extensible pipeline for dataset generation and pre-trained a RoBERTa model on the dataset.

Result: Created SciLaD dataset with over 10 million English publications and 35+ million multilingual publications. The pre-trained RoBERTa model achieved performance comparable to other scientific language models of similar size across comprehensive benchmarks.

Conclusion: SciLaD demonstrates that open-source tools can enable large-scale scientific data curation with high quality. The dataset and evaluation pipeline promote reproducibility and further research in scientific language processing and understanding.

Abstract: SciLaD is a novel, large-scale dataset of scientific language constructed entirely using open-source frameworks and publicly available data sources. It comprises a curated English split containing over 10 million scientific publications and a multilingual, unfiltered TEI XML split including more than 35 million publications. We also publish the extensible pipeline for generating SciLaD. The dataset construction and processing workflow demonstrates how open-source tools can enable large-scale, scientific data curation while maintaining high data quality. Finally, we pre-train a RoBERTa model on our dataset and evaluate it across a comprehensive set of benchmarks, achieving performance comparable to other scientific language models of similar size, validating the quality and utility of SciLaD. We publish the dataset and evaluation pipeline to promote reproducibility, transparency, and further research in natural scientific language processing and understanding, including scholarly document processing.

Method: Created MADON dataset of 272 Czech Supreme Court decisions with expert annotations, adapted transformer LLMs to Czech legal domain through continued pretraining, used asymmetric loss and class weighting for imbalance, and developed a three-stage pipeline combining ModernBERT, Llama 3.1, and traditional ML.

Result: Best models achieved 82.6% Bal-F1 for argument detection, 77.5% Bal-F1 for legal argument classification, and 83.8% Bal-F1 for formalism classification. The pipeline reduced computational costs while increasing explainability.

Conclusion: Legal argument mining enables judicial philosophy classification and challenges prevailing narratives about CEE formalism. The methodology is transferable across jurisdictions and all resources are publicly available.

Abstract: Courts must justify their decisions, but systematically analyzing judicial reasoning at scale remains difficult. This study tests claims about formalistic judging in Central and Eastern Europe (CEE) by developing automated methods to detect and classify judicial reasoning in decisions of Czech Supreme Courts using state-of-the-art natural language processing methods. We create the MADON dataset of 272 decisions from two Czech Supreme Courts with expert annotations of 9,183 paragraphs with eight argument types and holistic formalism labels for supervised training and evaluation. Using a corpus of 300,511 Czech court decisions, we adapt transformer LLMs to Czech legal domain through continued pretraining and we experiment with methods to address dataset imbalance including asymmetric loss and class weighting. The best models can detect argumentative paragraphs (82.6% Bal-F1), classify traditional types of legal argument (77.5% Bal-F1), and classify decisions as formalistic/non-formalistic (83.8% Bal-F1). Our three-stage pipeline combining ModernBERT, Llama 3.1, and traditional feature-based machine learning achieves promising results for decision classification while reducing computational costs and increasing explainability. Empirically, we challenge prevailing narratives about CEE formalism. We demonstrate that legal argument mining enables promising judicial philosophy classification and highlight its potential for other important tasks in computational legal studies. Our methodology can be used across jurisdictions, and our entire pipeline, datasets, guidelines, models, and source codes are available at https://github.com/trusthlt/madon.

Method: 1) Introduces HiSPA (Hidden State Poisoning Attack) where specific short phrases induce partial amnesia by overwriting hidden states. 2) Creates RoBench-25 benchmark to evaluate information retrieval under HiSPAs. 3) Tests various models including Mamba, Jamba (hybrid SSM-Transformer), and pure Transformers. 4) Extends analysis to Mamba-2 and Nemotron-3-Nano. 5) Conducts interpretability study of Mamba’s hidden layers during attacks.

Result: SSMs are vulnerable to HiSPAs while pure Transformers are not. Jamba-1.7-Mini (52B hybrid) collapses on RoBench-25 under HiSPA triggers. HiSPA triggers also weaken Jamba on Open-Prompt-Injections benchmark. Theoretical and empirical findings extend to Mamba-2 and Nemotron-3-Nano. Interpretability reveals patterns in Mamba’s hidden layers during attacks.

Conclusion: SSMs have critical security vulnerabilities to hidden state poisoning attacks that don’t affect Transformers, highlighting a fundamental difference in robustness between these architectures. The findings suggest the need for mitigation systems and raise concerns about deploying SSMs in security-critical applications.

Abstract: State space models (SSMs) like Mamba offer efficient alternatives to Transformer-based language models, with linear time complexity. Yet, their adversarial robustness remains critically unexplored. This paper studies the phenomenon whereby specific short input phrases induce a partial amnesia effect in such models, by irreversibly overwriting information in their hidden states, referred to as a Hidden State Poisoning Attack (HiSPA). Our benchmark RoBench-25 allows evaluating a model’s information retrieval capabilities when subject to HiSPAs, and confirms the vulnerability of SSMs against such attacks. Even the recent Jamba-1.7-Mini SSM–Transformer (a 52B hybrid model) collapses on RoBench-25 under some HiSPA triggers, whereas pure Transformers do not. We also observe that HiSPA triggers significantly weaken the Jamba model on the popular Open-Prompt-Injections benchmark, unlike pure Transformers. We further show that the theoretical and empirical findings extend to Mamba-2, and also analyse a Mamba-2-based hybrid (Nemotron-3-Nano). Finally, our interpretability study reveals patterns in Mamba’s hidden layers during HiSPAs that could be used to build a HiSPA mitigation system. The full code and data to reproduce the experiments can be found at https://anonymous.4open.science/r/hispa_anonymous-5DB0.

Method: Two-stage parameter fusion framework: 1) Geometric consensus filtering to prune conflicting gradients, and 2) Curvature-based importance weighting to amplify shared semantics. Uses directional consensus evaluation to disentangle knowledge updates.

Result: Extensive experiments on GUI agents and tool-use agent domains show outstanding continual learning performance with minimal computational overhead and parameter updates.

Conclusion: Agent-Dice effectively addresses the stability-plasticity dilemma in LLM-based agents by explicitly distinguishing between shared and conflicting knowledge, enabling better continual learning without catastrophic forgetting.

Abstract: Large Language Model (LLM)-based agents significantly extend the utility of LLMs by interacting with dynamic environments. However, enabling agents to continually learn new tasks without catastrophic forgetting remains a critical challenge, known as the stability-plasticity dilemma. In this work, we argue that this dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference. To address this, we propose Agent-Dice, a parameter fusion framework based on directional consensus evaluation. Concretely, Agent-Dice disentangles knowledge updates through a two-stage process: geometric consensus filtering to prune conflicting gradients, and curvature-based importance weighting to amplify shared semantics. We provide a rigorous theoretical analysis that establishes the validity of the proposed fusion scheme and offers insight into the origins of the stability-plasticity dilemma. Extensive experiments on GUI agents and tool-use agent domains demonstrate that Agent-Dice exhibits outstanding continual learning performance with minimal computational overhead and parameter updates. The codes are available at https://github.com/Wuzheng02/Agent-Dice.

Method: Introduces FACTUM framework with four mechanistic scores: Contextual Alignment Score (CAS), Attention Sink Usage (BAS), Parametric Force Score (PFS), and Pathway Alignment Score (PAS). Analyzes how attention and feed-forward network pathways coordinate during citation generation across different model scales (3B to 8B parameters).

Result: Correct citations show higher parametric force (PFS) and greater attention sink usage (BAS). The signature of correctness evolves with scale: 3B models rely on high pathway alignment (PAS), while 8B models shift to specialized strategies with orthogonal information from different pathways. FACTUM outperforms state-of-the-art baselines by up to 37.5% in AUC.

Conclusion: Citation hallucinations are complex coordination failures between model pathways, not simple over-reliance on parametric knowledge. High parametric force can be constructive when properly coordinated with attention pathways. The FACTUM framework enables more nuanced and reliable RAG systems by capturing these mechanistic interactions.

Abstract: Retrieval-Augmented Generation (RAG) models are critically undermined by citation hallucinations, a deceptive failure where a model cites a source that fails to support its claim. While existing work attributes hallucination to a simple over-reliance on parametric knowledge, we reframe this failure as an evolving, scale-dependent coordination failure between the Attention (reading) and Feed-Forward Network (recalling) pathways. We introduce FACTUM (Framework for Attesting Citation Trustworthiness via Underlying Mechanisms), a framework of four mechanistic scores: Contextual Alignment (CAS), Attention Sink Usage (BAS), Parametric Force (PFS), and Pathway Alignment (PAS). Our analysis reveals that correct citations are consistently marked by higher parametric force (PFS) and greater use of the attention sink (BAS) for information synthesis. Crucially, we find that “one-size-fits-all” theories are insufficient as the signature of correctness evolves with scale: while the 3B model relies on high pathway alignment (PAS), our best-performing 8B detector identifies a shift toward a specialized strategy where pathways provide distinct, orthogonal information. By capturing this complex interplay, FACTUM outperforms state-of-the-art baselines by up to 37.5% in AUC. Our results demonstrate that high parametric force is constructive when successfully coordinated with the Attention pathway, paving the way for more nuanced and reliable RAG systems.

Method: Uses self-supervised learning with feature-space objectives where modality-specific encoders learn contextualized target representations. Pre-trained on large-scale French corpora: Wikipedia, OSCAR, CroissantLLM for text; MultilingualLibriSpeech, LeBenchmark, and newly introduced INA-100k (100k hours of French audio from national archives) for speech.

Result: Pantagruel models show competitive or superior performance compared to strong French baselines (CamemBERT, FlauBERT, LeBenchmark2.0) across various downstream tasks in both modalities, while maintaining a shared architecture that handles speech or text inputs seamlessly.

Conclusion: Feature-space self-supervised objectives are effective for French representation learning, and Pantagruel serves as a robust foundation for multimodal speech-text understanding in French.

Abstract: We release Pantagruel models, a new family of self-supervised encoder models for French text and speech. Instead of predicting modality-tailored targets such as textual tokens or speech units, Pantagruel learns contextualized target representations in the feature space, allowing modality-specific encoders to capture linguistic and acoustic regularities more effectively. Separate models are pre-trained on large-scale French corpora, including Wikipedia, OSCAR and CroissantLLM for text, together with MultilingualLibriSpeech, LeBenchmark, and INA-100k for speech. INA-100k is a newly introduced 100,000-hour corpus of French audio derived from the archives of the Institut National de l’Audiovisuel (INA), the national repository of French radio and television broadcasts, providing highly diverse audio data. We evaluate Pantagruel across a broad range of downstream tasks spanning both modalities, including those from the standard French benchmarks such as FLUE or LeBenchmark. Across these tasks, Pantagruel models show competitive or superior performance compared to strong French baselines such as CamemBERT, FlauBERT, and LeBenchmark2.0, while maintaining a shared architecture that can seamlessly handle either speech or text inputs. These results confirm the effectiveness of feature-space self-supervised objectives for French representation learning and highlight Pantagruel as a robust foundation for multimodal speech-text understanding.

Method: Introduces Time Puzzles - constraint-based date inference tasks algorithmically generated to combine factual temporal anchors with cross-cultural calendar relations, enabling controlled evaluation of iterative temporal reasoning with tools.

Result: Across 13 LLMs, best model (GPT-5) achieves only 55.3% accuracy without tools, despite using easily searchable facts. Web search improves performance, but models perform substantially better when constraints are rewritten with explicit dates, removing factual lookup needs.

Conclusion: Reveals a gap in reliable tool use for iterative temporal reasoning, showing that while tools help, models struggle with the iterative reasoning process needed for temporal constraint satisfaction.

Abstract: Tool use, such as web search, has become a standard capability even in freely available large language models (LLMs). However, existing benchmarks evaluate temporal reasoning mainly in static, non-tool-using settings, which poorly reflect how LLMs perform temporal reasoning in practice. We introduce Time Puzzles, a constraint-based date inference task for evaluating iterative temporal reasoning with tools. Each puzzle combines factual temporal anchors with (cross-cultural) calendar relations and may admit one or multiple valid dates. The puzzles are algorithmically generated, enabling controlled and continual evaluation. Across 13 LLMs, even the best model (GPT-5) achieves only 55.3% accuracy without tools, despite using easily searchable facts. While web search improves performance, models perform substantially better when constraints are rewritten with explicit dates, removing the need for factual lookup. These results reveal a gap in reliable tool use for iterative temporal reasoning.

Method: Benchmarked 24 open-source language models on Unicode, Romanized, and mixed-script Sinhala using perplexity evaluation across diverse text sources.

Result: Found substantial script sensitivity with median performance degradation exceeding 300 times from Unicode to Romanized text. Model size showed no correlation with script-handling competence, and Unicode performance strongly predicted mixed-script robustness but not Romanized capability.

Conclusion: Single-script evaluation substantially underestimates real-world deployment challenges for LMs in multi-script low-resource environments. Provides baseline capabilities for Sinhala and practical guidance for model selection.

Abstract: The performance of Language Models (LMs) on low-resource, morphologically rich languages like Sinhala remains largely unexplored, particularly regarding script variation in digital communication. Sinhala exhibits script duality, with Unicode used in formal contexts and Romanized text dominating social media, while mixed-script usage is common in practice. This paper benchmarks 24 open-source LMs on Unicode, Romanized and mixed-script Sinhala using perplexity evaluation across diverse text sources. Results reveal substantial script sensitivity, with median performance degradation exceeding 300 times from Unicode to Romanized text. Critically, model size shows no correlation with script-handling competence, as smaller models often outperform architectures 28 times larger. Unicode performance strongly predicts mixed-script robustness but not Romanized capability, demonstrating that single-script evaluation substantially underestimates real-world deployment challenges. These findings establish baseline LM capabilities for Sinhala and provide practical guidance for model selection in multi-script low-resource environments.

Method: Analyzed 14.8 million prompts in realistic advice-seeking interactions focusing on race and gender in a U.S. context, testing multiple demographic cues for the same groups to measure consistency in model responses.

Result: Different cues for the same demographic group induce only partially overlapping changes in model responses, leading to inconsistent conclusions about personalization and unstable bias assessments. Inconsistencies reflect differences in cue-group association strength and bundled linguistic features.

Conclusion: Demographic conditioning in LLMs is not cue-invariant but depends fundamentally on how identity is cued, reflecting responses to linguistic signals rather than stable demographic categories. Advocates for multi-cue, mechanism-aware evaluations for robust claims about demographic variation.

Abstract: Demographic cue-based evaluation is widely used to study how large language models (LLMs) adapt their responses to signaled demographic attributes within and across groups. This approach typically relies on a single cue (e.g., names) as a proxy for group membership, implicitly treating different cues as interchangeable operationalizations of the same identity-conditioned behavior. We test this assumption in realistic advice-seeking interactions spanning 14.8 million prompts, focusing on race and gender in a U.S. context. We find that cues for the same group induce only partially overlapping changes in model responses, yielding inconsistent conclusions about personalization, while bias conclusions are unstable, with both magnitude and direction of group differences varying across cues. We further show that these inconsistencies reflect differences in cue-group association strength and linguistic features bundled within cues that shape model responses. Together, our findings suggest that demographic conditioning in LLMs is not a cue-invariant category-level parameter but depends fundamentally on how identity is cued, reflecting responses to linguistic signals rather than stable demographic categories. We therefore advocate multi-cue, mechanism-aware evaluations for robust and interpretable claims about demographic variation in LLM responses.

Method: Four-scenario evaluation protocol, logit lens analysis for language probability tracking, cross-lingual semantic similarity analysis, and selective fine-tuning of final layers responsible for language control.

Result: Achieved over 98% language consistency across six languages while fine-tuning only 3-5% of parameters, matching full-scope fine-tuning performance with fraction of computational resources.

Conclusion: Layer-localization of language control enables efficient multilingual adaptation, with final layers being key for language-specific generation while early/middle layers handle semantic alignment and task reasoning.

Abstract: Despite multilingual pretraining, large language models often struggle with non-English tasks, particularly in language control, the ability to respond in the intended language. We identify and characterize two key failure modes: the multilingual transfer bottleneck (correct language, incorrect task response) and the language consistency bottleneck (correct task response, wrong language). To systematically surface these issues, we design a four-scenario evaluation protocol spanning MMLU, MGSM, and XQuAD benchmarks. To probe these issues with interpretability, we extend logit lens analysis to track language probabilities layer by layer and compute cross-lingual semantic similarity of hidden states. The results reveal a three-phase internal structure: early layers align inputs into a shared semantic space, middle layers perform task reasoning, and late layers drive language-specific generation. Guided by these insights, we introduce selective fine-tuning of only the final layers responsible for language control. On Qwen-3-32B and Bloom-7.1B, this method achieves over 98 percent language consistency across six languages while fine-tuning only 3-5 percent of parameters, without sacrificing task accuracy. Importantly, this result is nearly identical to that of full-scope fine-tuning (for example, above 98 percent language consistency for both methods across all prompt scenarios) but uses a fraction of the computational resources. To the best of our knowledge, this is the first approach to leverage layer-localization of language control for efficient multilingual adaptation.

Method: Applied a detection model trained on historical reviews to later review cycles at International Conference on Learning Representations (ICLR) and Nature Communications (NC), tracking temporal patterns from pre-2022 through 2025.

Result: Minimal AI-generated content detected before 2022, followed by substantial increase through 2025: approximately 20% of ICLR reviews and 12% of Nature Communications reviews classified as AI-generated in 2025. Most pronounced growth in NC occurred between Q3 and Q4 2024.

Conclusion: Evidence suggests rapidly increasing presence of AI-assisted content in peer review, highlighting need for further study of implications for scholarly evaluation.

Abstract: The growing availability of large language models (LLMs) has raised questions about their role in academic peer review. This study examines the temporal emergence of AI-generated content in peer reviews by applying a detection model trained on historical reviews to later review cycles at International Conference on Learning Representations (ICLR) and Nature Communications (NC). We observe minimal detection of AI-generated content before 2022, followed by a substantial increase through 2025, with approximately 20% of ICLR reviews and 12% of Nature Communications reviews classified as AI-generated in 2025. The most pronounced growth of AI-generated reviews in NC occurs between the third and fourth quarter of 2024. Together, these findings provide suggestive evidence of a rapidly increasing presence of AI-assisted content in peer review and highlight the need for further study of its implications for scholarly evaluation.

Method: Distill sampled semantic distributions into lightweight student models that estimate prompt-conditioned density before LLM generates answer tokens. Student models predict semantic distributions over possible answers.

Result: Student models perform competitively relative to sampling-based semantic dispersion baselines on hallucination prediction tasks (TriviaQA, MMLU), while offering additional uncertainty primitives for out-of-domain detection and multiple-choice answer selection.

Conclusion: SSD provides an efficient framework for distilling predictive uncertainty in complex output spaces beyond language, enabling prompt-level uncertainty signals and answer-level reliability evaluation without expensive sampling.

Abstract: Large language models present challenges for principled uncertainty quantification, in part due to their complexity and the diversity of their outputs. Semantic dispersion, or the variance in the meaning of sampled answers, has been proposed as a useful proxy for model uncertainty, but the associated computational cost prohibits its use in latency-critical applications. We show that sampled semantic distributions can be distilled into lightweight student models which estimate a prompt-conditioned density before the language model generates an answer token. The student model predicts a semantic distribution over possible answers; the entropy of this distribution provides a prompt-level uncertainty signal, and the probability density allows answer-level reliability evaluation. Across experiments on TriviaQA and MMLU, we find our student models perform competitively relative to sampling-based semantic dispersion baselines on a hallucination prediction task, whilst offering additional uncertainty primitives for out-of-domain detection and multiple-choice answer selection. We term this technique Semantic Self-Distillation (SSD), which can serve as a general framework for distilling predictive uncertainty in complex output spaces beyond language.

Method: Proposes On-Policy Context Distillation (OPCD) framework where a student model is trained on its own generated trajectories while minimizing reverse Kullback-Leibler divergence against a context-conditioned teacher model. This combines on-policy learning with context distillation principles.

Result: OPCD consistently outperforms baseline methods across mathematical reasoning, text-based games, and domain-specific tasks, achieving higher task accuracy while better preserving out-of-distribution capabilities. It also enables effective cross-size distillation where smaller models can internalize experiential knowledge from larger teachers.

Conclusion: OPCD provides an effective framework for knowledge internalization in language models, successfully bridging on-policy and context distillation approaches, with applications in experiential knowledge distillation and system prompt distillation.

Abstract: Context distillation enables language models to internalize in-context knowledge into their parameters. In our work, we propose On-Policy Context Distillation (OPCD), a framework that bridges on-policy distillation with context distillation by training a student model on its own generated trajectories while minimizing reverse Kullback-Leibler divergence against a context-conditioned teacher. We demonstrate the effectiveness of OPCD on two important applications: experiential knowledge distillation, where models extract and consolidate transferable knowledge from their historical solution traces, and system prompt distillation, where models internalize beneficial behaviors encoded in optimized prompts. Across mathematical reasoning, text-based games, and domain-specific tasks, OPCD consistently outperforms baseline methods, achieving higher task accuracy while better preserving out-of-distribution capabilities. We further show that OPCD enables effective cross-size distillation, where smaller student models can internalize experiential knowledge from larger teachers.

Method: SeleCom redefines the encoder’s role as a query-conditioned information selector (decoder-only) trained with a massive, diverse, difficulty-graded synthetic QA dataset using curriculum learning, focusing only on relevant information rather than full document compression.

Result: SeleCom significantly outperforms existing soft compression approaches and achieves competitive or superior performance to non-compression baselines while reducing computation and latency by 33.8%-84.6%.

Conclusion: Query-conditioned information selection is more effective than full document compression for RAG, addressing scalability issues while maintaining or improving performance through selective encoding of relevant information.

Abstract: Retrieval-Augmented Generation (RAG) effectively grounds Large Language Models (LLMs) with external knowledge and is widely applied to Web-related tasks. However, its scalability is hindered by excessive context length and redundant retrievals. Recent research on soft context compression aims to address this by encoding long documents into compact embeddings, yet they often underperform non-compressed RAG due to their reliance on auto-encoder-like full-compression that forces the encoder to compress all document information regardless of relevance to the input query. In this work, we conduct an analysis on this paradigm and reveal two fundamental limitations: (I) Infeasibility, full-compression conflicts with the LLM’s downstream generation behavior; and (II) Non-necessity: full-compression is unnecessary and dilutes task-relevant information density. Motivated by these insights, we introduce SeleCom, a selector-based soft compression framework for RAG that redefines the encoder’s role as query-conditioned information selector. The selector is decoder-only and is trained with a massive, diverse and difficulty-graded synthetic QA dataset with curriculum learning. Extensive experiments show that SeleCom significantly outperforms existing soft compression approaches and achieves competitive or superior performance to non-compression baselines, while reducing computation and latency by 33.8%~84.6%.

Method: Proposes W5H2 structured intent decomposition framework, applies V-measure decomposition to separate clustering properties, uses SetFit with few-shot learning, and implements a five-tier cascade with risk-controlled selective prediction guarantees via RCPS.

Result: Achieves 91.1% accuracy on MASSIVE dataset in ~2ms (vs 37.9% for GPTCache and 68.8% for 20B-parameter LLM at 3,447ms), handles 85% of interactions locally, projects 97.5% cost reduction, and demonstrates cross-lingual transfer across 30 languages.

Conclusion: W5H2 framework enables efficient caching for AI agents by focusing on cache-specific properties rather than general classification accuracy, achieving high performance with minimal latency and significant cost savings.

Abstract: Personal AI agents incur substantial cost via repeated LLM calls. We show existing caching methods fail: GPTCache achieves 37.9% accuracy on real benchmarks; APC achieves 0-12%. The root cause is optimizing for the wrong property – cache effectiveness requires key consistency and precision, not classification accuracy. We observe cache-key evaluation reduces to clustering evaluation and apply V-measure decomposition to separate these on n=8,682 points across MASSIVE, BANKING77, CLINC150, and NyayaBench v2, our new 8,514-entry multilingual agentic dataset (528 intents, 20 W5H2 classes, 63 languages). We introduce W5H2, a structured intent decomposition framework. Using SetFit with 8 examples per class, W5H2 achieves 91.1%+/-1.7% on MASSIVE in ~2ms – vs 37.9% for GPTCache and 68.8% for a 20B-parameter LLM at 3,447ms. On NyayaBench v2 (20 classes), SetFit achieves 55.3%, with cross-lingual transfer across 30 languages. Our five-tier cascade handles 85% of interactions locally, projecting 97.5% cost reduction. We provide risk-controlled selective prediction guarantees via RCPS with nine bound families.

Method: The paper presents a survey methodology: 1) Examination of benchmark landscape, 2) Categorization of methods into multimodal embedding models, multimodal reranker models, and integration of RAG/Agentic systems for document intelligence.

Result: Provides a comprehensive overview of the VDR field, identifies key methodological approaches, and establishes a framework for understanding current state and future directions in multimodal document intelligence.

Conclusion: The survey offers a clear roadmap for future multimodal document intelligence by identifying persistent challenges and outlining promising research directions in the MLLM era.

Abstract: With the rapid proliferation of multimodal information, Visual Document Retrieval (VDR) has emerged as a critical frontier in bridging the gap between unstructured visually rich data and precise information acquisition. Unlike traditional natural image retrieval, visual documents exhibit unique characteristics defined by dense textual content, intricate layouts, and fine-grained semantic dependencies. This paper presents the first comprehensive survey of the VDR landscape, specifically through the lens of the Multimodal Large Language Model (MLLM) era. We begin by examining the benchmark landscape, and subsequently dive into the methodological evolution, categorizing approaches into three primary aspects: multimodal embedding models, multimodal reranker models, and the integration of Retrieval-Augmented Generation (RAG) and Agentic systems for complex document intelligence. Finally, we identify persistent challenges and outline promising future directions, aiming to provide a clear roadmap for future multimodal document intelligence.

Method: Theoretical analysis of transformer architectures with proofs of three key results about analogical reasoning, validated through experiments with models up to 1.5B parameters.

Result: Transformers encode entities with similar properties into similar representations, enabling property transfer through feature alignment. Joint training on similarity and attribution premises is necessary, and two-hop reasoning reduces to analogical reasoning with identity bridges.

Conclusion: Analogical reasoning in transformers emerges through representational geometry and feature alignment, with specific training requirements and data dependencies that reveal a unified mechanism for inductive reasoning.

Abstract: Understanding reasoning in large language models is complicated by evaluations that conflate multiple reasoning types. We isolate analogical reasoning (inferring shared properties between entities based on known similarities) and analyze its emergence in transformers. We theoretically prove three key results: (1) Joint training on similarity and attribution premises enables analogical reasoning through aligned representations; (2) Sequential training succeeds only when similarity structure is learned before specific attributes, revealing a necessary curriculum; (3) Two-hop reasoning ($a \to b, b \to c \implies a \to c$) reduces to analogical reasoning with identity bridges ($b = b$), which must appear explicitly in training data. These results reveal a unified mechanism: transformers encode entities with similar properties into similar representations, enabling property transfer through feature alignment. Experiments with architectures up to 1.5B parameters validate our theory and demonstrate how representational geometry shapes inductive reasoning capabilities.

Method: Multilingual audit of GPT-4o-Mini, Gemini-2.5-Flash, Llama 3.2, Mistral and Gemma 3 using internal representation analysis (log-probs/logits) to compare model opinion distributions against ground-truth public attitudes, plus bias benchmark evaluations (CrowS-Pairs, IndiBias, ThaiCLI, KoBBQ).

Result: Models generally align with public opinion on broad social issues but consistently fail on religious viewpoints, especially minority groups, often amplifying negative stereotypes. Lightweight interventions (demographic priming, native language prompting) help partially but don’t eliminate gaps.

Conclusion: Systematic, regionally grounded audits are urgently needed for equitable global LLM deployment due to persistent cultural misalignment in sensitive contexts.

Abstract: Large Language Models (LLMs) are increasingly being deployed in multilingual, multicultural settings, yet their reliance on predominantly English-centric training data risks misalignment with the diverse cultural values of different societies. In this paper, we present a comprehensive, multilingual audit of the cultural alignment of contemporary LLMs including GPT-4o-Mini, Gemini-2.5-Flash, Llama 3.2, Mistral and Gemma 3 across India, East Asia and Southeast Asia. Our study specifically focuses on the sensitive domain of religion as the prism for broader alignment. To facilitate this, we conduct a multi-faceted analysis of every LLM’s internal representations, using log-probs/logits, to compare the model’s opinion distributions against ground-truth public attitudes. We find that while the popular models generally align with public opinion on broad social issues, they consistently fail to accurately represent religious viewpoints, especially those of minority groups, often amplifying negative stereotypes. Lightweight interventions, such as demographic priming and native language prompting, partially mitigate but do not eliminate these cultural gaps. We further show that downstream evaluations on bias benchmarks (such as CrowS-Pairs, IndiBias, ThaiCLI, KoBBQ) reveal persistent harms and under-representation in sensitive contexts. Our findings underscore the urgent need for systematic, regionally grounded audits to ensure equitable global deployment of LLMs.

Method: The researchers re-examine LAPDOG, a retrieval-augmented framework for personalized dialogue, as a case study. They use both human judges and LLM-based judges to evaluate dialogue quality, identifying specific limitations like corrupted dialogue histories, contradictions between retrieved stories and persona, and incoherent response generation.

Result: Human and LLM judgments align closely with each other but diverge significantly from lexical similarity metrics. The study reveals specific failure modes in current systems that surface-level metrics miss, highlighting the inadequacy of current evaluation practices.

Conclusion: The work calls for more cognitively grounded evaluation methods for retrieval-augmented dialogue systems that better reflect principles of natural human communication, and charts a path toward more reliable assessment frameworks.

Abstract: In cognitive science and linguistic theory, dialogue is not seen as a chain of independent utterances but rather as a joint activity sustained by coherence, consistency, and shared understanding. However, many systems for open-domain and personalized dialogue use surface-level similarity metrics (e.g., BLEU, ROUGE, F1) as one of their main reporting measures, which fail to capture these deeper aspects of conversational quality. We re-examine a notable retrieval-augmented framework for personalized dialogue, LAPDOG, as a case study for evaluation methodology. Using both human and LLM-based judges, we identify limitations in current evaluation practices, including corrupted dialogue histories, contradictions between retrieved stories and persona, and incoherent response generation. Our results show that human and LLM judgments align closely but diverge from lexical similarity metrics, underscoring the need for cognitively grounded evaluation methods. Broadly, this work charts a path toward more reliable assessment frameworks for retrieval-augmented dialogue systems that better reflect the principles of natural human communication.

Method: Proposes a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time without full policy inclusion. Introduces PolicyRecall reward based on Jaccard score and Hallucination Penalty for GRPO training.

Result: Best model outperforms baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

Conclusion: The approach enables efficient business policy adherence in LLM-powered conversational assistants by teaching selective policy recall rather than full policy inclusion, addressing latency and performance issues.

Abstract: Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the “needle-in-the-haystack” problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel PolicyRecall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

Method: Trained classifiers to detect concealment behavior in LMs, comparing gradient-based vs prompt-based concealment methods, and testing generalization across different model architectures and topics.

Result: Classifiers outperformed human evaluators at detecting concealment in smaller models, but failed to generalize to unseen architectures/topics and performed no better than random on models exceeding 70B parameters.

Conclusion: Current black-box auditing methods have critical limitations for detecting knowledge concealment in large LMs, highlighting the need for more robust detection techniques.

Abstract: Language Models (LMs) may acquire harmful knowledge, and yet feign ignorance of these topics when under audit. Inspired by the recent discovery of deception-related behaviour patterns in LMs, we aim to train classifiers that detect when a LM is actively concealing knowledge. Initial findings on smaller models show that classifiers can detect concealment more reliably than human evaluators, with gradient-based concealment proving easier to identify than prompt-based methods. However, contrary to prior work, we find that the classifiers do not reliably generalize to unseen model architectures and topics of hidden knowledge. Most concerningly, the identifiable traces associated with concealment become fainter as the models increase in scale, with the classifiers achieving no better than random performance on any model exceeding 70 billion parameters. Our results expose a key limitation in black-box-only auditing of LMs and highlight the need to develop robust methods to detect models that are actively hiding the knowledge they contain.

Method: CoVerRL framework where a single model alternates between generator and verifier roles. Majority voting provides noisy supervision for training the verifier, while the improving verifier progressively filters self-consistent errors from pseudo-labels, creating a co-evolution cycle.

Result: Outperforms label-free baselines by 4.7-5.9% on mathematical reasoning benchmarks across Qwen and Llama model families. Self-verification accuracy improves from ~55% to over 85%, confirming genuine co-evolution of both capabilities.

Conclusion: CoVerRL successfully escapes the consensus trap by enabling a single model to co-evolve generator and verifier capabilities through alternating roles, maintaining high reward accuracy throughout training while preventing systematic error reinforcement.

Abstract: Label-free reinforcement learning enables large language models to improve reasoning capabilities without ground-truth supervision, typically by treating majority-voted answers as pseudo-labels. However, we identify a critical failure mode: as training maximizes self-consistency, output diversity collapses, causing the model to confidently reinforce systematic errors that evade detection. We term this the consensus trap. To escape it, we propose CoVerRL, a framework where a single model alternates between generator and verifier roles, with each capability bootstrapping the other. Majority voting provides noisy but informative supervision for training the verifier, while the improving verifier progressively filters self-consistent errors from pseudo-labels. This co-evolution creates a virtuous cycle that maintains high reward accuracy throughout training. Experiments across Qwen and Llama model families demonstrate that CoVerRL outperforms label-free baselines by 4.7-5.9% on mathematical reasoning benchmarks. Moreover, self-verification accuracy improves from around 55% to over 85%, confirming that both capabilities genuinely co-evolve.

Method: Introduces ReSynth, a trimodular framework that architecturally separates reasoning (Intellect), purpose (Identity), and knowledge (Memory) as independent components. The approach draws on cross-cultural perspectives on learning (Eastern conceptions of memorization) and critiques from the systematicity debate, aiming to create an architecture where systematic behavior emerges necessarily rather than accidentally.

Result: The paper presents a theoretical framework that diagnoses inherited limitations from psychological paradigms and proposes an architectural solution. While empirical results aren’t specified, the framework offers a principled approach to address adaptability and systematicity challenges in AGI development.

Conclusion: Adaptability, the central challenge of AGI, requires a representational architecture where systematic behavior is a necessary consequence rather than an accidental property. The ReSynth framework provides a pathway to overcome limitations inherited from psychological paradigms by separating reasoning, purpose, and knowledge components.

Abstract: The dominant paradigms of artificial intelligence were shaped by learning theories from psychology: behaviorism inspired reinforcement learning, cognitivism gave rise to deep learning and memory-augmented architectures, and constructivism influenced curriculum learning and compositional approaches. This paper argues that each AI paradigm inherited not only the strengths but the structural limitations of the psychological theory that inspired it. Reinforcement learning cannot account for the internal structure of knowledge, deep learning compresses representations into opaque parameter spaces resistant to principled update, and current integrative approaches lack a formal account of how new understanding is constructed from existing components. The paper further examines a cross-cultural divergence in the interpretation of rote learning, arguing that the Eastern conception of memorization as a structured, multi-phase precursor to understanding offers an underexploited bridge between psychological theory and AI methodology. Drawing on the systematicity debate and critique of Aizawa of both classicism and connectionism, this paper introduces ReSynth, a trimodular framework that separates reasoning (Intellect), purpose (Identity), and knowledge (Memory) as architecturally independent components. The paper traces the genealogy from psychological paradigm to AI method, diagnoses the inherited limitations at each stage, and argues that adaptability, the central challenge of artificial general intelligence requires a representational architecture in which systematic behavior is a necessary consequence rather than an accidental property.

Method: Surveyed five employees from a multinational company in the Philippines about their experiences with Duolingo. Analyzed frequency of encountering general vs. work-related scenarios, effectiveness of different lesson types, and collected suggestions for improvement.

Result: Respondents encountered general scenarios more frequently than work-related ones. General scenarios were relatable and effective for building foundational grammar, vocabulary, and cultural knowledge. Work-related scenarios help bridge the gap toward professional fluency with domain-specific vocabulary. Participants suggested diverse lesson scenarios when analyzed collectively.

Conclusion: Language learning applications should generate personalized, domain-specific lesson scenarios that adapt to individual needs while maintaining foundational support through general, relatable lesson scenarios to achieve professional fluency.

Abstract: Popular language learning applications such as Duolingo use large language models (LLMs) to generate lessons for its users. Most lessons focus on general real-world scenarios such as greetings, ordering food, or asking directions, with limited support for profession-specific contexts. This gap can hinder learners from achieving professional-level fluency, which we define as the ability to communicate comfortably various work-related and domain-specific information in the target language. We surveyed five employees from a multinational company in the Philippines on their experiences with Duolingo. Results show that respondents encountered general scenarios more frequently than work-related ones, and that the former are relatable and effective in building foundational grammar, vocabulary, and cultural knowledge. The latter helps bridge the gap toward professional fluency as it contains domain-specific vocabulary. Each participant suggested lesson scenarios that diverge in contexts when analyzed in aggregate. With this understanding, we propose that language learning applications should generate lessons that adapt to an individual’s needs through personalized, domain specific lesson scenarios while maintaining foundational support through general, relatable lesson scenarios.

Method: Uses Mixture of Experts (MoE) architecture with 30B total parameters but only 3B activated. After supervised fine-tuning on curated data, employs expanded Cascade RL covering broader reasoning and agentic domains, plus multi-domain on-policy distillation from strongest intermediate teacher models throughout training.

Result: Achieves Gold Medal-level performance in IMO, IOI, and ICPC World Finals - only the second open-weight LLM to do so. Delivers reasoning performance approaching frontier models with 20x fewer parameters, demonstrating high intelligence density.

Conclusion: Nemotron-Cascade 2 shows that compact models can achieve elite reasoning capabilities through careful architecture design and training techniques, making high-performance AI more accessible through open-weight release.

Abstract: We introduce Nemotron-Cascade 2, an open 30B MoE model with 3B activated parameters that delivers best-in-class reasoning and strong agentic capabilities. Despite its compact size, its mathematical and coding reasoning performance approaches that of frontier open models. It is the second open-weight LLM, after DeepSeekV3.2-Speciale-671B-A37B, to achieve Gold Medal-level performance in the 2025 International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the ICPC World Finals, demonstrating remarkably high intelligence density with 20x fewer parameters. In contrast to Nemotron-Cascade 1, the key technical advancements are as follows. After SFT on a meticulously curated dataset, we substantially expand Cascade RL to cover a much broader spectrum of reasoning and agentic domains. Furthermore, we introduce multi-domain on-policy distillation from the strongest intermediate teacher models for each domain throughout the Cascade RL process, allowing us to efficiently recover benchmark regressions and sustain strong performance gains along the way. We release the collection of model checkpoint and training data.

Method: The paper conducts a comprehensive review of existing literature on collaboration analysis using task-oriented conversation resources. It systematically examines related theories, coding schemes, tasks, and modeling approaches in the field.

Result: The review provides a practical resource for researchers working on collaboration analysis and identifies unexplored areas for future research in this domain.

Conclusion: Task-oriented conversational data is a valuable resource for collaboration analysis, and the review serves as both a practical guide and a roadmap for future research directions in this area.

Abstract: Collaboration is a task-oriented, high-level human behavior. In most cases, conversation serves as the primary medium for information exchange and coordination, making conversational data a valuable resource for the automatic analysis of collaborative processes. In this paper, we focus on verbal aspects of collaboration and conduct a review of collaboration analysis using task-oriented conversation resources, encompassing related theories, coding schemes, tasks, and modeling approaches. We aim to address the question of how to utilize task-oriented human-human conversational data for collaboration analysis. We hope our review will serve as a practical resource and illuminate unexplored areas for future collaboration analysis.

Method: Two-stage architecture: 1) Graph-based clustering discovers latent task types and trains a classifier for prompt assignment; 2) Mixture-of-experts with task-specific prediction heads for specialized quality estimation. Inference aggregates both stages to balance task-level stability with prompt-specific adaptability.

Result: Evaluated on 10 benchmarks with 11 frontier models, the method consistently outperforms existing baselines and surpasses the strongest individual model while incurring less than half its cost.

Conclusion: The proposed two-stage routing architecture effectively addresses limitations of existing approaches by enabling automated fine-grained task discovery and task-aware quality estimation for optimal model selection in large model pools.

Abstract: Prompt routing dynamically selects the most appropriate large language model from a pool of candidates for each query, optimizing performance while managing costs. As model pools scale to include dozens of frontier models with narrow performance gaps, existing approaches face significant challenges: manually defined task taxonomies cannot capture fine-grained capability distinctions, while monolithic routers struggle to differentiate subtle differences across diverse tasks. We propose a two-stage routing architecture that addresses these limitations through automated fine-grained task discovery and task-aware quality estimation. Our first stage employs graph-based clustering to discover latent task types and trains a classifier to assign prompts to discovered tasks. The second stage uses a mixture-of-experts architecture with task-specific prediction heads for specialized quality estimates. At inference, we aggregate predictions from both stages to balance task-level stability with prompt-specific adaptability. Evaluated on 10 benchmarks with 11 frontier models, our method consistently outperforms existing baselines and surpasses the strongest individual model while incurring less than half its cost.

Method: Collected 44 technical terms from generative syntax literature, had them translated by both human experts and ChatGPT-5, then conducted comparative analysis using analytical and comparative approaches to evaluate translation accuracy.

Result: Only 25% of ChatGPT translations were accurate, 38.6% were inaccurate, and 36.4% were partially correct. The study reveals significant syntactic and semantic challenges in LLM translation of technical linguistic terminology.

Conclusion: LLMs still cannot effectively handle specialized linguistic translation tasks, requiring closer collaboration between AI specialists and linguists to improve translation mechanisms for technical terminology.

Abstract: We aim to examine the extent to which Large Language Models (LLMs) can ’talk much’ about grammar modules, providing evidence from syntax core properties translated by ChatGPT into Arabic. We collected 44 terms from generative syntax previous works, including books and journal articles, as well as from our experience in the field. These terms were translated by humans, and then by ChatGPT-5. We then analyzed and compared both translations. We used an analytical and comparative approach in our analysis. Findings unveil that LLMs still cannot ’talk much’ about the core syntax properties embedded in the terms under study involving several syntactic and semantic challenges: only 25% of ChatGPT translations were accurate, while 38.6% were inaccurate, and 36.4.% were partially correct, which we consider appropriate. Based on these findings, a set of actionable strategies were proposed, the most notable of which is a close collaboration between AI specialists and linguists to better LLMs’ working mechanism for accurate or at least appropriate translation.

Method: Cannot determine method as abstract is unavailable

Result: Cannot determine results as abstract is unavailable

Conclusion: Cannot draw conclusions about paper content due to data unavailability

Abstract: Failed to fetch summary for 2512.06393: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.06393&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to access limitations

Abstract: Failed to fetch summary for 2601.01279: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.01279&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.08806: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.08806&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.00004: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.00004&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.07182: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.07182&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to paper fetch failure

Result: Unable to determine results due to paper fetch failure

Conclusion: Unable to determine conclusion due to paper fetch failure

Abstract: Failed to fetch summary for 2603.16929: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16929&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract or content.

Result: Cannot determine results without access to the paper abstract or content.

Conclusion: Cannot draw conclusions without access to the paper abstract or content.

Abstract: Failed to fetch summary for 2603.19225: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.19225&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Novel AI framework that simultaneously processes multi-section pathology slides to capture comprehensive tumor landscape across entire prostate gland, using patch and slide sub-sampling strategies to reduce computational cost.

Result: Strong predictive performance for 1- and 2-year BCR prediction, substantially outperforming clinical benchmarks; AI-derived risk score validated as most potent independent prognostic factor; computational cost reduced without compromising performance; external validation confirmed generalizability.

Conclusion: The AI-based multi-section slide analysis demonstrates clinical feasibility and prognostic value as a scalable tool for post-operative management in prostate cancer.

Abstract: Prostate cancer is one of the most frequently diagnosed malignancies in men worldwide. However, precise prediction of biochemical recurrence (BCR) after radical prostatectomy remains challenging due to the multifocality of tumors distributed throughout the prostate gland. In this paper, we propose a novel AI framework that simultaneously processes a series of multi-section pathology slides to capture the comprehensive tumor landscape across the entire prostate gland. To develop this predictive AI model, we curated a large-scale dataset of 23,451 slides from 789 patients. The proposed framework demonstrated strong predictive performance for 1- and 2-year BCR prediction, substantially outperforming established clinical benchmarks. The AI-derived risk score was validated as the most potent independent prognostic factor in a multivariable Cox proportional hazards analysis, surpassing conventional clinical markers such as pre-operative PSA and Gleason score. Furthermore, we demonstrated that integrating patch and slide sub-sampling strategies significantly reduces computational cost during both training and inference without compromising predictive performance, and generalizability of AI was confirmed through external validation. Collectively, these results highlight the clinical feasibility and prognostic value of the proposed AI-based multi-section slide analysis as a scalable tool for post-operative management in prostate cancer.

Method: Study structured decoder layer pruning through domain-aware activation similarity, measuring how strongly each layer transforms representations for math vs non-math inputs. Develop math-aware, non-math-aware, and mixed ranking criteria to identify layers with least activation change within target domains.

Result: Across two state-of-the-art VLMs and math/general multimodal benchmarks, discover consistent three-regime structure: low pruning budgets show high sensitivity to which layers are removed; moderate budgets see method convergence as structural damage accumulates; high budgets favor spacing-aware strategies. Domain-aware rankings achieve strongest stability in ranking-sensitive regime while matching/exceeding structure-aware baselines at larger budgets.

Conclusion: Results provide clearer picture of how depth contributes to domain-specific behavior in VLMs and offer practical, interpretable approach to reducing model depth without sacrificing essential mathematical or general vision-language capabilities.

Abstract: Transformer-based vision-language models (VLMs) contain substantial depth redundancy, yet the effect of removing specific decoder layers remains poorly understood, especially for domains that require tight coupling between perception and multi-step reasoning. We study structured decoder layer pruning through the lens of domain-aware activation similarity, measuring how strongly each layer transforms representations for math versus non-math inputs. This yields simple math-aware, non-math-aware, and mixed ranking criteria that identify layers whose input-output activations change least within a target domain. Across two state-of-the-art VLMs and a broad suite of math and general multimodal benchmarks, we uncover a consistent three-regime structure: at low pruning budgets, performance is highly sensitive to which layers are removed; at moderate budgets, methods converge as structural damage accumulates; and at high budgets, structural continuity dominates, favoring spacing-aware strategies. Our domain-aware rankings achieve the strongest stability in the ranking-sensitive regime, while matching or exceeding structure-aware baselines at larger budgets. These results provide a clearer picture of how depth contributes to domain-specific behavior in VLMs and offer a practical, interpretable approach to reducing model depth without sacrificing essential mathematical or general vision-language capabilities.

Method: Uses sensitivity scores (magnitude, gradient, or combination) and architectural rules to derive architecture-aware sparsity ranges. Systematically samples these ranges to produce ten strategies per sensitivity signal, eliminating repeated pruning runs.

Result: Demonstrates Pareto-optimal results on CNNs and Vision Transformers, reducing accuracy degradation on Swin-Tiny by 40% relative to standard single-criterion pruning.

Conclusion: Coordinating existing pruning signals enables more reliable and efficient compressed models than introducing new criteria, offering deployment-ready accuracy-sparsity trade-offs.

Abstract: Deploying deep neural networks (DNNs) on edge devices requires strong compression with minimal accuracy loss. This paper introduces Mix-and-Match Pruning, a globally guided, layer-wise sparsification framework that leverages sensitivity scores and simple architectural rules to generate diverse, high-quality pruning configurations. The framework addresses a key limitation that different layers and architectures respond differently to pruning, making single-strategy approaches suboptimal. Mix-and-Match derives architecture-aware sparsity ranges, e.g., preserving normalization layers while pruning classifiers more aggressively, and systematically samples these ranges to produce ten strategies per sensitivity signal (magnitude, gradient, or their combination). This eliminates repeated pruning runs while offering deployment-ready accuracy-sparsity trade-offs. Experiments on CNNs and Vision Transformers demonstrate Pareto-optimal results, with Mix-and-Match reducing accuracy degradation on Swin-Tiny by 40% relative to standard single-criterion pruning. These findings show that coordinating existing pruning signals enables more reliable and efficient compressed models than introducing new criteria.

Method: STAC framework with three components: (1) Working Temporal Token Caching using decayed cumulative attention scores to preserve long-term informative tokens; (2) Long-term Spatial Token Caching that compresses spatially redundant tokens into voxel-aligned representations; (3) Chunk-based Multi-frame Optimization for joint processing of consecutive frames to improve temporal coherence and GPU efficiency.

Result: Achieves state-of-the-art reconstruction quality while reducing memory consumption by nearly 10x and accelerating inference by 4x, substantially improving scalability for real-time 3D reconstruction in streaming settings.

Conclusion: STAC effectively addresses memory bottlenecks in streaming 3D reconstruction by exploiting intrinsic spatio-temporal sparsity in attention mechanisms, enabling efficient and scalable real-time 3D reconstruction.

Abstract: Online 3D reconstruction from streaming inputs requires both long-term temporal consistency and efficient memory usage. Although causal VGGT transformers address this challenge through a key-value (KV) cache mechanism, the cache grows linearly with the stream length, creating a major memory bottleneck. Under limited memory budgets, early cache eviction significantly degrades reconstruction quality and temporal consistency. In this work, we observe that attention in causal transformers for 3D reconstruction exhibits intrinsic spatio-temporal sparsity. Based on this insight, we propose STAC, a Spatio-Temporally Aware Cache Compression framework for streaming 3D reconstruction with large causal transformers. STAC consists of three key components: (1) a Working Temporal Token Caching mechanism that preserves long-term informative tokens using decayed cumulative attention scores; (2) a Long-term Spatial Token Caching scheme that compresses spatially redundant tokens into voxel-aligned representations for memory-efficient storage; and (3) a Chunk-based Multi-frame Optimization strategy that jointly processes consecutive frames to improve temporal coherence and GPU efficiency. Extensive experiments show that STAC achieves state-of-the-art reconstruction quality while reducing memory consumption by nearly 10x and accelerating inference by 4x, substantially improving the scalability of real-time 3D reconstruction in streaming settings.

Method: Uses Low-Rank Adaptation (LoRA) to fine-tune minimal parameters of Wan2.1 model, designs lightweight pose encoder with stacked 3D convolutional layers, and integrates reference appearance via concatenation with pose information for better alignment.

Result: Achieves visually appealing and temporally consistent high-fidelity animations with strong generalization capabilities, successfully upscaling from 480p training to 720p inference.

Conclusion: UniAnimate-DiT demonstrates effective human image animation with efficient parameter tuning and good generalization, making it a practical solution for high-quality video generation.

Abstract: This report presents UniAnimate-DiT, an advanced project that leverages the cutting-edge and powerful capabilities of the open-source Wan2.1 model for consistent human image animation. Specifically, to preserve the robust generative capabilities of the original Wan2.1 model, we implement Low-Rank Adaptation (LoRA) technique to fine-tune a minimal set of parameters, significantly reducing training memory overhead. A lightweight pose encoder consisting of multiple stacked 3D convolutional layers is designed to encode motion information of driving poses. Furthermore, we adopt a simple concatenation operation to integrate the reference appearance into the model and incorporate the pose information of the reference image for enhanced pose alignment. Experimental results show that our approach achieves visually appearing and temporally consistent high-fidelity animations. Trained on 480p (832x480) videos, UniAnimate-DiT demonstrates strong generalization capabilities to seamlessly upscale to 720P (1280x720) during inference. The training and inference code is publicly available at https://github.com/ali-vilab/UniAnimate-DiT.

Method: Two efficient VAD methods: PatchCore-Lite uses product-quantized memory bank with coarse search followed by exact search on decoded subset. Padim-Lite uses diagonal covariance to convert Mahalanobis distance into efficient element-wise computation.

Result: PatchCore-Lite achieves 79% reduction in total memory footprint. Padim-Lite achieves 77% reduction in total memory and 31% decrease in inference time. Both maintain effectiveness on MVTec AD and VisA benchmarks.

Conclusion: VAD can be effectively deployed on edge devices, enabling real-time, private, and cost-efficient industrial inspection through efficient algorithm design that reduces computational and memory requirements.

Abstract: Visual Anomaly Detection (VAD) is essential for industrial quality control, enabling automatic defect detection in manufacturing. In real production lines, VAD systems must satisfy strict real-time and privacy requirements, necessitating a shift from cloud-based processing to local edge deployment. However, processing data locally on edge devices introduces new challenges because edge hardware has limited memory and computational resources. To overcome these limitations, we propose two efficient VAD methods designed for edge deployment: PatchCore-Lite and Padim-Lite, based on the popular PatchCore and PaDiM models. PatchCore-Lite runs first a coarse search on a product-quantized memory bank, then an exact search on a decoded subset. Padim-Lite is sped up using diagonal covariance, turning Mahalanobis distance into efficient element-wise computation. We evaluate our methods on the MVTec AD and VisA benchmarks and show their suitability for edge environments. PatchCore-Lite achieves a remarkable 79% reduction in total memory footprint, while PaDiM-Lite achieves substantial efficiency gains with a 77% reduction in total memory and a 31% decrease in inference time. These results show that VAD can be effectively deployed on edge devices, enabling real-time, private, and cost-efficient industrial inspection.

Method: Proposes EARTalking with novel frame-by-frame in-context audio-driven streaming generation. Introduces Sink Frame Window Attention (SFA) for variable-length video generation with identity consistency, and streaming Frame Condition In-Context (FCIC) scheme for injecting diverse control signals in streaming manner.

Result: Outperforms existing autoregressive methods and achieves performance comparable to diffusion-based methods. Demonstrates feasibility of in-context streaming autoregressive control for flexible, efficient generation.

Conclusion: EARTalking unlocks scalable direction for flexible, efficient audio-driven talking head generation with interactive control at every frame and arbitrary moments.

Abstract: Audio-driven talking head generation aims to create vivid and realistic videos from a static portrait and speech. Existing AR-based methods rely on intermediate facial representations, which limit their expressiveness and realism. Meanwhile, diffusion-based methods generate clip-by-clip, lacking fine-grained control and causing inherent latency due to overall denoising across the window. To address these limitations, we propose EARTalking, a novel end-to-end, GPT-style autoregressive model for interactive audio-driven talking head generation. Our method introduces a novel frame-by-frame, in-context, audio-driven streaming generation paradigm. For inherently supporting variable-length video generation with identity consistency, we propose the Sink Frame Window Attention (SFA) mechanism. Furthermore, to avoid the complex, separate networks that prior works required for diverse control signals, we propose a streaming Frame Condition In-Context (FCIC) scheme. This scheme efficiently injects diverse control signals in a streaming, in-context manner, enabling interactive control at every frame and at arbitrary moments. Experiments demonstrate that EARTalking outperforms existing autoregressive methods and achieves performance comparable to diffusion-based methods. Our work demonstrates the feasibility of in-context streaming autoregressive control, unlocking a scalable direction for flexible, efficient generation. The code will be released for reproducibility.

Method: The study categorizes existing approaches into three evolutionary stages: handcrafted physical priors, data-driven deep restoration, and hybrid physical-intelligent generation. It summarizes over 30 representative methods across CNNs, GANs, Transformers, and diffusion models. Large-scale quantitative experiments were conducted on five public datasets using 12 evaluation metrics.

Result: Transformer- and diffusion-based models improve SSIM by 12%~18% and reduce perceptual errors by 20%~35% on average. Hybrid physics-guided designs achieve higher radiometric stability, with models using explicit transmission or airlight constraints reducing color bias by up to 27%.

Conclusion: The survey identifies open challenges including dynamic atmospheric modeling, multimodal fusion, lightweight deployment, data scarcity, and joint degradations. It outlines future directions for developing trustworthy, controllable, and efficient (TCE) dehazing systems for remote sensing applications.

Abstract: Remote sensing images (RSIs) are frequently degraded by haze, fog, and thin clouds, which obscure surface reflectance and hinder downstream applications. This study presents the first systematic and unified survey of RSIs dehazing, integrating methodological evolution, benchmark assessment, and physical consistency analysis. We categorize existing approaches into a three-stage progression: from handcrafted physical priors, to data-driven deep restoration, and finally to hybrid physical-intelligent generation, and summarize more than 30 representative methods across CNNs, GANs, Transformers, and diffusion models. To provide a reliable empirical reference, we conduct large-scale quantitative experiments on five public datasets using 12 metrics, including PSNR, SSIM, CIEDE, LPIPS, FID, SAM, ERGAS, UIQI, QNR, NIQE, and HIST. Cross-domain comparison reveals that recent Transformer- and diffusion-based models improve SSIM by 12%~18% and reduce perceptual errors by 20%~35% on average, while hybrid physics-guided designs achieve higher radiometric stability. A dedicated physical radiometric consistency experiment further demonstrates that models with explicit transmission or airlight constraints reduce color bias by up to 27%. Based on these findings, we summarize open challenges: dynamic atmospheric modeling, multimodal fusion, lightweight deployment, data scarcity, and joint degradations, and outline promising research directions for future development of trustworthy, controllable, and efficient (TCE) dehazing systems. All reviewed resources, including source code, benchmark datasets, evaluation metrics, and reproduction configurations are publicly available at https://github.com/VisionVerse/RemoteSensing-Restoration-Survey.

Method: 1) Created TransFrag27K dataset with multiscene synthetic data of broken transparent fragments and scalable generation pipeline. 2) Developed TransFragNet for visual grasping position detection. 3) Used Gelsight Mini sensors on two-finger gripper to obtain tactile information of fragment edges. 4) Proposed visual-tactile fusion material classifier. 5) Introduced visual-tactile fusion contour estimation framework inspired by human perception. 6) Developed multi-dimensional similarity metrics for contour matching and reassembly.

Result: The framework demonstrates strong performance in real-world validation, providing a reproducible benchmark for evaluating visual-tactile contour estimation and fragment reassembly. Experimental results validate the proposed approach.

Conclusion: The proposed visual-tactile fusion framework effectively addresses the challenging problem of transparent fragment contour estimation, offering a comprehensive solution with dataset, detection network, fusion classifier, and reassembly algorithm for practical applications in reassembly tasks.

Abstract: The contour estimation of transparent fragments is very important for autonomous reassembly, especially in the fields of precision optical instrument repair, cultural relic restoration, and identification of other precious device broken accidents. Different from general intact transparent objects, the contour estimation of transparent fragments face greater challenges due to strict optical properties, irregular shapes and edges. To address this issue, a general transparent fragments contour estimation framework based on visual-tactile fusion is proposed in this paper. First, we construct the transparent fragment dataset named TransFrag27K, which includes a multiscene synthetic data of broken fragments from multiple types of transparent objects, and a scalable synthetic data generation pipeline. Secondly, we propose a visual grasping position detection network named TransFragNet to identify, locate and segment the sampling grasping position. And, we use a two-finger gripper with Gelsight Mini sensors to obtain reconstructed tactile information of the lateral edge of the fragments. By fusing this tactile information with visual cues, a visual-tactile fusion material classifier is proposed. Inspired by the way humans estimate a fragment’s contour combining vision and touch, we introduce a general transparent fragment contour estimation framework based on visual-tactile fusion, demonstrates strong performance in real-world validation. Finally, a multi-dimensional similarity metrics based contour matching and reassembly algorithm is proposed, providing a reproducible benchmark for evaluating visual-tactile contour estimation and fragment reassembly. The experimental results demonstrate the validity of the proposed framework. The dataset and codes are available at https://github.com/Keithllin/Transparent-Fragments-Contour-Estimation.

Method: Develops a 360° saliency graph representation that explicitly encodes visual, contextual, semantic, and geometric information as nodes, edges, edge weights, and angular positions. This representation is robust to view changes and environmental challenges. Uses this representation for vision-based navigation by first localizing query scenes in topological maps, then estimating movement directions using embedded geometric information.

Result: Experimental results demonstrate the efficacy of the proposed 360° saliency graph representation in enhancing both scene localization and vision-based indoor navigation compared to existing methods.

Conclusion: The 360° saliency graph provides a rich, efficient representation that addresses limitations of traditional scene representations and improves performance in vision-based navigation applications.

Abstract: A Scene, represented visually using different formats such as RGB-D, LiDAR scan, keypoints, rectangular, spherical, multi-views, etc., contains information implicitly embedded relevant to applications such as scene indexing, vision-based navigation. Thus, these representations may not be efficient for such applications. This paper proposes a novel 360° saliency graph representation of the scenes. This rich representation explicitly encodes the relevant visual, contextual, semantic, and geometric information of the scene as nodes, edges, edge weights, and angular position in the 360° graph. Also, this representation is robust against scene view change and addresses challenges of indoor environments such as varied illumination, occlusions, and shadows as in the case of existing traditional methods. We have utilized this rich and efficient representation for vision-based navigation and compared it with existing navigation methods using 360° scenes. However, these existing methods suffer from limitations of poor scene representation, lacking scene-specific information. This work utilizes the proposed representation first to localize the query scene in the given topological map, and then facilitate 2D navigation by estimating the next required movement directions towards the target destination in the topological map by using the embedded geometric information in the 360° saliency graph. Experimental results demonstrate the efficacy of the proposed 360° saliency graph representation in enhancing both scene localization and vision-based indoor navigation.

Method: Teacher-based knowledge retention method that uses incremental category samples instead of old ones, combined with mask-based partial category knowledge distillation that decouples and filters potentially misleading information to prevent misguiding the student model.

Result: Comparative and ablation experiments show robust performance, demonstrating effectiveness in mitigating catastrophic forgetting and enhancing overall accuracy in incremental hyperspectral image classification.

Conclusion: The proposed approach successfully addresses catastrophic forgetting in incremental hyperspectral image classification without requiring old category samples, using teacher-based knowledge retention and selective knowledge distillation.

Abstract: In incremental classification tasks for hyperspectral images, catastrophic forgetting is an unavoidable challenge. While memory recall methods can mitigate this issue, they heavily rely on samples from old categories. This paper proposes a teacher-based knowledge retention method for incremental image classification. It alleviates model forgetting of old category samples by utilizing incremental category samples, without depending on old category samples. Additionally, this paper introduces a mask-based partial category knowledge distillation algorithm. By decoupling knowledge distillation, this approach filters out potentially misleading information that could misguide the student model, thereby enhancing overall accuracy. Comparative and ablation experiments demonstrate the proposed method’s robust performance.

Method: Proposes Dynamic Sparse Compressed Sensing Network (DSCSNet), a deep-unfolded network that couples ADMM with learnable parameters. Uses ℓ₁-norm sparsity constraints in auxiliary variable updates instead of traditional ℓ₂-norm, and integrates self-attention-based dynamic thresholding for adaptive sparsification. Modules are jointly optimized end-to-end across three ADMM iterative steps.

Result: Extensive experiments on synthetic infrared dataset CSIST-100K show DSCSNet outperforms state-of-the-art methods in key metrics like CSO-mAP and sub-pixel localization error.

Conclusion: DSCSNet achieves robust sparsity induction and scene adaptability while retaining compressed sensing physical logic, enhancing unmixing accuracy and generalization in complex infrared scenarios.

Abstract: Due to the limitations of optical lens focal length and detector resolution, distant clustered infrared small targets often appear as mixed spots. The Close Small Object Unmixing (CSOU) task aims to recover the number, sub-pixel positions, and radiant intensities of individual targets from these spots, which is a highly ill-posed inverse problem. Existing methods struggle to balance the rigorous sparsity guarantees of model-driven approaches and the dynamic scene adaptability of data-driven methods. To address this dilemma, this paper proposes a Dynamic Sparse Compressed Sensing Network (DSCSNet), a deep-unfolded network that couples the Alternating Direction Method of Multipliers (ADMM) with learnable parameters. Specifically, we embed a strict $\ell_1$-norm sparsity constraint into the auxiliary variable update step of ADMM to replace the traditional $\ell_2$-norm smoothness-promoting terms, which effectively preserves the discrete energy peaks of small targets. We also integrate a self-attention-based dynamic thresholding mechanism into the reconstruction stage, which adaptively adjusts the sparsification intensity using the sparsity-enhanced information from the iterative process. These modules are jointly optimized end-to-end across the three iterative steps of ADMM. Retaining the physical logic of compressed sensing, DSCSNet achieves robust sparsity induction and scene adaptability, thus enhancing the unmixing accuracy and generalization in complex infrared scenarios. Extensive experiments on the synthetic infrared dataset CSIST-100K demonstrate that DSCSNet outperforms state-of-the-art methods in key metrics such as CSO-mAP and sub-pixel localization error.

Method: Introduces a lightweight preprocessing and splatting pipeline that expands usable dynamic range and stabilizes per-frame photometry. The approach specifically addresses thermal imagery’s unique challenges without requiring any dataset-specific tuning.

Result: Achieves state-of-the-art performance across thermal-only novel view synthesis benchmarks, demonstrating effective handling of thermal imagery’s limitations for view synthesis tasks.

Conclusion: The proposed preprocessing and splatting pipeline successfully addresses the core challenges of thermal imagery for novel view synthesis, enabling reliable thermal-only view synthesis without RGB guidance beyond camera pose.

Abstract: Thermal cameras provide reliable visibility in darkness and adverse conditions, but thermal imagery remains significantly harder to use for novel view synthesis (NVS) than visible-light images. This difficulty stems primarily from two characteristics of affordable thermal sensors. First, thermal images have extremely low dynamic range, which weakens appearance cues and limits the gradients available for optimization. Second, thermal data exhibit rapid frame-to-frame photometric fluctuations together with slow radiometric drift, both of which destabilize correspondence estimation and create high-frequency floater artifacts during view synthesis, particularly when no RGB guidance (beyond camera pose) is available. Guided by these observations, we introduce a lightweight preprocessing and splatting pipeline that expands usable dynamic range and stabilizes per-frame photometry. Our approach achieves state-of-the-art performance across thermal-only NVS benchmarks, without requiring any dataset-specific tuning.

Method: Introduces InjectFlow, a training-free method that injects orthogonal semantics during the initial velocity field computation without changing random seeds. This prevents latent drift toward majority modes while maintaining generative quality.

Result: Extensive experiments show InjectFlow successfully fixes 75% of prompts that standard flow matching models fail to generate correctly on the GenEval dataset.

Conclusion: Provides theoretical analysis and algorithm for building more fair and robust visual foundation models, offering a ready-to-use solution for bias mitigation in Flow Matching.

Abstract: Flow Matching (FM) has recently emerged as a leading approach for high-fidelity visual generation, offering a robust continuous-time alternative to ordinary differential equation (ODE) based models. However, despite their success, FM models are highly sensitive to dataset biases, which cause severe semantic degradation when generating out-of-distribution or minority-class samples. In this paper, we provide a rigorous mathematical formalization of the ``Bias Manifold’’ within the FM framework. We identify that this performance drop is driven by conditional expectation smoothing, a mechanism that inevitably leads to trajectory lock-in during inference. To resolve this, we introduce InjectFlow, a novel, training-free method by injecting orthogonal semantics during the initial velocity field computation, without requiring any changes to the random seeds. This design effectively prevents the latent drift toward majority modes while maintaining high generative quality. Extensive experiments demonstrate the effectiveness of our approach. Notably, on the GenEval dataset, InjectFlow successfully fixes 75% of the prompts that standard flow matching models fail to generate correctly. Ultimately, our theoretical analysis and algorithm provide a ready-to-use solution for building more fair and robust visual foundation models.

Method: Introduces StreamingEval framework with fixed-capacity memory bank to normalize accessible historical visual context. Jointly evaluates visual encoding efficiency, text decoding latency, and task performance to quantify overall system deployability under standardized protocol.

Result: Extensive experiments across multiple datasets reveal substantial gaps between current Video-LLMs and requirements of realistic streaming applications, providing systematic basis for future research.

Conclusion: StreamingEval addresses the need for practical evaluation of streaming video understanding, highlighting current limitations and providing a standardized framework for future Video-LLM development in real-world interactive applications.

Abstract: Real-time, continuous understanding of visual signals is essential for real-world interactive AI applications, and poses a fundamental system-level challenge. Existing research on streaming video understanding, however, typically focuses on isolated aspects such as question-answering accuracy under limited visual context or improvements in encoding efficiency, while largely overlooking practical deployability under realistic resource constraints. To bridge this gap, we introduce StreamingEval, a unified evaluation framework for assessing the streaming video understanding capabilities of Video-LLMs under realistic constraints. StreamingEval benchmarks both mainstream offline models and recent online video models under a standardized protocol, explicitly characterizing the trade-off between efficiency, storage and accuracy. Specifically, we adopt a fixed-capacity memory bank to normalize accessible historical visual context, and jointly evaluate visual encoding efficiency, text decoding latency, and task performance to quantify overall system deployability. Extensive experiments across multiple datasets reveal substantial gaps between current Video-LLMs and the requirements of realistic streaming applications, providing a systematic basis for future research in this direction. Codes will be released at https://github.com/wwgTang-111/StreamingEval1.

Method: Introduced NoiseZoo dataset containing per-image DDIM-inverted diffusion noise and matching encoder representations (CLIP, DINO). Used linear probes to analyze semantic information in both spaces and performed orthogonalization for disentangled image editing.

Result: Linear probes show strong, aligned attribute predictions in both spaces, revealing that generative noise encodes meaningful semantics. Simple orthogonalization enables faithful, controllable image edits (smile, gender, age) without architectural changes.

Conclusion: Provides empirical support for the UNE hypothesis, revealing a shared Gaussian-like latent geometry that concretely links encoding and generation, enabling new editing capabilities.

Abstract: Generative models and vision encoders have largely advanced on separate tracks, optimized for different goals and grounded in different mathematical principles. Yet, they share a fundamental property: latent space Gaussianity. Generative models map Gaussian noise to images, while encoders map images to semantic embeddings whose coordinates empirically behave as Gaussian. We hypothesize that both are views of a shared latent source, the Universal Normal Embedding (UNE): an approximately Gaussian latent space from which encoder embeddings and DDIM-inverted noise arise as noisy linear projections. To test our hypothesis, we introduce NoiseZoo, a dataset of per-image latents comprising DDIM-inverted diffusion noise and matching encoder representations (CLIP, DINO). On CelebA, linear probes in both spaces yield strong, aligned attribute predictions, indicating that generative noise encodes meaningful semantics along linear directions. These directions further enable faithful, controllable edits (e.g., smile, gender, age) without architectural changes, where simple orthogonalization mitigates spurious entanglements. Taken together, our results provide empirical support for the UNE hypothesis and reveal a shared Gaussian-like latent geometry that concretely links encoding and generation. Code and data are available https://rbetser.github.io/UNE/

Method: Instead of encoding watermarks into initial noise, DiffMark injects a persistent learned perturbation at every denoising step of a frozen diffusion model. Uses Latent Consistency Models as a differentiable training bridge to avoid backpropagating through full denoising chain, reducing gradient steps from 50 DDIM to 4 LCM.

Result: Achieves single-pass detection at 16.4 ms (45x speedup over sampling-based methods), maintains competitive watermark robustness against distortion/regeneration/adversarial attacks, and provides per-image key flexibility with cross-model transferability.

Conclusion: DiffMark offers a practical plug-and-play watermarking solution for diffusion models that balances efficiency, flexibility, and robustness while enabling cross-architecture compatibility without per-model fine-tuning.

Abstract: As diffusion models (DMs) enable photorealistic image generation at unprecedented scale, watermarking techniques have become essential for provenance establishment and accountability. Existing methods face challenges: sampling-based approaches operate on frozen models but require costly $N$-step Denoising Diffusion Implicit Models (DDIM) inversion (typically N=50) for zero-bit-only detection; fine-tuning-based methods achieve fast multi-bit extraction but couple the watermark to a specific model checkpoint, requiring retraining for each architecture. We propose DiffMark, a plug-and-play watermarking method that offers three key advantages over existing approaches: single-pass multi-bit detection, per-image key flexibility, and cross-model transferability. Rather than encoding the watermark into the initial noise vector, DiffMark injects a persistent learned perturbation $δ$ at every denoising step of a completely frozen DM. The watermark signal accumulates in the final denoised latent $z_0$ and is recovered in a single forward pass. The central challenge of backpropagating gradients through a frozen UNet without traversing the full denoising chain is addressed by employing Latent Consistency Models (LCM) as a differentiable training bridge. This reduces the number of gradient steps from 50 DDIM to 4 LCM and enables a single-pass detection at 16.4 ms, a 45x speedup over sampling-based methods. Moreover, by this design, the encoder learns to map any runtime secret to a unique perturbation at inference time, providing genuine per-image key flexibility and transferability to unseen diffusion-based architectures without per-model fine-tuning. Although achieving these advantages, DiffMark also maintains competitive watermark robustness against distortion, regeneration, and adversarial attacks.

Method: 1) Superpixel Generation module extracts structural priors from source domain using Simple Linear Iterative Clustering. 2) Resolution-adaptive Fusion aligns source structures with target backgrounds via texture similarity. 3) Pseudo-label re-Synthesize mechanism uses multi-stage noise generation to simulate realistic rain streaks. Framework is plug-and-play for arbitrary deraining architectures.

Result: Achieves remarkable PSNR gains of 32% to 59% in OOD domains while significantly accelerating training convergence. Extensive experiments on state-of-the-art models demonstrate effectiveness.

Conclusion: Proposed framework successfully addresses domain adaptation challenges in image deraining without requiring paired rainy observations, enabling robust performance in unseen scenarios through realistic pseudo-data generation.

Abstract: Image deraining plays a pivotal role in low-level computer vision, serving as a prerequisite for robust outdoor surveillance and autonomous driving systems. While deep learning paradigms have achieved remarkable success in firmly aligned settings, they often suffer from severe performance degradation when generalized to unseen Out-of-Distribution (OOD) scenarios. This failure stems primarily from the significant domain discrepancy between synthetic training datasets and the complex physical dynamics of real-world rain. To address these challenges, this paper proposes a pioneering cross-scenario deraining adaptation framework. Diverging from conventional approaches, our method obviates the requirements for paired rainy observations in the target domain, leveraging exclusively rain-free background images. We design a Superpixel Generation (Sup-Gen) module to extract stable structural priors from the source domain using Simple Linear Iterative Clustering. Subsequently, a Resolution-adaptive Fusion strategy is introduced to align these source structures with target backgrounds through texture similarity, ensuring the synthesis of diverse and realistic pseudo-data. Finally, we implement a pseudo-label re-Synthesize mechanism that employs multi-stage noise generation to simulate realistic rain streaks. This framework functions as a versatile plug-and-play module capable of seamless integration into arbitrary deraining architectures. Extensive experiments on state-of-the-art models demonstrate that our approach yields remarkable PSNR gains of up to 32% to 59% in OOD domains while significantly accelerating training convergence.

Method: Proposes a latent representation-based framework using either direct one-step training or a two-step strategy combining VAE pretraining with parameter-to-latent interpolation for both spectrum-level and spatial-spectral emulation.

Result: Outperforms classical regression-based emulators in reconstruction accuracy, spectral fidelity, and robustness to real-world spatial variability on PROSAIL-simulated vegetation data and Sentinel-3 OLCI imagery.

Conclusion: The method enables efficient hyperspectral emulation that preserves downstream biophysical parameter retrieval performance, making it practically relevant for remote sensing applications.

Abstract: Synthetic hyperspectral image (HSI) generation is essential for large-scale simulation, algorithm development, and mission design, yet traditional radiative transfer models remain computationally expensive and often limited to spectrum-level outputs. In this work, we propose a latent representation-based framework for hyperspectral emulation that learns a latent generative representation of hyperspectral data. The proposed approach supports both spectrum-level and spatial-spectral emulation and can be trained either in a direct one-step formulation or in a two-step strategy that couples variational autoencoder (VAE) pretraining with parameter-to-latent interpolation. Experiments on PROSAIL-simulated vegetation data and Sentinel-3 OLCI imagery demonstrate that the method outperforms classical regression-based emulators in reconstruction accuracy, spectral fidelity, and robustness to real-world spatial variability. We further show that emulated HSIs preserve performance in downstream biophysical parameter retrieval, highlighting the practical relevance of emulated data for remote sensing applications.

Method: Proposes establishing relevant interaction schemes through illustrative use cases and discusses under-explored research directions for leveraging available data through multiple extents and communities.

Result: Identifies gaps in current geospatial data fusion approaches and proposes a framework for more symmetric, mutually beneficial integration of diverse data sources.

Conclusion: Current geospatial data fusion is limited by asymmetric approaches; future research should focus on creating more balanced, mutually beneficial integration schemes across scales and communities.

Abstract: We face a unprecedented amount of geospatial data, describing directly or indirectly the Earth Surface at multiple spatial, temporal, and semantic scales, and stemming from numerous contributors, from satellites to citizens. The main challenge in all the geospatial-related communities lies in suitably leveraging a combination of some of the sources for either a generic or a thematic application. Certain data fusion schemes are predominantly exploited: they correspond to popular tasks with mainstream data sources, e.g., free archives of Sentinel images coupled with OpenStreetMap data under an open and widespread deep-learning backbone for land-cover mapping purposes. Most of these approaches unfortunately operate under a “master-slave” paradigm, where one source is basically integrated to help processing the “main” source, without mutual advantages (e.g., large-scale estimation of a given biophysical variable using in-situ observations) and under a specific community bias. We argue that numerous key data fusion configurations, and in particular the effort in symmetrizing the exploitation of multiple data sources, are insufficiently addressed while being highly beneficial for generic or thematic applications. Bridges and retroactions between scales, communities and their respective sources are lacking, neglecting the utmost potential of such a “global-local loop”. In this paper, we propose to establish the most relevant interaction schemes through illustrative use cases. We subsequently discuss under-explored research directions that could take advantage of leveraging available data through multiples extents and communities.

Method: Proposes GraphiContact framework that transfers complementary human priors from two pretrained Transformer encoders to predict per-vertex human-scene contact on reconstructed mesh. Introduces Single-Image Multi-Infer Uncertainty (SIMU) training strategy with token-level adaptive routing to simulate occlusion and noisy observations during training while preserving efficient single-branch inference at test time.

Result: Experiments on five benchmark datasets show that GraphiContact achieves consistent gains on both contact prediction and 3D human reconstruction.

Conclusion: GraphiContact provides a comprehensive solution for 3D human reconstruction and interaction analysis, addressing the gap between contact prediction and mesh reconstruction with improved robustness to real-world challenges like occlusion and noise.

Abstract: Monocular vertex-level human-scene contact prediction is a fundamental capability for interactive systems such as assistive monitoring, embodied AI, and rehabilitation analysis. In this work, we study this task jointly with single-image 3D human mesh reconstruction, using reconstructed body geometry as a scaffold for contact reasoning. Existing approaches either focus on contact prediction without sufficiently exploiting explicit 3D human priors, or emphasize pose/mesh reconstruction without directly optimizing robust vertex-level contact inference under occlusion and perceptual noise. To address this gap, we propose GraphiContact, a pose-aware framework that transfers complementary human priors from two pretrained Transformer encoders and predicts per-vertex human-scene contact on the reconstructed mesh. To improve robustness in real-world scenarios, we further introduce a Single-Image Multi-Infer Uncertainty (SIMU) training strategy with token-level adaptive routing, which simulates occlusion and noisy observations during training while preserving efficient single-branch inference at test time. Experiments on five benchmark datasets show that GraphiContact achieves consistent gains on both contact prediction and 3D human reconstruction. Our code, based on the GraphiContact method, provides comprehensive 3D human reconstruction and interaction analysis, and will be publicly available at https://github.com/Aveiro-Lin/GraphiContact.

Method: Four-backbone ViT ensemble (CLIP and SigLIP variants) with expanded training data and degradation modeling. Two-stage training: binary classification followed by feature-level self-distillation for representation alignment.

Result: Achieves strong robustness and generalization in NTIRE challenge setting, handling diverse “in-the-wild” conditions with stable predictions across unseen generators and complex degradations.

Conclusion: FeatDistill offers an effective practical solution for real-world deepfake detection, balancing performance with efficiency (10GB GPU memory).

Abstract: The rapid iteration and widespread dissemination of deepfake technology have posed severe challenges to information security, making robust and generalizable detection of AI-generated forged images increasingly important. In this paper, we propose FeatDistill, an AI-generated image detection framework that integrates feature distillation with a multi-expert ensemble, developed for the NTIRE Challenge on Robust AI-Generated Image Detection in the Wild. The framework explicitly targets three practical bottlenecks in real-world forensics: degradation interference, insufficient feature representation, and limited generalization. Concretely, we build a four-backbone Vision Transformer (ViT) ensemble composed of CLIP and SigLIP variants to capture complementary forensic cues. To improve data coverage, we expand the training set and introduce comprehensive degradation modeling, which exposes the detector to diverse quality variations and synthesis artifacts commonly encountered in unconstrained scenarios. We further adopt a two-stage training paradigm: the model is first optimized with a standard binary classification objective, then refined by dense feature-level self-distillation for representation alignment. This design effectively mitigates overfitting and enhances semantic consistency of learned features. At inference time, the final prediction is obtained by averaging the probabilities from four independently trained experts, yielding stable and reliable decisions across unseen generators and complex degradations. Despite the ensemble design, the framework remains efficient, requiring only about 10 GB peak GPU memory. Extensive evaluations in the NTIRE challenge setting demonstrate that FeatDistill achieves strong robustness and generalization under diverse ``in-the-wild’’ conditions, offering an effective and practical solution for real-world deepfake image detection.

Method: Proposes Visual Grounding Score Guided Decoding (VGS-Decoding) which introduces a Visual Grounding Score (VGS) that measures each token’s visual dependency by comparing probability distributions from original vs. distorted images. During decoding, probabilities are reweighted to amplify visually grounded tokens and suppress hallucinations.

Result: Experiments on MIMIC-Diff-VQA and VQA-RAD datasets across LLaVA-Med, CheXagent, and MedGemma models show consistent improvements: up to +9.12% overall gain and +8.98% in open-ended recall, with only 2× inference overhead and no additional training.

Conclusion: VGS-Decoding effectively mitigates hallucinations in medical VLMs through adaptive per-token probability reweighting based on visual grounding scores, offering a practical, training-free solution for clinical deployment.

Abstract: Medical Vision-Language Models (VLMs) often hallucinate by generating responses based on language priors rather than visual evidence, posing risks in clinical applications. We propose Visual Grounding Score Guided Decoding (VGS-Decoding), a training-free method to mitigate hallucinations during inference. Our key insight is that hallucinated tokens maintain or increase their probability when visual information is degraded, while visually grounded tokens decrease in probability. We introduce the Visual Grounding Score (VGS), which measures each token’s visual dependency by comparing distributions from original and distorted images. During decoding, we reweight probabilities by amplifying visually grounded tokens while suppressing hallucinations. Unlike fixed-weight contrastive methods, VGS-Decoding provides per-token adaptive control. Experiments on MIMIC-Diff-VQA and VQA-RAD across LLaVA-Med, CheXagent, and MedGemma demonstrate consistent improvements, with up to +9.12% overall gain and $+8.98%$ in open-ended recall, while introducing only $2\times$ inference overhead and no additional training, making it practical for clinical deployment. Upon acceptance, code will be released publicly to facilitate reproducibility.

Method: Uses interpretability-driven analysis to identify that cross-modal attention heads fail to distinguish video tokens across different timestamps. Proposes Temporally Conditioned Attention Sharpening (TCAS) - an attention enhancement method that constructs an objective based on attention distinctions to improve temporal resolution capability.

Result: TCAS significantly enhances temporal logic consistency of Video-LLMs. Further analysis shows it improves temporal discriminability of attention heads. The method also achieves performance improvements in general video temporal grounding tasks.

Conclusion: Temporal logic consistency is crucial for temporal understanding in Video-LLMs. The proposed TCAS method effectively addresses inconsistency issues by enhancing attention mechanisms’ temporal discrimination capabilities.

Abstract: Large language models (LLMs) often generate self-contradictory outputs, which severely impacts their reliability and hinders their adoption in practical applications. In video-language models (Video-LLMs), this phenomenon recently draws the attention of researchers. Specifically, these models fail to provide logically consistent responses to rephrased questions based on their grounding outputs. However, the underlying causes of this phenomenon remain underexplored. In this work, we adopt an interpretability-driven approach to analyze, statistically summarize, and intervention the potential factors of the phenomenon. We find that one of the primary reasons for the inconsistency in responses lies in the inability of cross-modal attention heads to effectively distinguish video tokens across different timestamps. To address this, we propose an attention enhancement method called Temporally Conditioned Attention Sharpening (TCAS), which constructs an enhancement objective based on attention distinctions to enhance the model’s temporal resolution capability, thereby improving its temporal understanding logic consistency. Experimental results demonstrate that our method significantly enhances the temporal logic consistency of Video-LLMs. Further analyses reveal that our method indeed improves the temporal discriminability of attention heads, validating our conclusions. Additionally, our method even achieves performance improvements in general video temporal grounding tasks, suggesting that temporal logic consistency is an important factor in temporal understanding.

Method: Proposes a workload-centric framework with phased adoption model tied to orbital data center maturity. Grounds the framework with in-orbit semantic-reduction prototypes: 1) Earth-observation pipeline on Sentinel-2 imagery converting raw imagery to compact semantic artifacts, 2) Multi-pass stereo reconstruction prototype reducing raw data to derived 3D representations.

Result: Achieved 99.7-99.99% payload reduction for Earth-observation pipeline on Seattle and Bengaluru imagery. Multi-pass stereo reconstruction reduced ~306 MB to ~1.57 MB (99.49% reduction). Demonstrates semantic abstraction drives early workload suitability more than raw compute scale.

Conclusion: Supports a workload-first view where semantic abstraction, not raw compute scale, determines early suitability for space-based processing. The framework helps decide orbit vs. terrestrial allocation for AI workloads based on data reduction potential.

Abstract: Space-based compute is becoming plausible as launch costs fall and data-intensive AI workloads grow. This paper proposes a workload-centric framework for deciding which tasks belong in orbit versus terrestrial cloud, along with a phased adoption model tied to orbital data center maturity. We ground the framework with in-orbit semantic-reduction prototypes. An Earth-observation pipeline on Sentinel-2 imagery from Seattle and Bengaluru (formerly Bangalore) achieves 99.7-99.99% payload reduction by converting raw imagery to compact semantic artifacts. A multi-pass stereo reconstruction prototype reduces ~306 MB to ~1.57 MB of derived 3D representations (99.49% reduction). These results support a workload-first view in which semantic abstraction, not raw compute scale, drives early workload suitability.

Method: 1) Visuo-temporal velocity-based joint embedding fusing sub-pixel joint cues and inter-frame motion; 2) Attention-driven pose-query encoder mapping joint representations to pose-aware node space; 3) Dual-branch decoupled spatio-temporal attention graph modeling temporal propagation and spatial constraints; 4) Node-space expert fusion module adaptively combining branch outputs.

Result: Extensive experiments on three widely used video pose benchmarks demonstrate superior performance over state-of-the-art methods, highlighting the value of explicit node-centric reasoning.

Conclusion: The proposed node-centric framework offers a new perspective for advancing video-based human pose estimation by explicitly integrating visual, temporal, and structural reasoning, addressing key challenges in motion blur, occlusion, and spatiotemporal dynamics.

Abstract: Video-based human pose estimation remains challenged by motion blur, occlusion, and complex spatiotemporal dynamics. Existing methods often rely on heatmaps or implicit spatio-temporal feature aggregation, which limits joint topology expressiveness and weakens cross-frame consistency. To address these problems, we propose a novel node-centric framework that explicitly integrates visual, temporal, and structural reasoning for accurate pose estimation. First, we design a visuo-temporal velocity-based joint embedding that fuses sub-pixel joint cues and inter-frame motion to build appearance- and motion-aware representations. Then, we introduce an attention-driven pose-query encoder, which applies attention over joint-wise heatmaps and frame-wise features to map the joint representations into a pose-aware node space, generating image-conditioned joint-aware node embeddings. Building upon these node embeddings, we propose a dual-branch decoupled spatio-temporal attention graph that models temporal propagation and spatial constraint reasoning in specialized local and global branches. Finally, a node-space expert fusion module is proposed to adaptively fuse the complementary outputs from both branches, integrating local and global cues for final joint predictions. Extensive experiments on three widely used video pose benchmarks demonstrate that our method outperforms state-of-the-art methods. The results highlight the value of explicit node-centric reasoning, offering a new perspective for advancing video-based human pose estimation.

Method: Proposes DCG-Net with: 1) Dual Cross-Attention module for bidirectional attention between visual tokens and textual concept-value prototypes, enabling localized evidence attribution; 2) Parametric Concept Graph initialized with Positive Pointwise Mutual Information priors and refined through sparsity-controlled message passing to capture relational structure among clinical concepts.

Result: Experiments on white blood cell morphology and skin lesion diagnosis show DCG-Net achieves state-of-the-art classification performance while producing clinically interpretable diagnostic explanations.

Conclusion: DCG-Net provides an end-to-end interpretable framework that integrates multimodal alignment with structured concept reasoning, offering both high performance and clinically meaningful explanations for medical image analysis.

Abstract: Deep learning models have achieved strong performance in medical image analysis, but their internal decision processes remain difficult to interpret. Concept Bottleneck Models (CBMs) partially address this limitation by structuring predictions through human-interpretable clinical concepts. However, existing CBMs typically overlook the contextual dependencies among concepts. To address these issues, we propose an end-to-end interpretable framework \emph{DCG-Net} that integrates multimodal alignment with structured concept reasoning. DCG-Net introduces a Dual Cross-Attention module that replaces cosine similarity matching with bidirectional attention between visual tokens and canonicalized textual concept-value prototypes, enabling spatially localized evidence attribution. To capture the relational structure inherent to clinical concepts, we develop a Parametric Concept Graph initialized with Positive Pointwise Mutual Information priors and refined through sparsity-controlled message passing. This formulation models inter-concept dependencies in a manner consistent with clinical domain knowledge. Experiments on white blood cell morphology and skin lesion diagnosis demonstrate that DCG-Net achieves state-of-the-art classification performance while producing clinically interpretable diagnostic explanations.

Method: The method uses the rotation direction of linear polarization as a discriminative factor between direct and inter-reflection. By actively rotating the linear polarization of incident light and analyzing the rotation direction of reflected light, it decomposes reflectance components in metal objects.

Result: Evaluation with synthetic and real data demonstrates effectiveness in decomposing specular inter-reflections of metal objects. The method can be combined with other decomposition techniques for detailed light transport analysis and improves 3D measurement accuracy against strong specular inter-reflection.

Conclusion: The proposed polarization-based decomposition method successfully addresses the limitation of existing methods by handling specular inter-reflection in metal objects, offering practical applications in 3D measurement and detailed light transport analysis.

Abstract: Polarization is well known for its ability to decompose diffuse and specular reflections. However, the existing decomposition methods only focus on direct reflection and overlook multiple reflections, especially specular inter-reflection. In this paper, we propose a novel decomposition method for handling specular inter-reflection of metal objects by using a unique polarimetric feature: the rotation direction of linear polarization. This rotation direction serves as a discriminative factor between direct and inter-reflection on specular surfaces. To decompose the reflectance components, we actively rotate the linear polarization of incident light and analyze the rotation direction of the reflected light. We evaluate our method using both synthetic and real data, demonstrating its effectiveness in decomposing specular inter-reflections of metal objects. Furthermore, we demonstrate that our method can be combined with other decomposition methods for a detailed analysis of light transport. As a practical application, we show its effectiveness in improving the accuracy of 3D measurement against strong specular inter-reflection.

Method: Proposes a prompt-free SAM adaptation that leverages multi-level encoder features and residual decoding. Only fine-tunes LoRA modules within the frozen SAM encoder, requiring just 4.1M trainable parameters for efficient adaptation.

Result: Achieves state-of-the-art performance on three benchmark datasets (TNBC, MoNuSeg, PanNuke) and demonstrates strong cross-dataset generalization capabilities.

Conclusion: The proposed framework provides an effective and practical solution for histopathology nuclei segmentation with computational efficiency and strong generalization across datasets.

Abstract: Histopathology nuclei segmentation is crucial for quantitative tissue analysis and cancer diagnosis. Although existing segmentation methods have achieved strong performance, they are often computationally heavy and show limited generalization across datasets, which constrains their practical deployment. Recent SAM-based approaches have shown great potential in general and medical imaging, but typically rely on prompt guidance or complex decoders, making them less suitable for histopathology images with dense nuclei and heterogeneous appearances. We propose a prompt-free and lightweight SAM adaptation that leverages multi-level encoder features and residual decoding for accurate and efficient nuclei segmentation. The framework fine-tunes only LoRA modules within the frozen SAM encoder, requiring just 4.1M trainable parameters. Experiments on three benchmark datasets TNBC, MoNuSeg, and PanNuke demonstrate state-of-the-art performance and strong cross-dataset generalization, highlighting the effectiveness and practicality of the proposed framework for histopathology applications.

Method: ADE-CoT incorporates three strategies: 1) difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty, 2) edit-specific verification using region localization and caption consistency for early pruning, and 3) depth-first opportunistic stopping guided by an instance-specific verifier.

Result: Experiments on three state-of-the-art editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show ADE-CoT achieves superior performance-efficiency trade-offs, obtaining better performance with more than 2x speedup over Best-of-N with comparable sampling budgets.

Conclusion: ADE-CoT successfully adapts Image-CoT principles to image editing by addressing the unique challenges of goal-directed editing, demonstrating significant improvements in both efficiency and performance through adaptive resource allocation and verification strategies.

Abstract: Image Chain-of-Thought (Image-CoT) is a test-time scaling paradigm that improves image generation by extending inference time. Most Image-CoT methods focus on text-to-image (T2I) generation. Unlike T2I generation, image editing is goal-directed: the solution space is constrained by the source image and instruction. This mismatch causes three challenges when applying Image-CoT to editing: inefficient resource allocation with fixed sampling budgets, unreliable early-stage verification using general MLLM scores, and redundant edited results from large-scale sampling. To address this, we propose ADaptive Edit-CoT (ADE-CoT), an on-demand test-time scaling framework to enhance editing efficiency and performance. It incorporates three key strategies: (1) a difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty; (2) edit-specific verification in early pruning that uses region localization and caption consistency to select promising candidates; and (3) depth-first opportunistic stopping, guided by an instance-specific verifier, that terminates when intent-aligned results are found. Extensive experiments on three SOTA editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show that ADE-CoT achieves superior performance-efficiency trade-offs. With comparable sampling budgets, ADE-CoT obtains better performance with more than 2x speedup over Best-of-N.

Method: Proposes scale-encoded neural surface representation incorporating pixel coverage area, associating each 3D point with spatial scale and calculating normals from hybrid grid-based encoding. Includes mesh extraction module assigning optimal local scale to each vertex based on training observations.

Result: Method consistently yields high-fidelity surface reconstruction from normals observed at varying distances, outperforming existing multi-view normal integration methods.

Conclusion: Scale-aware approach effectively addresses multi-view normal inconsistency caused by varying camera distances, enabling preservation of high-frequency surface details in reconstruction.

Abstract: Previous multi-view normal integration methods typically sample a single ray per pixel, without considering the spatial area covered by each pixel, which varies with camera intrinsics and the camera-to-object distance. Consequently, when the target object is captured at different distances, the normals at corresponding pixels may differ across views. This multi-view surface normal inconsistency results in the blurring of high-frequency details in the reconstructed surface. To address this issue, we propose a scale-encoded neural surface representation that incorporates the pixel coverage area into the neural representation. By associating each 3D point with a spatial scale and calculating its normal from a hybrid grid-based encoding, our method effectively represents multi-scale surface normals captured at varying distances. Furthermore, to enable scale-aware surface reconstruction, we introduce a mesh extraction module that assigns an optimal local scale to each vertex based on the training observations. Experimental results demonstrate that our approach consistently yields high-fidelity surface reconstruction from normals observed at varying distances, outperforming existing multi-view normal integration methods.

Method: Proposes MV-BrainFM with anatomical distance information incorporated into Transformer-based modeling to guide inter-regional interactions, and introduces unsupervised cross-view consistency learning to align representations from multiple atlases of the same subject in a shared latent space.

Result: Extensive experiments on 20K+ subjects from 17 fMRI datasets show MV-BrainFM consistently outperforms 14 existing brain network foundation models and task-specific baselines under both single-atlas and multi-atlas settings.

Conclusion: MV-BrainFM provides a scalable, atlas-agnostic foundation model for brain network analysis that effectively captures complementary information across heterogeneous network views while maintaining anatomical awareness.

Abstract: Brain network analysis provides an interpretable framework for characterizing brain organization and has been widely used for neurological disorder identification. Recent advances in self-supervised learning have motivated the development of brain network foundation models. However, existing approaches are often limited by atlas dependency, insufficient exploitation of multiple network views, and weak incorporation of anatomical priors. In this work, we propose MV-BrainFM, a multi-view brain network foundation model designed to learn generalizable and scalable representations from brain networks constructed with arbitrary atlases. MV-BrainFM explicitly incorporates anatomical distance information into Transformer-based modeling to guide inter-regional interactions, and introduces an unsupervised cross-view consistency learning strategy to align representations from multiple atlases of the same subject in a shared latent space. By jointly enforcing within-view robustness and cross-view alignment during pretraining, the model effectively captures complementary information across heterogeneous network views while remaining atlas-aware. In addition, MV-BrainFM adopts a unified multi-view pretraining paradigm that enables simultaneous learning from multiple datasets and atlases, significantly improving computational efficiency compared to conventional sequential training strategies. The proposed framework also demonstrates strong scalability, consistently benefiting from increasing data diversity while maintaining stable performance across unseen atlas configurations. Extensive experiments on more than 20K subjects from 17 fMRI datasets show that MV-BrainFM consistently outperforms 14 existing brain network foundation models and task-specific baselines under both single-atlas and multi-atlas settings.

Method: Uni-C leverages video diffusion priors to guide the denoising process of preceding models, aligning their outputs with downstream requirements. It can be used as a plug-and-play module in workflows or independently to enhance individual model outputs.

Result: Extensive experiments across video and 3D generation tasks show Uni-C consistently improves generation quality in both workflow-based and standalone settings.

Conclusion: Uni-C demonstrates versatility and strong generalization capability for improving generative model outputs and addressing distributional mismatches in chained AI workflows.

Abstract: In practical AI workflows, complex tasks often involve chaining multiple generative models, such as using a video or 3D generation model after a 2D image generator. However, distributional mismatches between the output of upstream models and the expected input of downstream models frequently degrade overall generation quality. To address this issue, we propose Uni-Classifier (Uni-C), a simple yet effective plug-and-play module that leverages video diffusion priors to guide the denoising process of preceding models, thereby aligning their outputs with downstream requirements. Uni-C can also be applied independently to enhance the output quality of individual generative models. Extensive experiments across video and 3D generation tasks demonstrate that Uni-C consistently improves generation quality in both workflow-based and standalone settings, highlighting its versatility and strong generalization capability.

Method: Multi-stage fine-tuning of DINOBloom-base model with three classifier head families (linear, cosine, MLP), then constructing ensemble where MLP is primary predictor with confusion pair resolution when two other heads agree

Result: Different heads specialized for different granulocyte classes: cosine best for mature (BNE F1=0.470), linear for immature (MMY F1=0.585), MLP for most immature (PMY F1=0.733). Ensemble improved performance on confusion pairs

Conclusion: Head-diverse ensemble leverages class-specific specialization to address morphological continuum confusion, with consistent misclassifications revealing probable labeling errors or inherent ambiguity

Abstract: The classification of white blood cells (WBCs) from peripheral blood smears is critical for the diagnosis of leukemia. However, automated approaches still struggle due to challenges including class imbalance, domain shift, and morphological continuum confusion, where adjacent maturation stages exhibit subtle, overlapping features. We present a multi-stage fine-tuning methodology for 13-class WBC classification in the WBCBench 2026 Challenge (ISBI 2026). Our best-performing model is a fine-tuned DINOBloom-base, on which we train multiple classifier head families (linear, cosine, and multilayer perceptron (MLP)). The cosine head performed best on the mature granulocyte boundary (Band neutrophil (BNE) F1 = 0.470), the linear head on more immature granulocyte classes (Metamyelocyte (MMY) F1 = 0.585), and the MLP head on the most immature granulocyte (Promyelocyte (PMY) F1 = 0.733), revealing class-specific specialization. Based on this specialization, we construct a head-diverse ensemble, where the MLP head acts as the primary predictor, and its predictions within the four predefined confusion pairs are replaced only when two other head families agree. We further show that cases consistently misclassified by all models are substantially enriched for probable labeling errors or inherent morphological ambiguity.

Method: Two novel components: (1) using vision foundation model embeddings to capture local spatial structure within each patch, and (2) employing graph neural networks with jigsaw regularization to achieve across-patch spatial awareness in tissue images.

Result: The combination of vision foundation model embeddings and jigsaw regularization markedly improves classification performance over state-of-the-art attention-based MIL approaches on benchmark datasets for breast, head-and-neck, and colon cancer.

Conclusion: Incorporating both local patch-level spatial structure (via foundation model embeddings) and global across-patch spatial relationships (via graph neural networks with jigsaw regularization) significantly enhances computational pathology classification by better leveraging the inherent spatial organization of tissue images.

Abstract: Computational pathology involves the digitization of stained tissues into whole-slide images (WSIs) that contain billions of pixels arranged as contiguous patches. Statistical analysis of WSIs largely focuses on classification via multiple instance learning (MIL), in which slide-level labels are inferred from unlabeled patches. Most MIL methods treat patches as exchangeable, overlooking the rich spatial and topological structure that underlies tissue images. This work builds on recent graph-based methods that aim to incorporate spatial awareness into MIL. Our approach is new in two regards: (1) we deploy vision \emph{foundation-model embeddings} to incorporate local spatial structure within each patch, and (2) achieve across-patch spatial awareness using graph neural networks together with a novel {\em jigsaw regularization}. We find that a combination of these two features markedly improves classification over state-of-the-art attention-based MIL approaches on benchmark datasets in breast, head-and-neck, and colon cancer.

Method: Proposes a training-free framework leveraging pretrained single-view HMR models as strong priors. First constructs robust multi-view initialization from single-view predictions, then refines via test-time optimization guided by multi-view consistency and anatomical constraints.

Result: Extensive experiments demonstrate state-of-the-art performance on standard benchmarks, surpassing multi-view models trained with explicit multi-view supervision.

Conclusion: The proposed training-free framework enables calibration-free multi-view human mesh recovery that generalizes to arbitrary camera setups, overcoming limitations of existing geometry-based and learning-based approaches.

Abstract: Multi-view human mesh recovery (HMR) is broadly deployed in diverse domains where high accuracy and strong generalization are essential. Existing approaches can be broadly grouped into geometry-based and learning-based methods. However, geometry-based methods (e.g., triangulation) rely on cumbersome camera calibration, while learning-based approaches often generalize poorly to unseen camera configurations due to the lack of multi-view training data, limiting their performance in real-world scenarios. To enable calibration-free reconstruction that generalizes to arbitrary camera setups, we propose a training-free framework that leverages pretrained single-view HMR models as strong priors, eliminating the need for multi-view training data. Our method first constructs a robust and consistent multi-view initialization from single-view predictions, and then refines it via test-time optimization guided by multi-view consistency and anatomical constraints. Extensive experiments demonstrate state-of-the-art performance on standard benchmarks, surpassing multi-view models trained with explicit multi-view supervision.

Method: FAAR uses Performance-Driven Rank Shrinking (PDRS) to allocate optimal rank per adapter location and task, and Task-Spectral Pyramidal Decoder (TS-PD) that analyzes image frequency spectrum to inject input-specific context into spatial bias learning for cross-task relationships.

Result: FAAR reduces parameters by up to 9x compared to traditional MTL fine-tuning while improving overall performance on dense visual task benchmarks, outperforming other PEFT methods in both accuracy and efficiency.

Conclusion: FAAR provides an effective parameter-efficient fine-tuning approach for multi-task learning that dynamically adapts ranks and captures spatial inter-task relationships through frequency analysis.

Abstract: Adapting models pre-trained on large-scale datasets is a proven way to reach strong performance quickly for down-stream tasks. However, the growth of state-of-the-art mod-els makes traditional full fine-tuning unsuitable and difficult, especially for multi-task learning (MTL) where cost scales with the number of tasks. As a result, recent studies investigate parameter-efficient fine-tuning (PEFT) using low-rank adaptation to significantly reduce the number of trainable parameters. However, these existing methods use a single, fixed rank, which may not be optimal for differ-ent tasks or positions in the MTL architecture. Moreover, these methods fail to learn spatial information that cap-tures inter-task relationships and helps to improve diverse task predictions. This paper introduces Frequency-Aware and Automatic Rank (FAAR) for efficient MTL fine-tuning. Our method introduces Performance-Driven Rank Shrink-ing (PDRS) to allocate the optimal rank per adapter location and per task. Moreover, by analyzing the image frequency spectrum, FAAR proposes a Task-Spectral Pyramidal Decoder (TS-PD) that injects input-specific context into spatial bias learning to better reflect cross-task relationships. Experiments performed on dense visual task benchmarks show the superiority of our method in terms of both accuracy and efficiency compared to other PEFT methods in MTL. FAAR reduces the number of parameters by up to 9 times compared to traditional MTL fine-tuning whilst improving overall performance. Our code is available.

Method: Formally defines PSVU task, creates PEARL-Bench with 132 videos and 2,173 fine-grained annotations, and proposes PEARL - a plug-and-play, training-free strategy for streaming video personalization.

Result: PEARL achieves state-of-the-art performance across 8 offline and online models, brings consistent PSVU improvements to 3 distinct architectures, proving to be effective and robust.

Conclusion: This work advances vision-language model personalization and inspires research into streaming personalized AI assistants, with PEARL serving as a strong baseline for PSVU.

Abstract: Human cognition of new concepts is inherently a streaming process: we continuously recognize new objects or identities and update our memories over time. However, current multimodal personalization methods are largely limited to static images or offline videos. This disconnects continuous visual input from instant real-world feedback, limiting their ability to provide the real-time, interactive personalized responses essential for future AI assistants. To bridge this gap, we first propose and formally define the novel task of Personalized Streaming Video Understanding (PSVU). To facilitate research in this new direction, we introduce PEARL-Bench, the first comprehensive benchmark designed specifically to evaluate this challenging setting. It evaluates a model’s ability to respond to personalized concepts at exact timestamps under two modes: (1) Frame-level, focusing on a specific person or object in discrete frames, and (2) a novel Video-level, focusing on personalized actions unfolding across continuous frames. PEARL-Bench comprises 132 unique videos and 2,173 fine-grained annotations with precise timestamps. Concept diversity and annotation quality are strictly ensured through a combined pipeline of automated generation and human verification. To tackle this challenging new setting, we further propose PEARL, a plug-and-play, training-free strategy that serves as a strong baseline. Extensive evaluations across 8 offline and online models demonstrate that PEARL achieves state-of-the-art performance. Notably, it brings consistent PSVU improvements when applied to 3 distinct architectures, proving to be a highly effective and robust strategy. We hope this work advances vision-language model (VLM) personalization and inspires further research into streaming personalized AI assistants. Code is available at https://github.com/Yuanhong-Zheng/PEARL.

Method: Develops a benchmark for SfM in novel view synthesis using existing datasets. Tests two strategies: 1) using fewer features to accelerate classical SfM, and 2) using feed-forward neural networks for initial estimates followed by classical refinement.

Result: 1) Fewer features significantly accelerate classical SfM while maintaining high pose accuracy. 2) Neural network initialization with classical refinement provides the best efficiency-effectiveness trade-off.

Conclusion: The benchmark demonstrates practical strategies for efficient SfM that maintain accuracy, with hybrid approaches offering optimal performance for novel view synthesis applications.

Abstract: Novel view synthesis (NVS) approaches such as NeRFs or 3DGS can produce photo-realistic 3D scene representation from a set of images with known extrinsic and intrinsic parameters. The necessary camera poses and calibrations are typically obtained from the images via Structure-from-Motion (SfM). Classical SfM approaches rely on local feature matches between the images to estimate both the poses and a sparse 3D model of the scene, using bundle adjustment to refine initial pose, intrinsics, and geometry estimates. In order to increase run-time efficiency, recent SfM systems forgo optimization via bundle adjustment. Instead, they train feed-forward (transformer-based) neural networks to directly regress camera parameters and the 3D structure. While orders of magnitude more efficient, such recent works produce significantly less accurate estimates. To stimulate research on developing SfM approaches that are both efficient \emph{and} effective, this paper develops a benchmark focused on SfM for novel view synthesis. Using existing datasets and two simple strategies for making the reconstruction process more efficient, we show that: (1) simply using fewer features already significantly accelerates classical SfM methods while maintaining high pose accuracy. (2) using feed-forward networks to obtain initial estimates and refining them using classical SfM techniques leads to the best efficiency-effectiveness trade-off. We will make our benchmark and code publicly available.

Method: Two general methods: 1) Forward pass method using root-finding algorithm and Jacobian with respect to input images, 2) Backward pass method that iteratively inverts each layer from the top, adding random vectors sampled from the null-space of each linear layer.

Result: The methods successfully produce random-like input images that yield near-perfect classification scores on both transformer architectures and sequential networks based on linear layers, revealing vulnerabilities in the underlying networks.

Conclusion: The proposed methods provide more comprehensive coverage of input image spaces that solve the inverse mapping problem, demonstrating that networks can be fooled by seemingly random images that achieve high classification confidence.

Abstract: Neural network systems describe complex mappings that can be very difficult to understand. In this paper, we study the inverse problem of determining the input images that get mapped to specific neural network classes. Ultimately, we expect that these images contain recognizable features that are associated with their corresponding class classifications. We introduce two general methods for solving the inverse problem. In our forward pass method, we develop an inverse method based on a root-finding algorithm and the Jacobian with respect to the input image. In our backward pass method, we iteratively invert each layer, at the top. During the inversion process, we add random vectors sampled from the null-space of each linear layer. We demonstrate our new methods on both transformer architectures and sequential networks based on linear layers. Unlike previous methods, we show that our new methods are able to produce random-like input images that yield near perfect classification scores in all cases, revealing vulnerabilities in the underlying networks. Hence, we conclude that the proposed methods provide a more comprehensive coverage of the input image spaces that solve the inverse mapping problem.

Method: CREG projects multi-layer contrastive Grad-times-Act attributions into a reference-centered polar coordinate system, creating directional evidence distributions over compass sectors. It’s training-free and uses three evaluation metrics: Direction Alignment Error, Edge Accuracy, and Causal Occlusion Score.

Result: On Qwen2-VL-7B across VSR and COCO-Pairs, CREG consistently outperforms standard attribution baselines. On COCO-Pairs, it achieves DAE of 55.5° and EA of 0.553, improving over attention rollout by 16.1° in angular error and 0.120 in EA. Causal occlusion experiments show COS ≥ +0.42, supporting faithfulness.

Conclusion: Contrastive, multi-layer attribution can expose directional evidence more faithfully than standard saliency-based explanations in VLM spatial reasoning. CREG benefits from more structured spatial representations that emerge at larger model scales.

Abstract: Vision-language models (VLMs) perform strongly on spatial reasoning benchmarks, yet how they encode directional relations remains poorly understood. Existing attribution methods such as GradCAM and attention rollout reveal where a model attends, but not what direction it infers between objects. We introduce CREG (Compass Relational Evidence Graph), a training-free interpretability framework that projects multi-layer contrastive Grad-times-Act attributions into a reference-centered polar coordinate system, producing a directional evidence distribution over compass sectors. To evaluate directional explanations, we propose three metrics: Direction Alignment Error (DAE), Edge Accuracy (EA), and Causal Occlusion Score (COS). On Qwen2-VL-7B across VSR and COCO-Pairs, CREG consistently outperforms standard attribution baselines; on COCO-Pairs, prediction-targeted CREG achieves a DAE of 55.5 degrees and an EA of 0.553, improving over attention rollout by 16.1 degrees in angular error and 0.120 in EA. Causal occlusion experiments on 540 samples across both datasets further support the faithfulness of these directional explanations, with COS greater than or equal to +0.42. The gains are smaller on Qwen2-VL-2B, suggesting that CREG benefits from the more structured spatial representations that emerge at larger scales. Overall, our results show that contrastive, multi-layer attribution can expose directional evidence more faithfully than standard saliency-based explanations in VLM spatial reasoning.

Method: Created realistic benchmark with 546 camera traps using streaming protocol evaluating models over chronologically ordered intervals; studied site-specific adaptation, model updating with past data, and identified drivers of difficulty including class imbalance and temporal shifts.

Result: Found biological foundation models underperform at many sites; adaptation is challenging and can degrade performance below zero-shot; identified class imbalance and temporal shifts as key difficulties; effective integration of model-update and post-processing improves accuracy but gaps remain.

Conclusion: Provides actionable deployment guidelines for ecological practitioners and establishes new research directions for vision and ML, highlighting need for better temporal adaptation methods and understanding when zero-shot models succeed.

Abstract: Camera traps are vital for large-scale biodiversity monitoring, yet accurate automated analysis remains challenging due to diverse deployment environments. While the computer vision community has mostly framed this challenge as cross-domain generalization, this perspective overlooks a primary challenge faced by ecological practitioners: maintaining reliable recognition at the fixed site over time, where the dynamic nature of ecosystems introduces profound temporal shifts in both background and animal distributions. To bridge this gap, we present the first unified study of camera-trap species recognition over time. We introduce a realistic benchmark comprising 546 camera traps with a streaming protocol that evaluates models over chronologically ordered intervals. Our end-user-centric study yields four key findings. (1) Biological foundation models (e.g., BioCLIP 2) underperform at numerous sites even in initial intervals, underscoring the necessity of site-specific adaptation. (2) Adaptation is challenging under realistic evaluation: when models are updated using past data and evaluated on future intervals (mirrors real deployment lifecycles), naive adaptation can even degrade below zero-shot performance. (3) We identify two drivers of this difficulty: severe class imbalance and pronounced temporal shift in both species distribution and backgrounds between consecutive intervals. (4) We find that effective integration of model-update and post-processing techniques can largely improve accuracy, though a gap from the upper bounds remains. Finally, we highlight critical open questions, such as predicting when zero-shot models will succeed at a new site and determining whether/when model updates are necessary. Our benchmark and analysis provide actionable deployment guidelines for ecological practitioners while establishing new directions for future research in vision and machine learning.

Method: Proposes an end-to-end optimization framework that jointly learns a material classifier and determines optimal combinations of rotation angles for polarization elements controlling both incident and reflected light states. Uses Mueller-matrix material dataset.

Result: Demonstrates high-accuracy material classification even with a limited number of measurements.

Conclusion: The proposed framework enables efficient material classification by optimizing polarization measurement angles, reducing measurement time while maintaining accuracy.

Abstract: Material classification is a fundamental problem in computer vision and plays a crucial role in scene understanding. Previous studies have explored various material recognition methods based on reflection properties such as color, texture, specularity, and scattering. Among these cues, polarization is particularly valuable because it provides rich material information and enables recognition even at distances where capturing high-resolution texture is impractical. However, measuring polarimetric reflectance properties typically requires multiple modulations of the polarization state of the incident light, making the process time-consuming and often unnecessary for certain recognition tasks. While material classification can be achieved using only a subset of polarimetric measurements, the optimal configuration of measurement angles remains unclear. In this study, we propose an end-to-end optimization framework that jointly learns a material classifier and determines the optimal combinations of rotation angles for polarization elements that control both the incident and reflected light states. Using our Mueller-matrix material dataset, we demonstrate that our method achieves high-accuracy material classification even with a limited number of measurements.

Method: 1) Identifies limitations of existing retrieval-based evaluation metrics; 2) Proposes Multimodal LLMs-as-a-judge evaluation framework using yes/no questions about image content; 3) Investigates whether negation direction exists in CLIP embedding space; 4) Uses representation engineering for test-time intervention to steer CLIP toward negation-aware behavior without fine-tuning; 5) Tests generalization on non-common image-text samples.

Result: Found evidence that a negation direction exists in CLIP’s embedding space. Showed that representation engineering can manipulate this direction to improve negation understanding without fine-tuning. Demonstrated the proposed MLLM-based evaluation provides fair assessment of negation understanding compared to flawed retrieval metrics.

Conclusion: Negation understanding in CLIP can be improved through test-time intervention using discovered negation directions in embedding space, bypassing need for large-scale fine-tuning. New MLLM-based evaluation framework provides more reliable assessment of negation understanding capabilities.

Abstract: Joint Vision-Language Embedding models such as CLIP typically fail at understanding negation in text queries - for example, failing to distinguish “no” in the query: “a plain blue shirt with no logos”. Prior work has largely addressed this limitation through data-centric approaches, fine-tuning CLIP on large-scale synthetic negation datasets. However, these efforts are commonly evaluated using retrieval-based metrics that cannot reliably reflect whether negation is actually understood. In this paper, we identify two key limitations of such evaluation metrics and investigate an alternative evaluation framework based on Multimodal LLMs-as-a-judge, which typically excel at understanding simple yes/no questions about image content, providing a fair evaluation of negation understanding in CLIP models. We then ask whether there already exists a direction in the CLIP embedding space associated with negation. We find evidence that such a direction exists, and show that it can be manipulated through test-time intervention via representation engineering to steer CLIP toward negation-aware behavior without any fine-tuning. Finally, we test negation understanding on non-common image-text samples to evaluate generalization under distribution shifts.

Method: First conducts synthetic comparisons to isolate effective regimes of CFG and AG from weak-to-strong principle perspective. Proposes SGG as hybrid instantiation combining benefits of both. Also migrates W2S principle into training objective to improve unguided diffusion models.

Result: SGG outperforms existing training-free guidance variants on SD3 and SD3.5. Training-time experiments on transformer architectures show effective migration and performance gains in both conditional and unconditional settings.

Conclusion: SGG provides effective hybrid guidance for diffusion models, addressing gradient error accumulation and improving generalization through weak-to-strong principle integration.

Abstract: Diffusion models generate synthetic images through an iterative refinement process. However, the misalignment between the simulation-free objective and the iterative process often causes accumulated gradient error along the sampling trajectory, which leads to unsatisfactory results and a failure to generalize. Guidance techniques like Classifier Free Guidance (CFG) and AutoGuidance (AG) alleviate this by extrapolating between the main and inferior signal for stronger generalization. Despite empirical success, the effective operational regimes of prevalent guidance methods are still under-explored, leading to ambiguity when selecting the appropriate guidance method given a precondition. In this work, we first conduct synthetic comparisons to isolate and demonstrate the effective regime of guidance methods represented by CFG and AG from the perspective of weak-to-strong principle. Based on this, we propose a hybrid instantiation called SGG under the principle, taking the benefits of both. Furthermore, we demonstrate that the W2S principle along with SGG can be migrated into the training objective, improving the generalization ability of unguided diffusion models. We validate our approach with comprehensive experiments. At inference time, evaluations on SD3 and SD3.5 confirm that SGG outperforms existing training-free guidance variants. Training-time experiments on transformer architectures demonstrate the effective migration and performance gains in both conditional and unconditional settings. Code is available at https://github.com/851695e35/SGG.

Method: Proposes RayMap3R with dual-branch inference: identifies dynamic regions by contrasting RayMap predictions (which exhibit static-scene bias) with image predictions, suppresses dynamic interference during memory updates, and introduces reset metric alignment and state-aware smoothing for consistency.

Result: Achieves state-of-the-art performance among streaming approaches on dynamic scene reconstruction across multiple benchmarks.

Conclusion: RayMap3R provides an effective training-free solution for dynamic scene reconstruction in streaming settings by leveraging inherent static-scene bias in RayMap predictions for dynamic identification.

Abstract: Streaming feed-forward 3D reconstruction enables real-time joint estimation of scene geometry and camera poses from RGB images. However, without explicit dynamic reasoning, streaming models can be affected by moving objects, causing artifacts and drift. In this work, we propose RayMap3R, a training-free streaming framework for dynamic scene reconstruction. We observe that RayMap-based predictions exhibit a static-scene bias, providing an internal cue for dynamic identification. Based on this observation, we construct a dual-branch inference scheme that identifies dynamic regions by contrasting RayMap and image predictions, suppressing their interference during memory updates. We further introduce reset metric alignment and state-aware smoothing to preserve metric consistency and stabilize predicted trajectories. Our method achieves state-of-the-art performance among streaming approaches on dynamic scene reconstruction across multiple benchmarks.

Method: Three key innovations: 1) slice-aware piling strategy positioning anisotropic 3D Gaussians, 2) differentiable projection operator encoding finite-thickness point spread function, 3) compact encoding and joint optimization pipeline for simultaneous reconstruction and compression.

Result: Reduces storage and reconstruction costs, sustains diagnostic fidelity, enables fast 2D visualization and 3D voxelization. Delivers results 11x faster than NeRF-based approaches with 16x compression over voxel grids.

Conclusion: GaussianPile offers practical path to deployable compression and exploration of slice-based volumetric datasets with high efficiency and quality preservation.

Abstract: Slice-based volumetric imaging is widely applied and it demands representations that compress aggressively while preserving internal structure for analysis. We introduce GaussianPile, unifying 3D Gaussian splatting with an imaging system-aware focus model to address this challenge. Our proposed method introduces three key innovations: (i) a slice-aware piling strategy that positions anisotropic 3D Gaussians to model through-slice contributions, (ii) a differentiable projection operator that encodes the finite-thickness point spread function of the imaging acquisition system, and (iii) a compact encoding and joint optimization pipeline that simultaneously reconstructs and compresses the Gaussian sets. Our CUDA-based design retains the compression and real-time rendering efficiency of Gaussian primitives while preserving high-frequency internal volumetric detail. Experiments on microscopy and ultrasound datasets demonstrate that our method reduces storage and reconstruction cost, sustains diagnostic fidelity, and enables fast 2D visualization, along with 3D voxelization. In practice, it delivers high-quality results in as few as 3 minutes, up to 11x faster than NeRF-based approaches, and achieves consistent 16x compression over voxel grids, offering a practical path to deployable compression and exploration of slice-based volumetric datasets.

Method: Three-component pipeline: 1) source image expansion with world-knowledge infusion, 2) adaptive multi-agent editing instruction-image synthesis, and 3) task-aware data quality verification mechanism.

Result: Created ScaleEdit-12M dataset spanning 23 task families; fine-tuning UniWorld-V1 and Bagel models yielded up to 10.4% improvement on ImgEdit, 35.1% on GEdit, and up to 150.0% on knowledge-infused benchmarks.

Conclusion: Open-source agentic pipelines can approach commercial-grade data quality while maintaining cost-effectiveness and scalability for multimodal model training.

Abstract: Instruction-based image editing has emerged as a key capability for unified multimodal models (UMMs), yet constructing large-scale, diverse, and high-quality editing datasets without costly proprietary APIs remains challenging. Previous image editing datasets either rely on closed-source models for annotation, which prevents cost-effective scaling, or employ fixed synthetic editing pipelines, which suffer from limited quality and generalizability. To address these challenges, we propose ScaleEditor, a fully open-source hierarchical multi-agent framework for end-to-end construction of large-scale, high-quality image editing datasets. Our pipeline consists of three key components: source image expansion with world-knowledge infusion, adaptive multi-agent editing instruction-image synthesis, and a task-aware data quality verification mechanism. Using ScaleEditor, we curate ScaleEdit-12M, the largest open-source image editing dataset to date, spanning 23 task families across diverse real and synthetic domains. Fine-tuning UniWorld-V1 and Bagel on ScaleEdit yields consistent gains, improving performance by up to 10.4% on ImgEdit and 35.1% on GEdit for general editing benchmarks and by up to 150.0% on RISE and 26.5% on KRIS-Bench for knowledge-infused benchmarks. These results demonstrate that open-source, agentic pipelines can approach commercial-grade data quality while retaining cost-effectiveness and scalability. Both the framework and dataset will be open-sourced.

Method: Proposes MCL-FIR: multi-head design for evolving classes, reformulates triplets into doublets with InfoNCE for simpler training, and uses EMA distillation for efficient knowledge transfer.

Result: Outperforms CIL baselines with similar training cost, achieves comparable performance to static methods using only ~30% training cost across four datasets.

Conclusion: MCL-FIR provides scalable, efficient solution for dynamic fashion retrieval scenarios with strong balance between efficiency and accuracy.

Abstract: Most fine-grained fashion image retrieval (FIR) methods assume a static setting, requiring full retraining when new attributes appear, which is costly and impractical for dynamic scenarios. Although pretrained models support zero-shot inference, their accuracy drops without supervision, and no prior work explores class-incremental learning (CIL) for fine-grained FIR. We propose a multihead continual learning framework for fine-grained fashion image retrieval with contrastive learning and exponential moving average (EMA) distillation (MCL-FIR). MCL-FIR adopts a multi-head design to accommodate evolving classes across increments, reformulates triplet inputs into doublets with InfoNCE for simpler and more effective training, and employs EMA distillation for efficient knowledge transfer. Experiments across four datasets demonstrate that, beyond its scalability, MCL-FIR achieves a strong balance between efficiency and accuracy. It significantly outperforms CIL baselines under similar training cost, and compared with static methods, it delivers comparable performance while using only about 30% of the training cost. The source code is publicly available in https://github.com/Dr-LingXiao/MCL-FIR.

Method: Proposes IBCapsNet using Information Bottleneck principle with one-pass variational aggregation: primary capsules compressed into global context, then processed by class-specific VAEs to infer latent capsules regularized by KL divergence.

Result: Matches CapsNet accuracy on clean data (99.41% MNIST, 92.01% SVHN), significantly outperforms under noise (+17.10% clamped additive, +14.54% multiplicative), 2.54x faster training, 3.64x higher inference throughput, 4.66% fewer parameters.

Conclusion: Bridges information-theoretic representation learning with capsule networks for robust, efficient, interpretable deep models.

Abstract: Capsule networks (CapsNets) are superior at modeling hierarchical spatial relationships but suffer from two critical limitations: high computational cost due to iterative dynamic routing and poor robustness under input corruptions. To address these issues, we propose IBCapsNet, a novel capsule architecture grounded in the Information Bottleneck (IB) principle. Instead of iterative routing, IBCapsNet employs a one-pass variational aggregation mechanism, where primary capsules are first compressed into a global context representation and then processed by class-specific variational autoencoders (VAEs) to infer latent capsules regularized by the KL divergence. This design enables efficient inference while inherently filtering out noise. Experiments on MNIST, Fashion-MNIST, SVHN and CIFAR-10 show that IBCapsNet matches CapsNet in clean-data accuracy (achieving 99.41% on MNIST and 92.01% on SVHN), yet significantly outperforms it under four types of synthetic noise - demonstrating average improvements of +17.10% and +14.54% for clamped additive and multiplicative noise, respectively. Moreover, IBCapsNet achieves 2.54x faster training and 3.64x higher inference throughput compared to CapsNet, while reducing model parameters by 4.66%. Our work bridges information-theoretic representation learning with capsule networks, offering a principled path toward robust, efficient, and interpretable deep models. Code is available at https://github.com/cxiang26/IBCapsnet

Method: Uses MeanFlow as learning target to approximate average velocity between arbitrary PF-ODE states, capturing teacher dynamics without explicit rollouts. Improves CFG formulation with teacher CFG distillation to better leverage pretrained generative priors.

Result: Achieves efficient, flexible, high-quality super-resolution on synthetic and real-world benchmarks, matching or exceeding multi-step teacher performance with much lower computational cost.

Conclusion: MFSR provides a practical distillation framework that balances efficiency and quality for real-world image super-resolution, enabling single-step inference while preserving optional refinement capability.

Abstract: Diffusion- and flow-based models have advanced Real-world Image Super-Resolution (Real-ISR), but their multi-step sampling makes inference slow and hard to deploy. One-step distillation alleviates the cost, yet often degrades restoration quality and removes the option to refine with more steps. We present Mean Flows for Super-Resolution (MFSR), a new distillation framework that produces photorealistic results in a single step while still allowing an optional few-step path for further improvement. Our approach uses MeanFlow as the learning target, enabling the student to approximate the average velocity between arbitrary states of the Probability Flow ODE (PF-ODE) and effectively capture the teacher’s dynamics without explicit rollouts. To better leverage pretrained generative priors, we additionally improve original MeanFlow’s Classifier-Free Guidance (CFG) formulation with teacher CFG distillation strategy, which enhances restoration capability and preserves fine details. Experiments on both synthetic and real-world benchmarks demonstrate that MFSR achieves efficient, flexible, and high-quality super-resolution, delivering results on par with or even better than multi-step teachers while requiring much lower computational cost.

Method: Two generative strategies: VLM-guided approach and damage-sensitive Mixture-of-Experts method. Benchmark against Pix2Pix and ControlNet using Structure-Aware Evaluation Framework with pixel-level assessment, ResNet-based semantic consistency, and VLM-as-a-Judge.

Result: Experiments on 300 disaster scenarios reveal realism-fidelity trade-off: diffusion-based approaches achieve high perceptual realism but hallucinate details. ControlNet achieves highest semantic accuracy (0.71), while VLM-enhanced and MoE models excel in textural plausibility but struggle with semantic clarity.

Conclusion: Establishes baseline for trustworthy cross-view synthesis, emphasizing that visually realistic generations may fail to preserve critical structural information needed for reliable disaster assessment.

Abstract: In the immediate aftermath of natural disasters, rapid situational awareness is critical. Traditionally, satellite observations are widely used to estimate damage extent. However, they lack the ground-level perspective essential for characterizing specific structural failures and impacts. Meanwhile, ground-level data (e.g., street-view imagery) remains largely inaccessible during time-sensitive events. This study investigates Satellite-to-Street View Synthesis to bridge this data gap. We introduce two generative strategies to synthesize post-disaster street views from satellite imagery: a Vision-Language Model (VLM)-guided approach and a damage-sensitive Mixture-of-Experts (MoE) method. We benchmark these against general-purpose baselines (Pix2Pix, ControlNet) using a proposed Structure-Aware Evaluation Framework. This multi-tier protocol integrates (1) pixel-level quality assessment, (2) ResNet-based semantic consistency verification, and (3) a novel VLM-as-a-Judge for perceptual alignment. Experiments on 300 disaster scenarios reveal a critical realism–fidelity trade-off: while diffusion-based approaches (e.g., ControlNet) achieve high perceptual realism, they often hallucinate structural details. Quantitative results show that standard ControlNet achieves the highest semantic accuracy, 0.71, whereas VLM-enhanced and MoE models excel in textural plausibility but struggle with semantic clarity. This work establishes a baseline for trustworthy cross-view synthesis, emphasizing that visually realistic generations may still fail to preserve critical structural information required for reliable disaster assessment.

Method: 1) Construct hierarchical clinical cognition dataset and use Supervised Fine-Tuning to internalize expert diagnostic logic (anatomical localization → morphological evaluation → microvascular analysis). 2) Counterfactual-driven reinforcement learning strategy using lesion masking to generate counterfactual normal samples and clinical-cognition-centric rewards to enforce causal rectification.

Result: Achieves State-of-the-Art performance across multiple benchmarks, significantly enhancing diagnostic accuracy in complex clinical scenarios.

Conclusion: The CogAlign framework successfully addresses clinical-cognitive misalignment and visual bias in medical MLLMs, providing a robust approach for gastrointestinal endoscopy analysis with publicly available code and datasets.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated remarkable potential in medical image analysis. However, their application in gastrointestinal endoscopy is currently hindered by two critical limitations: the misalignment between general model reasoning and standardized clinical cognitive pathways, and the lack of causal association between visual features and diagnostic outcomes. In this paper, we propose a novel Clinical-Cognitive-Aligned (CogAlign) framework to address these challenges. First, we endow the model with rigorous clinical analytical capabilities by constructing the hierarchical clinical cognition dataset and employing Supervised Fine-Tuning (SFT). Unlike conventional approaches, this strategy internalizes the hierarchical diagnostic logic of experts, ranging from anatomical localization and morphological evaluation to microvascular analysis, directly into the model. Second, to eliminate visual bias, we provide a theoretical analysis demonstrating that standard supervised tuning inevitably converges to spurious background correlations. Guided by this insight, we propose a counterfactual-driven reinforcement learning strategy to enforce causal rectification. By generating counterfactual normal samples via lesion masking and optimizing through clinical-cognition-centric rewards, we constrain the model to strictly ground its diagnosis in causal lesion features. Extensive experiments demonstrate that our approach achieves State-of-the-Art (SoTA) performance across multiple benchmarks, significantly enhancing diagnostic accuracy in complex clinical scenarios. All source code and datasets will be made publicly available.

Method: Two complementary modules: 1) polarity-weighted gradients for scene refinement using turbulence-induced event polarity alternation patterns, and 2) event-tube constraints for motion decoupling using spatiotemporally coherent event patterns from dynamic objects.

Result: Outperforms state-of-the-art methods, especially with dynamic objects, while reducing data overhead by ~77.3% and system latency by ~89.5%. Two real-world event-frame turbulence datasets created.

Conclusion: Event cameras enable high-quality turbulence mitigation with few frames by exploiting event polarity patterns and motion priors, overcoming traditional accuracy-efficiency tradeoffs.

Abstract: Turbulence mitigation (TM) is highly ill-posed due to the stochastic nature of atmospheric turbulence. Most methods rely on multiple frames recorded by conventional cameras to capture stable patterns in natural scenarios. However, they inevitably suffer from a trade-off between accuracy and efficiency: more frames enhance restoration at the cost of higher system latency and larger data overhead. Event cameras, equipped with microsecond temporal resolution and efficient sensing of dynamic changes, offer an opportunity to break the bottleneck. In this work, we present EHETM, a high-quality and efficient TM method inspired by the superiority of events to model motions in continuous sequences. We discover two key phenomena: (1) turbulence-induced events exhibit distinct polarity alternation correlated with sharp image gradients, providing structural cues for restoring scenes; and (2) dynamic objects form spatiotemporally coherent ``event tubes’’ in contrast to irregular patterns within turbulent events, providing motion priors for disentangling objects from turbulence. Based on these insights, we design two complementary modules that respectively leverage polarity-weighted gradients for scene refinement and event-tube constraints for motion decoupling, achieving high-quality restoration with few frames. Furthermore, we construct two real-world event-frame turbulence datasets covering atmospheric and thermal cases. Experiments show that EHETM outperforms SOTA methods, especially under scenes with dynamic objects, while reducing data overhead and system latency by approximately 77.3% and 89.5%, respectively. Our code is available at: https://github.com/Xavier667/EHETM.

Method: Created a new benchmark to systematically study combinations of different initialization types (dense laser scans, dense multi-view stereo point clouds, dense monocular depth estimates, sparse SfM point clouds) with various densification schemes. Analyzed how well current densification approaches can improve reconstruction quality when starting from different initialization densities.

Result: Current densification approaches are unable to take full advantage of dense initialization data and often fail to significantly improve over sparse SfM-based initialization. The benchmark reveals limitations in existing densification schemes when provided with higher-quality initial point clouds.

Conclusion: There is a need for improved densification methods that can better leverage dense initialization data to achieve higher quality 3D reconstructions with 3D Gaussian Splatting.

Abstract: 3D Gaussian Splatting (3DGS) has become the method of choice for photo-realistic 3D reconstruction of scenes, due to being able to efficiently and accurately recover the scene appearance and geometry from images. 3DGS represents the scene through a set of 3D Gaussians, parameterized by their position, spatial extent, and view-dependent color. Starting from an initial point cloud, 3DGS refines the Gaussians’ parameters as to reconstruct a set of training images as accurately as possible. Typically, a sparse Structure-from-Motion point cloud is used as initialization. In order to obtain dense Gaussian clouds, 3DGS methods thus rely on a densification stage. In this paper, we systematically study the relation between densification and initialization. Proposing a new benchmark, we study combinations of different types of initializations (dense laser scans, dense (multi-view) stereo point clouds, dense monocular depth estimates, sparse SfM point clouds) and different densification schemes. We show that current densification approaches are not able to take full advantage of dense initialization as they are often unable to (significantly) improve over sparse SfM-based initialization. We will make our benchmark publicly available.

Method: Proposes a Cross-modal Fuzzy Alignment Network with Fuzzy Token Alignment module using fuzzy membership functions to model token-level associations, and Context-Aware Dynamic Alignment module incorporating ground-view images as bridge agents for adaptive alignment.

Result: Experiments on the constructed AERI-PEDES dataset and TBAPR benchmark demonstrate the superiority of the proposed method.

Conclusion: The approach effectively addresses the challenges of text-aerial person retrieval by improving semantic alignment through fuzzy logic and bridge agents, with validation on new benchmark datasets.

Abstract: Text-aerial person retrieval aims to identify targets in UAV-captured images from eyewitness descriptions, supporting intelligent transportation and public security applications. Compared to ground-view text–image person retrieval, UAV-captured images often suffer from degraded visual information due to drastic variations in viewing angles and flight altitudes, making semantic alignment with textual descriptions very challenging. To address this issue, we propose a novel Cross-modal Fuzzy Alignment Network, which quantifies the token-level reliability by fuzzy logic to achieve accurate fine-grained alignment and incorporates ground-view images as a bridge agent to further mitigate the gap between aerial images and text descriptions, for text–aerial person retrieval. In particular, we design the Fuzzy Token Alignment module that employs the fuzzy membership function to dynamically model token-level association strength and suppress the influence of unobservable or noisy tokens. It can alleviate the semantic inconsistencies caused by missing visual cues and significantly enhance the robustness of token-level semantic alignment. Moreover, to further mitigate the gap between aerial images and text descriptions, we design a Context-Aware Dynamic Alignment module to incorporate the ground-view agent as a bridge in text–aerial alignment and adaptively combine direct alignment and agent-assisted alignment to improve the robustness. In addition, we construct a large-scale benchmark dataset called AERI-PEDES by using a chain-of-thought to decompose text generation into attribute parsing, initial captioning, and refinement, thus boosting textual accuracy and semantic consistency. Experiments on AERI-PEDES and TBAPR demonstrate the superiority of our method.

Method: Premier represents each user’s preference as a learnable embedding and introduces a preference adapter that fuses user embeddings with text prompts. It modulates the generative process with fused preference embeddings and uses a dispersion loss to enhance distinctness of individual preferences and improve alignment with user-specific styles. For new users with scarce data, it represents them as linear combinations of existing preference embeddings.

Result: Experiments show Premier outperforms prior methods under the same history length, achieving stronger preference alignment and superior performance on text consistency, ViPer proxy metrics, and expert evaluations.

Conclusion: Premier provides an effective framework for personalized image generation that better captures and reflects user preferences through learnable embeddings and preference modulation.

Abstract: Text-to-image generation has advanced rapidly, yet it still struggles to capture the nuanced user preferences. Existing approaches typically rely on multimodal large language models to infer user preferences, but the derived prompts or latent codes rarely reflect them faithfully, leading to suboptimal personalization. We present Premier, a novel preference modulation framework for personalized image generation. Premier represents each user’s preference as a learnable embedding and introduces a preference adapter that fuses the user embedding with the text prompt. To enable accurate and fine-grained preference control, the fused preference embedding is further used to modulate the generative process. To enhance the distinctness of individual preference and improve alignment between outputs and user-specific styles, we incorporate a dispersion loss that enforces separation among user embeddings. When user data are scarce, new users are represented as linear combinations of existing preference embeddings learned during training, enabling effective generalization. Experiments show that Premier outperforms prior methods under the same history length, achieving stronger preference alignment and superior performance on text consistency, ViPer proxy metrics, and expert evaluations.

Method: Weakly supervised multimodal segmentation framework that refines threshold-guided pseudo-labels through learned models. Preserves annotation-free character of classical thresholding/clustering while adding denoising, confidence-aware pseudo-supervision, and physically structured fusion. Compares different fusion strategies: direct concatenation, depth-aware cross-attention, gated fusion, and confidence-aware modulation.

Result: Threshold-guided learned refinement provides most robust improvement over baselines. Best performing model is confidence-gated depth-aware cross-attention (CG-DCA), which consistently outperforms threshold-based, image-only, and earlier multimodal baselines. Performance depends on confidence-aware fusion and structured local depth interaction rather than model complexity alone. Cross-well analyses confirm stable performance.

Conclusion: Establishes practical, scalable framework for annotation-free segmentation showing multimodal improvement is maximized when auxiliary logs are incorporated selectively and depth-aware. Confidence-gated depth-aware cross-attention provides optimal performance for combining 2D image texture with 1D well-logs in weakly supervised setting.

Abstract: Acoustic borehole images provide high-resolution borehole-wall structure, but large-scale interpretation remains difficult because dense expert annotations are rarely available and subsurface information is intrinsically multimodal. The challenge is developing weakly supervised methods combining two-dimensional image texture with depth-aligned one-dimensional well-logs. Here, we introduce a weakly supervised multimodal segmentation framework that refines threshold-guided pseudo-labels through learned models. This preserves the annotation-free character of classical thresholding and clustering workflows while extending them with denoising, confidence-aware pseudo-supervision, and physically structured fusion. We establish that threshold-guided learned refinement provides the most robust improvement over raw thresholding, denoised thresholding, and latent clustering baselines. Multimodal performance depends strongly on fusion strategy: direct concatenation provides limited gains, whereas depth-aware cross-attention, gated fusion, and confidence-aware modulation substantially improve agreement with the weak supervisory reference. The strongest model, confidence-gated depth-aware cross-attention (CG-DCA), consistently outperforms threshold-based, image-only, and earlier multimodal baselines. Targeted ablations show its advantage depends specifically on confidence-aware fusion and structured local depth interaction rather than model complexity alone. Cross-well analyses confirm this performance is broadly stable. These results establish a practical, scalable framework for annotation-free segmentation, showing multimodal improvement is maximized when auxiliary logs are incorporated selectively and depth-aware.

Method: Proposes VSD-MOT framework with CLIP Image Encoder as teacher model for visual semantic extraction. Uses Dual-Constraint Semantic Distillation (DCSD) to transfer semantic extraction capabilities to student model. Includes Dynamic Semantic Weight Regulation (DSWR) module to adaptively allocate fusion weights based on real-time frame quality assessment.

Result: Extensive experiments demonstrate effectiveness and superiority in low-quality video scenarios while maintaining good performance in conventional scenarios.

Conclusion: The proposed VSD-MOT framework successfully addresses multi-object tracking challenges in low-quality videos through visual semantic distillation and dynamic weight regulation, improving robustness in real-world scenarios.

Abstract: Existing multi-object tracking algorithms typically fail to adequately address the issues in low-quality videos, resulting in a significant decline in tracking performance when image quality deteriorates in real-world scenarios. This performance degradation is primarily due to the algorithms’ inability to effectively tackle the problems caused by information loss in low-quality images. To address the challenges of low-quality video scenarios, inspired by vision-language models, we propose a multi-object tracking framework guided by visual semantic distillation (VSD-MOT). Specifically, we introduce the CLIP Image Encoder to extract global visual semantic information from images to compensate for the loss of information in low-quality images. However, direct integration can substantially impact the efficiency of the multi-object tracking algorithm. Therefore, this paper proposes to extract visual semantic information from images through knowledge distillation. This method adopts a teacher-student learning framework, with the CLIP Image Encoder serving as the teacher model. To enable the student model to acquire the capability of extracting visual semantic information suitable for multi-object tracking tasks from the teacher model, we have designed the Dual-Constraint Semantic Distillation method (DCSD). Furthermore, to address the dynamic variation of frame quality in low-quality videos, we propose the Dynamic Semantic Weight Regulation (DSWR) module, which adaptively allocates fusion weights based on real-time frame quality assessment. Extensive experiments demonstrate the effectiveness and superiority of the proposed method in low-quality video scenarios in the real world. Meanwhile, our method can maintain good performance in conventional scenarios.

Method: SATTC (Structure-Aware Test-Time Calibration) is a label-free calibration head that operates on similarity matrices of frozen EEG and image encoders. It combines: 1) Geometric expert with subject-adaptive whitening of EEG embeddings and adaptive CSLS, and 2) Structural expert using mutual nearest neighbors, bidirectional top-k ranks, and class popularity, fused via Product-of-Experts rule.

Result: On THINGS-EEG dataset with leave-one-subject-out protocol, SATTC improves Top-1 and Top-5 accuracy, reduces hubness and per-class imbalance, and produces more reliable small-k shortlists compared to baseline methods. Gains transfer across multiple EEG encoders.

Conclusion: SATTC serves as an encoder-agnostic, label-free test-time calibration layer for cross-subject neural decoding, effectively addressing subject shift and hubness problems in EEG-to-image retrieval.

Abstract: Cross-subject EEG-to-image retrieval for visual decoding is challenged by subject shift and hubness in the embedding space, which distort similarity geometry and destabilize top-k rankings, making small-k shortlists unreliable. We introduce SATTC (Structure-Aware Test-Time Calibration), a label-free calibration head that operates directly on the similarity matrix of frozen EEG and image encoders. SATTC combines a geometric expert, subject-adaptive whitening of EEG embeddings with an adaptive variant of Cross-domain Similarity Local Scaling (CSLS), and a structural expert built from mutual nearest neighbors, bidirectional top-k ranks, and class popularity, fused via a simple Product-of-Experts rule. On THINGS-EEG under a strict leave-one-subject-out protocol, standardized inference with cosine similarities, L2-normalized embeddings, and candidate whitening already yields a strong cross-subject baseline over the original ATM retrieval setup. Building on this baseline, SATTC further improves Top-1 and Top-5 accuracy, reduces hubness and per-class imbalance, and produces more reliable small-k shortlists. These gains transfer across multiple EEG encoders, supporting SATTC as an encoder-agnostic, label-free test-time calibration layer for cross-subject neural decoding.

Method: Proposes Structure-Aware Domain Generalization (SADG) with: 1) Structure-aware serialization (SAS) using centroid-based topology and geodesic curvature continuity for transformation-invariant sequences; 2) Hierarchical domain-aware modeling (HDM) to stabilize cross-domain reasoning; 3) Lightweight spectral graph alignment (SGA) for test-time feature shifting without parameter updates.

Result: The approach improves structural fidelity and consistently outperforms state-of-the-art methods across multiple tasks including reconstruction, denoising, and registration. Also introduces MP3DObject dataset for multi-task DG evaluation.

Conclusion: SADG effectively addresses structural drift in multi-task domain generalization for point clouds by preserving structural hierarchy through transformation-invariant serialization and hierarchical domain modeling.

Abstract: While recent Transformer and Mamba architectures have advanced point cloud representation learning, they are typically developed for single-task or single-domain settings. Directly applying them to multi-task domain generalization (DG) leads to degraded performance. Transformers effectively model global dependencies but suffer from quadratic attention cost and lack explicit structural ordering, whereas Mamba offers linear-time recurrence yet often depends on coordinate-driven serialization, which is sensitive to viewpoint changes and missing regions, causing structural drift and unstable sequential modeling. In this paper, we propose Structure-Aware Domain Generalization (SADG), a Mamba-based In-Context Learning framework that preserves structural hierarchy across domains and tasks. We design structure-aware serialization (SAS) that generates transformation-invariant sequences using centroid-based topology and geodesic curvature continuity. We further devise hierarchical domain-aware modeling (HDM) that stabilizes cross-domain reasoning by consolidating intra-domain structure and fusing inter-domain relations. At test time, we introduce a lightweight spectral graph alignment (SGA) that shifts target features toward source prototypes in the spectral domain without updating model parameters, ensuring structure-preserving test-time feature shifting. In addition, we introduce MP3DObject, a real-scan object dataset for multi-task DG evaluation. Comprehensive experiments demonstrate that the proposed approach improves structural fidelity and consistently outperforms state-of-the-art methods across multiple tasks including reconstruction, denoising, and registration.

Method: CTCal leverages reliable cross-attention maps from smaller timesteps (less noise) to calibrate representation learning at larger timesteps (more noise), with timestep-aware adaptive weighting to balance CTCal and diffusion loss.

Result: Extensive experiments on T2I-Compbench++ and GenEval benchmarks demonstrate effectiveness and generalizability across both diffusion-based (SD 2.1) and flow-based (SD 3) approaches.

Conclusion: CTCal provides explicit supervision for text-image alignment, is model-agnostic, and can be seamlessly integrated into existing text-to-image diffusion models to improve alignment quality.

Abstract: Recent advancements in text-to-image synthesis have been largely propelled by diffusion-based models, yet achieving precise alignment between text prompts and generated images remains a persistent challenge. We find that this difficulty arises primarily from the limitations of conventional diffusion loss, which provides only implicit supervision for modeling fine-grained text-image correspondence. In this paper, we introduce Cross-Timestep Self-Calibration (CTCal), founded on the supporting observation that establishing accurate text-image alignment within diffusion models becomes progressively more difficult as the timestep increases. CTCal leverages the reliable text-image alignment (i.e., cross-attention maps) formed at smaller timesteps with less noise to calibrate the representation learning at larger timesteps with more noise, thereby providing explicit supervision during training. We further propose a timestep-aware adaptive weighting to achieve a harmonious integration of CTCal and diffusion loss. CTCal is model-agnostic and can be seamlessly integrated into existing text-to-image diffusion models, encompassing both diffusion-based (e.g., SD 2.1) and flow-based approaches (e.g., SD 3). Extensive experiments on T2I-Compbench++ and GenEval benchmarks demonstrate the effectiveness and generalizability of the proposed CTCal. Our code is available at https://github.com/xiefan-guo/ctcal.

Method: Real-time video surveillance with YOLOv5 deep learning model tracks gauzes on “In” and “Out” trays; integrated model detects both humans and gauzes using 11,000 training images; supports manual count adjustments

Result: Improved accuracy and increased frame rate from 8 FPS to 15 FPS; system now supports manual adjustments based on doctor feedback

Conclusion: AI-based gauze counting system offers reliable prevention method for gossypiboma, addressing limitations of manual counting and previous iterations

Abstract: During surgeries, there is a risk of medical gauzes being left inside patients’ bodies, leading to “Gossypiboma” in patients and can cause serious complications in patients and also lead to legal problems for hospitals from malpractice lawsuits and regulatory penalties. Diagnosis depends on imaging methods such as X-rays or CT scans, and the usual treatment involves surgical excision. Prevention methods, such as manual counts and RFID-integrated gauzes, aim to minimize gossypiboma risks. However, manual tallying of 100s of gauzes by nurses is time-consuming and diverts resources from patient care. In partnership with Singapore General Hospital (SGH) we have developed a new prevention method, an AI-based system for gauze counting in surgical settings. Utilizing real-time video surveillance and object recognition technology powered by YOLOv5, a Deep Learning model was designed to monitor gauzes on two designated trays labelled “In” and “Out”. Gauzes are tracked from the “In” tray, prior to their use in the patient’s body & in the “Out” tray post-use, ensuring accurate counting and verifying that no gauze remains inside the patient at the end of the surgery. We have trained it using numerous images from Operation Theatres & augmented it to satisfy all possible scenarios. This study has also addressed the shortcomings of previous project iterations. Previously, the project employed two models: one for human detection and another for gauze detection, trained on a total of 2800 images. Now we have an integrated model capable of identifying both humans and gauzes, using a training set of 11,000 images. This has led to improvements in accuracy and increased the frame rate from 8 FPS to 15 FPS now. Incorporating doctor’s feedback, the system now also supports manual count adjustments, enhancing its reliability in actual surgeries.

Method: Proposes DiT-BlockSkip with two key techniques: 1) Timestep-aware dynamic patch sampling that adjusts patch sizes based on diffusion timesteps and resizes to lower resolution, 2) Block skipping mechanism that selectively fine-tunes essential transformer blocks while precomputing residual features for skipped blocks, using cross-attention masking for block selection.

Result: Achieves competitive personalization performance both qualitatively and quantitatively while substantially reducing memory usage, enabling on-device feasibility for large-scale diffusion transformers.

Conclusion: DiT-BlockSkip provides an effective memory-efficient fine-tuning framework that maintains performance while reducing computational requirements, moving toward practical deployment of personalized diffusion transformers on resource-constrained devices.

Abstract: Diffusion Transformers (DiTs) have significantly enhanced text-to-image (T2I) generation quality, enabling high-quality personalized content creation. However, fine-tuning these models requires substantial computational complexity and memory, limiting practical deployment under resource constraints. To tackle these challenges, we propose a memory-efficient fine-tuning framework called DiT-BlockSkip, integrating timestep-aware dynamic patch sampling and block skipping by precomputing residual features. Our dynamic patch sampling strategy adjusts patch sizes based on the diffusion timestep, then resizes the cropped patches to a fixed lower resolution. This approach reduces forward & backward memory usage while allowing the model to capture global structures at higher timesteps and fine-grained details at lower timesteps. The block skipping mechanism selectively fine-tunes essential transformer blocks and precomputes residual features for the skipped blocks, significantly reducing training memory. To identify vital blocks for personalization, we introduce a block selection strategy based on cross-attention masking. Evaluations demonstrate that our approach achieves competitive personalization performance qualitatively and quantitatively, while reducing memory usage substantially, moving toward on-device feasibility (e.g., smartphones, IoT devices) for large-scale diffusion transformers.

Method: 1) Dual-Thread Engine decouples map rendering from core localization for low latency and drift-free accuracy; 2) Large-scale synthetic dataset with precise annotations enables training lightweight network for zero-shot simulation-to-real generalization; 3) Joint Neural-Guided Stochastic-Gradient Optimizer (JNGO) ensures robust convergence under aggressive motion.

Result: PiLoT outperforms state-of-the-art methods on comprehensive benchmarks while running over 25 FPS on NVIDIA Jetson Orin platform.

Conclusion: PiLoT provides a robust, accurate, real-time solution for UAV geo-localization that works in GNSS-denied environments without expensive active sensors, enabled by novel architectural and optimization techniques.

Abstract: We present PiLoT, a unified framework that tackles UAV-based ego and target geo-localization. Conventional approaches rely on decoupled pipelines that fuse GNSS and Visual-Inertial Odometry (VIO) for ego-pose estimation, and active sensors like laser rangefinders for target localization. However, these methods are susceptible to failure in GNSS-denied environments and incur substantial hardware costs and complexity. PiLoT breaks this paradigm by directly registering live video stream against a geo-referenced 3D map. To achieve robust, accurate, and real-time performance, we introduce three key contributions: 1) a Dual-Thread Engine that decouples map rendering from core localization thread, ensuring both low latency while maintaining drift-free accuracy; 2) a large-scale synthetic dataset with precise geometric annotations (camera pose, depth maps). This dataset enables the training of a lightweight network that generalizes in a zero-shot manner from simulation to real data; and 3) a Joint Neural-Guided Stochastic-Gradient Optimizer (JNGO) that achieves robust convergence even under aggressive motion. Evaluations on a comprehensive set of public and newly collected benchmarks show that PiLoT outperforms state-of-the-art methods while running over 25 FPS on NVIDIA Jetson Orin platform. Our code and dataset is available at: https://github.com/Choyaa/PiLoT.

Method: Proposes Masked Edge Prediction Model (MEMO) with: 1) Large-scale synthetic edge dataset for pre-training, 2) Lightweight fine-tuning module (1.2% additional params), 3) Training with varying input masking ratios, 4) Progressive inference strategy that finalizes edge predictions based on confidence gradients (high center, lower boundaries).

Result: Achieves visually appealing, post-processing-free, human-like edge maps. Outperforms prior methods on crispness-aware evaluations.

Conclusion: Carefully designed training and inference strategy alone (without specialized losses or architecture changes) is sufficient to achieve crisp, human-like edge quality using only cross-entropy loss.

Abstract: Learning-based edge detection models trained with cross-entropy loss often suffer from thick edge predictions, which deviate from the crisp, single-pixel annotations typically provided by humans. While previous approaches to achieving crisp edges have focused on designing specialized loss functions or modifying network architectures, we show that a carefully designed training and inference strategy alone is sufficient to achieve human-like edge quality. In this work, we introduce the Masked Edge Prediction MOdel (MEMO), which produces both accurate and crisp edges using only cross-entropy loss. We first construct a large-scale synthetic edge dataset to pre-train MEMO, enhancing its generalization ability. Subsequent fine-tuning on downstream datasets requires only a lightweight module comprising 1.2% additional parameters. During training, MEMO learns to predict edges under varying ratios of input masking. A key insight guiding our inference is that thick edge predictions typically exhibit a confidence gradient: high in the center and lower toward the boundaries. Leveraging this, we propose a novel progressive prediction strategy that sequentially finalizes edge predictions in order of prediction confidence, resulting in thinner and more precise contours. Our method achieves visually appealing, post-processing-free, human-like edge maps and outperforms prior methods on crispness-aware evaluations.

Method: Plug-and-play test-time framework that: (1) builds memory bank and retrieves semantically/perceptually aligned neighbors using reasoning summaries, (2) reframes VLM as probabilistic comparator for pairwise preference probabilities, fusing ordinal evidence with initial scores using Thurstone’s Case V model, (3) performs gated reflection and consolidates memory for future decisions.

Result: Produces denser, distortion-sensitive predictions that mitigate discrete collapse. Experiments across multiple IQA benchmarks show consistent gains over reasoning-induced VLM baselines, existing non-reasoning IQA methods, and test-time scaling alternatives.

Conclusion: ME-IQA effectively addresses discrete collapse in reasoning-induced VLMs for IQA through memory-enhanced re-ranking, improving sensitivity and performance across benchmarks.

Abstract: Reasoning-induced vision-language models (VLMs) advance image quality assessment (IQA) with textual reasoning, yet their scalar scores often lack sensitivity and collapse to a few values, so-called discrete collapse. We introduce ME-IQA, a plug-and-play, test-time memory-enhanced re-ranking framework. It (i) builds a memory bank and retrieves semantically and perceptually aligned neighbors using reasoning summaries, (ii) reframes the VLM as a probabilistic comparator to obtain pairwise preference probabilities and fuse this ordinal evidence with the initial score under Thurstone’s Case V model, and (iii) performs gated reflection and consolidates memory to improve future decisions. This yields denser, distortion-sensitive predictions and mitigates discrete collapse. Experiments across multiple IQA benchmarks show consistent gains over strong reasoning-induced VLM baselines, existing non-reasoning IQA methods, and test-time scaling alternatives.

Method: Co-VLN is a minimalist, model-agnostic framework where independently navigating agents exchange structured perceptual memory when they identify common traversed locations. This expands each agent’s receptive field without requiring additional exploration. The framework is validated on the R2R benchmark using two representative paradigms: learning-based DUET and zero-shot MapGPT.

Result: Vision-sharing enabled models yield substantial performance improvements across both paradigms. The framework establishes a strong foundation for collaborative embodied navigation research and systematically reveals the underlying dynamics of peer observation sharing in VLN.

Conclusion: Co-VLN demonstrates that peer observation sharing can significantly benefit VLN systems by overcoming partial observability constraints, enabling agents to leverage each other’s perceptual experiences for improved navigation performance.

Abstract: Vision-Language Navigation (VLN) systems are fundamentally constrained by partial observability, as an agent can only accumulate knowledge from locations it has personally visited. As multiple robots increasingly coexist in shared environments, a natural question arises: can agents navigating the same space benefit from each other’s observations? In this work, we introduce Co-VLN, a minimalist, model-agnostic framework for systematically investigating whether and how peer observations from concurrently navigating agents can benefit VLN. When independently navigating agents identify common traversed locations, they exchange structured perceptual memory, effectively expanding each agent’s receptive field at no additional exploration cost. We validate our framework on the R2R benchmark under two representative paradigms (the learning-based DUET and the zero-shot MapGPT), and conduct extensive analytical experiments to systematically reveal the underlying dynamics of peer observation sharing in VLN. Results demonstrate that vision-sharing enabled model yields substantial performance improvements across both paradigms, establishing a strong foundation for future research in collaborative embodied navigation.

Method: Proposes Clifford-M, a lightweight backbone replacing feed-forward expansion and frequency-splitting modules with sparse geometric interaction. Uses Clifford-style rolling product to jointly capture alignment and structural variation with linear complexity, enabling efficient cross-scale fusion in a compact dual-resolution architecture.

Result: Without pre-training, achieves mean AUC-ROC of 0.8142 and mean macro-F1 of 0.5481 on ODIR-5K using only 0.85M parameters, outperforming larger CNN baselines. On RFMiD without fine-tuning, attains 0.7425 macro AUC and 0.7610 micro AUC, showing robustness to cross-dataset shift.

Conclusion: Competitive and efficient fundus diagnosis can be achieved without explicit frequency engineering by designing core feature interaction to capture multi-scale structure directly through sparse geometric interaction.

Abstract: Multi-label fundus diagnosis requires features that capture both fine-grained lesions and large-scale retinal structure. Many multi-scale medical vision models address this challenge through explicit frequency decomposition, but our ablation studies show that such heuristics provide limited benefit in this setting: replacing the proposed simple dual-resolution stem with Octave Convolution increased parameters by 35% and computation by a 2.23-fold increase in computation; without improving mean accuracy, while a fixed wavelet-based variant performed substantially worse. Motivated by these findings, we propose Clifford-M, a lightweight backbone that replaces both feed-forward expansion and frequency-splitting modules with sparse geometric interaction. The model is built on a Clifford-style rolling product that jointly captures alignment and structural variation with linear complexity, enabling efficient cross-scale fusion and self-refinement in a compact dual-resolution architecture. Without pre-training, Clifford-M achieves a mean AUC-ROC of 0.8142 and a mean macro-F1 (optimal threshold) of 0.5481 on ODIR-5K using only 0.85M parameters, outperforming substantially larger mid-scale CNN baselines under the same training protocol. When evaluated on RFMiD without fine-tuning, it attains 0.7425 +/- 0.0198 macro AUC and 0.7610 +/- 0.0344 micro AUC, indicating reasonable robustness to cross-dataset shift. These results suggest that competitive and efficient fundus diagnosis can be achieved without explicit frequency engineering, provided that the core feature interaction is designed to capture multi-scale structure directly.

Method: Conducts diagnostic analysis of visual representation degradation in MLLMs, identifies degradation in middle layers, and proposes Predictive Regularization (PRe) to force degraded intermediate features to predict initial visual features, maintaining visual attributes.

Result: Extensive experiments show that mitigating visual degradation through PRe effectively boosts vision-language performance, confirming the importance of robust internal visual representations for comprehensive multimodal understanding.

Conclusion: Visual representation degradation is a pervasive issue in MLLMs driven by text-generation optimization, and maintaining visual fidelity through techniques like PRe is crucial for building robust multimodal models with both strong cross-modal reasoning and core visual competence.

Abstract: While Multimodal Large Language Models (MLLMs) excel at vision-language tasks, the cost of their language-driven training on internal visual foundational competence remains unclear. In this paper, we conduct a detailed diagnostic analysis to unveil a pervasive issue: visual representation degradation in MLLMs. Specifically, we find that compared to the initial visual features, the visual representation in the middle layers of LLM exhibits both a degradation in global function and patch structure. We attribute this phenomenon to a visual sacrifice driven by the singular text-generation objective, where the model compromises its visual fidelity to optimize for answer generation. We argue that a robust MLLM requires both strong cross-modal reasoning and core visual competence, and propose Predictive Regularization (PRe) to force degraded intermediate features to predict initial visual features, thereby maintaining the inherent visual attributes of the MLLM’s internal representations. Extensive experiments confirm that mitigating this visual degradation effectively boosts vision-language performance, underscoring the critical importance of fostering robust internal visual representations within MLLMs for comprehensive multimodal understanding.

Method: Proposes EDC with: 1) Tri-stream encoder with Carrier Tokens for compact global context modeling, 2) Discrepancy-Conditioned Hybrid Fusion (DCHF) to selectively suppress unreliable regions during global aggregation, and 3) Auxiliary cloud removal branch with teacher-guided distillation.

Result: Achieves superior accuracy and efficiency: improves mIoU by 0.56% on M3M-CR and 0.88% on WHU-OPT-SAR datasets, while reducing parameters by 46.7% and accelerating inference by 1.98×.

Conclusion: EDC effectively addresses the efficiency-reliability trade-off in optical-SAR semantic segmentation under cloud occlusion through compact global modeling and selective fusion mechanisms.

Abstract: Cloud occlusion severely degrades the semantic integrity of optical remote sensing imagery. While incorporating Synthetic Aperture Radar (SAR) provides complementary observations, achieving efficient global modeling and reliable cross-modal fusion under cloud interference remains challenging. Existing methods rely on dense global attention to capture long-range dependencies, yet such aggregation indiscriminately propagates cloud-induced noise. Improving robustness typically entails enlarging model capacity, which further increases computational overhead. Given the large-scale and high-resolution nature of remote sensing applications, such computational demands hinder practical deployment, leading to an efficiency-reliability trade-off. To address this dilemma, we propose EDC, an efficiency-oriented and discrepancy-conditioned optical-SAR semantic segmentation framework. A tri-stream encoder with Carrier Tokens enables compact global context modeling with reduced complexity. To prevent noise contamination, we introduce a Discrepancy-Conditioned Hybrid Fusion (DCHF) mechanism that selectively suppresses unreliable regions during global aggregation. In addition, an auxiliary cloud removal branch with teacher-guided distillation enhances semantic consistency under occlusion. Extensive experiments demonstrate that EDC achieves superior accuracy and efficiency, improving mIoU by 0.56% and 0.88% on M3M-CR and WHU-OPT-SAR, respectively, while reducing the number of parameters by 46.7% and accelerating inference by 1.98$\times$. Our implementation is available at https://github.com/mengcx0209/EDC.

Method: Introduces PlanaReLoc, a plane-centric paradigm where a deep matcher associates planar primitives between query images and 3D planar maps in a learned unified embedding space. The 6-DoF pose is then solved and refined under a robust framework.

Result: Comprehensive experiments on ScanNet and 12Scenes datasets across hundreds of scenes show superiority of planar primitives in facilitating reliable cross-modal structural correspondences and achieving effective camera relocalization without requiring textured maps, pose priors, or per-scene training.

Conclusion: Planar primitives provide effective region-based representations for camera relocalization, enabling reliable structural correspondences and robust pose estimation in structured environments with minimal requirements.

Abstract: While structure-based relocalizers have long strived for point correspondences when establishing or regressing query-map associations, in this paper, we pioneer the use of planar primitives and 3D planar maps for lightweight 6-DoF camera relocalization in structured environments. Planar primitives, beyond being fundamental entities in projective geometry, also serve as region-based representations that encapsulate both structural and semantic richness. This motivates us to introduce PlanaReLoc, a streamlined plane-centric paradigm where a deep matcher associates planar primitives across the query image and the map within a learned unified embedding space, after which the 6-DoF pose is solved and refined under a robust framework. Through comprehensive experiments on the ScanNet and 12Scenes datasets across hundreds of scenes, our method demonstrates the superiority of planar primitives in facilitating reliable cross-modal structural correspondences and achieving effective camera relocalization without requiring realistically textured/colored maps, pose priors, or per-scene training. The code and data are available at https://github.com/3dv-casia/PlanaReLoc .

Method: Proposes EruDiff with two components: 1) Diffusion Knowledge Distribution Matching (DK-DM) to align knowledge distributions of implicit prompts with explicit anchors, and 2) Negative-Only Reinforcement Learning (NO-RL) to correct biases in explicit prompt rendering.

Result: Significantly enhances performance of leading diffusion models (FLUX and Qwen-Image) on both scientific knowledge (Science-T2I) and world knowledge (WISE) benchmarks, demonstrating effectiveness and generalizability.

Conclusion: EruDiff successfully addresses the knowledge dislocation problem in diffusion models, enabling better handling of implicit prompts requiring world knowledge through knowledge refactoring and bias correction.

Abstract: Text-to-image diffusion models have achieved remarkable fidelity in synthesizing images from explicit text prompts, yet exhibit a critical deficiency in processing implicit prompts that require deep-level world knowledge, ranging from natural sciences to cultural commonsense, resulting in counter-factual synthesis. This paper traces the root of this limitation to a fundamental dislocation of the underlying knowledge structures, manifesting as a chaotic organization of implicit prompts compared to their explicit counterparts. In this paper, we propose EruDiff, which aims to refactor the knowledge within diffusion models. Specifically, we develop the Diffusion Knowledge Distribution Matching (DK-DM) to register the knowledge distribution of intractable implicit prompts with that of well-defined explicit anchors. Furthermore, to rectify the inherent biases in explicit prompt rendering, we employ the Negative-Only Reinforcement Learning (NO-RL) strategy for fine-grained correction. Rigorous empirical evaluations demonstrate that our method significantly enhances the performance of leading diffusion models, including FLUX and Qwen-Image, across both the scientific knowledge benchmark (i.e., Science-T2I) and the world knowledge benchmark (i.e., WISE), underscoring the effectiveness and generalizability. Our code is available at https://github.com/xiefan-guo/erudiff.

Method: MERIT introduces a unified framework with: 1) sensor-aware noise modeling loss that aligns signal-dependent noise statistics of generated images with target domain, 2) conditional multi-scale large kernel attention module for improved context and sensor-aware feature modeling, and 3) MDRAW dataset for multi-domain RAW image translation evaluation.

Result: Extensive experiments show MERIT outperforms prior models with 5.56 dB improvement in quality and 80% reduction in training iterations. The framework demonstrates superior scalability and performance across multiple camera domains.

Conclusion: MERIT provides an effective unified solution for multi-domain RAW image translation that addresses the scalability limitations of previous approaches while improving translation quality through sensor-aware noise modeling and attention mechanisms.

Abstract: RAW images captured by different camera sensors exhibit substantial domain shifts due to varying spectral responses, noise characteristics, and tone behaviors, complicating their direct use in downstream computer vision tasks. Prior methods address this problem by training domain-specific RAW-to-RAW translators for each source-target pair, but such approaches do not scale to real-world scenarios involving multiple types of commercial cameras. In this work, we introduce MERIT, the first unified framework for multi-domain RAW image translation, which leverages a single model to perform translations across arbitrary camera domains. To address domain-specific noise discrepancies, we propose a sensor-aware noise modeling loss that explicitly aligns the signal-dependent noise statistics of the generated images with those of the target domain. We further enhance the generator with a conditional multi-scale large kernel attention module for improved context and sensor-aware feature modeling. To facilitate standardized evaluation, we introduce MDRAW, the first dataset tailored for multi-domain RAW image translation, comprising both paired and unpaired RAW captures from five diverse camera sensors across a wide range of scenes. Extensive experiments demonstrate that MERIT outperforms prior models in both quality (5.56 dB improvement) and scalability (80% reduction in training iterations).

Method: Combines CLIP-based hierarchical pre-ordering, neural ranking head, probabilistic ensemble (Elo, BTL, GP), epistemic-aleatoric uncertainty decomposition, and information-theoretic pair selection to intelligently select which pairs to compare.

Result: Achieves 11-16% annotation reduction while improving inter-rater reliability in visual ranking tasks (medical imaging, historical dating, aesthetics). Extracts 5-20× more ranking information per comparison than baselines in FG-NET with ground-truth ages.

Conclusion: Dodgersort provides Pareto-optimal accuracy-efficiency trade-offs for ranking tasks, with neural adaptation and ensemble uncertainty being key components for cross-domain performance gains.

Abstract: Pairwise comparison labeling is emerging as it yields higher inter-rater reliability than conventional classification labeling, but exhaustive comparisons require quadratic cost. We propose Dodgersort, which leverages CLIP-based hierarchical pre-ordering, a neural ranking head and probabilistic ensemble (Elo, BTL, GP), epistemic–aleatoric uncertainty decomposition, and information-theoretic pair selection. It reduces human comparisons while improving the reliability of the rankings. In visual ranking tasks in medical imaging, historical dating, and aesthetics, Dodgersort achieves a 11–16% annotation reduction while improving inter-rater reliability. Cross-domain ablations across four datasets show that neural adaptation and ensemble uncertainty are key to this gain. In FG-NET with ground-truth ages, the framework extracts 5–20$\times$ more ranking information per comparison than baselines, yielding Pareto-optimal accuracy–efficiency trade-offs.

Method: Developed GOLDMARK framework with curated TCGA cohort and clinically actionable biomarker labels. Provides structured intermediate representations including tile coordinate maps, per-slide feature embeddings from pathology foundation models, quality-control metadata, patient-level splits, trained models, and evaluation outputs.

Result: Across 33 tumor-biomarker tasks, mean AUROC was 0.689 (TCGA) and 0.630 (MSKCC). For eight highest-performing tasks, mean AUROCs were 0.831 and 0.801 respectively, showing stable cross-site performance for established morphologic-genomic associations.

Conclusion: GOLDMARK establishes a shared experimental substrate for computational pathology, enabling reproducible benchmarking and direct comparison of methods across datasets and models.

Abstract: Computational biomarkers (CBs) are histopathology-derived patterns extracted from hematoxylin-eosin (H&E) whole-slide images (WSIs) using artificial intelligence (AI) to predict therapeutic response or prognosis. Recently, slide-level multiple-instance learning (MIL) with pathology foundation models (PFMs) has become the standard baseline for CB development. While these methods have improved predictive performance, computational pathology lacks standardized intermediate data formats, provenance tracking, checkpointing conventions, and reproducible evaluation metrics required for clinical-grade deployment. We introduce GOLDMARK (https://artificialintelligencepathology.org), a standardized benchmarking framework built on a curated TCGA cohort with clinically actionable OncoKB level 1-3 biomarker labels. GOLDMARK releases structured intermediate representations, including tile coordinate maps, per-slide feature embeddings from canonical PFMs, quality-control metadata, predefined patient-level splits, trained slide-level models, and evaluation outputs. Models are trained on TCGA and evaluated on an independent MSKCC cohort with reciprocal testing. Across 33 tumor-biomarker tasks, mean AUROC was 0.689 (TCGA) and 0.630 (MSKCC). Restricting to the eight highest-performing tasks yielded mean AUROCs of 0.831 and 0.801, respectively. These tasks correspond to established morphologic-genomic associations (e.g., LGG IDH1, COAD MSI/BRAF, THCA BRAF/NRAS, BLCA FGFR3, UCEC PTEN) and showed the most stable cross-site performance. Differences between canonical encoders were modest relative to task-specific variability. GOLDMARK establishes a shared experimental substrate for computational pathology, enabling reproducible benchmarking and direct comparison of methods across datasets and models.

Method: 1) Novel 3D Gaussian hand model for temporal rendering consistency; 2) Diffusion-based hand restorer for seamless scene integration and handling complex interactions/deformations; 3) Framework translates multi-modal sensing glove videos to photorealistic bare hands.

Result: Created HandSense, the first multi-modal HOI dataset with glove-to-hand videos and synchronized tactile/IMU signals. Demonstrated significant enhancement of downstream applications including video-based contact estimation and hand tracking under severe occlusion.

Conclusion: Glove2Hand enables creation of physically-informed hand-object interaction datasets that overcome limitations of conventional hand videos, advancing computer vision, robotics, and AR/VR applications.

Abstract: Understanding hand-object interaction (HOI) is fundamental to computer vision, robotics, and AR/VR. However, conventional hand videos often lack essential physical information such as contact forces and motion signals, and are prone to frequent occlusions. To address the challenges, we present Glove2Hand, a framework that translates multi-modal sensing glove HOI videos into photorealistic bare hands, while faithfully preserving the underlying physical interaction dynamics. We introduce a novel 3D Gaussian hand model that ensures temporal rendering consistency. The rendered hand is seamlessly integrated into the scene using a diffusion-based hand restorer, which effectively handles complex hand-object interactions and non-rigid deformations. Leveraging Glove2Hand, we create HandSense, the first multi-modal HOI dataset featuring glove-to-hand videos with synchronized tactile and IMU signals. We demonstrate that HandSense significantly enhances downstream bare-hand applications, including video-based contact estimation and hand tracking under severe occlusion.

Method: Proposes a lightweight ensemble approach using pretrained SwinV2-Tiny, DinoBloom-Small, and ConvNeXT-V2-Tiny models. Uses dataset expansion to address class imbalance, trains 3 instantiations of each architecture in stratified 3-fold cross-validation, and aggregates predictions through logit averaging from all 9 models.

Result: The ensemble achieves excellent performance on the challenging white blood cell classification dataset, though the model shows weaknesses in confusing similar-looking myelocytes in granulopoiesis and lymphocytes in lymphopoiesis.

Conclusion: A simple ensemble of lightweight pretrained vision models can effectively automate white blood cell classification for leukemia diagnosis, demonstrating strong performance despite challenges with rare cell types and dataset imbalances.

Abstract: Automating white blood cell classification for diagnosis of leukaemia is a promising alternative to time-consuming and resource-intensive examination of cells by expert pathologists. However, designing robust algorithms for classification of rare cell types remains challenging due to variations in staining, scanning and inter-patient heterogeneity. We propose a lightweight ensemble approach for classification of cells during Haematopoiesis, with a focus on the biology of Granulopoiesis, Monocytopoiesis and Lymphopoiesis. Through dataset expansion to alleviate some class imbalance, we demonstrate that a simple ensemble of lightweight pretrained SwinV2-Tiny, DinoBloom-Small and ConvNeXT-V2-Tiny models achieves excellent performance on this challenging dataset. We train 3 instantiations of each architecture in a stratified 3-fold cross-validation framework; for an input image, we forward-pass through all 9 models and aggregate through logit averaging. We further reason on the weaknesses of our model in confusing similar-looking myelocytes in granulopoiesis and lymphocytes in lymphopoiesis. Code: https://gitlab.com/siddharthsrivastava/wbc-bench-2026.

Method: FRoG integrates per-Gaussian embedding with coarse-to-fine temporal embedding strategy for early fusion acceleration. It introduces depth- and error-guided sampling for robust initialization, and modulates opacity variations to mitigate local optima in dim scenes.

Result: Comprehensive experiments show FRoG achieves accelerated rendering speeds while maintaining state-of-the-art visual quality for dynamic scene reconstruction.

Conclusion: FRoG presents an efficient and robust framework that addresses key limitations in dynamic 3D Gaussian Splatting, achieving both speed and quality improvements.

Abstract: 3D Gaussian Splatting has demonstrated remarkable real-time rendering capabilities and superior visual quality in novel view synthesis for static scenes. Building upon these advantages, researchers have progressively extended 3D Gaussians to dynamic scene reconstruction. Deformation field-based methods have emerged as a promising approach among various techniques. These methods maintain 3D Gaussian attributes in a canonical field and employ the deformation field to transform this field across temporal sequences. Nevertheless, these approaches frequently encounter challenges such as suboptimal rendering speeds, significant dependence on initial point clouds, and vulnerability to local optima in dim scenes. To overcome these limitations, we present FRoG, an efficient and robust framework for high-quality dynamic scene reconstruction. FRoG integrates per-Gaussian embedding with a coarse-to-fine temporal embedding strategy, accelerating rendering through the early fusion of temporal embeddings. Moreover, to enhance robustness against sparse initializations, we introduce a novel depth- and error-guided sampling strategy. This strategy populates the canonical field with new 3D Gaussians at low-deviation initial positions, significantly reducing the optimization burden on the deformation field and improving detail reconstruction in both static and dynamic regions. Furthermore, by modulating opacity variations, we mitigate the local optima problem in dim scenes, improving color fidelity. Comprehensive experimental results validate that our method achieves accelerated rendering speeds while maintaining state-of-the-art visual quality.

Method: Proposes a targeted weight re-initialization strategy to restore neural plasticity before fine-tuning. The method selectively re-initializes certain weights in pretrained models to break saturation and enable better gradient flow during transfer learning.

Result: Both convolutional neural networks (CNNs) and vision transformers (ViTs) benefit from the approach, achieving higher test accuracy with faster convergence on several image classification benchmarks. The method adds negligible computational overhead and is compatible with standard transfer learning pipelines.

Conclusion: Targeted weight re-initialization effectively addresses neural plasticity loss in transfer learning, improving adaptation to downstream tasks, especially for atypical datasets, while maintaining computational efficiency.

Abstract: Transfer learning with models pretrained on ImageNet has become a standard practice in computer vision. Transfer learning refers to fine-tuning pretrained weights of a neural network on a downstream task, typically unrelated to ImageNet. However, pretrained weights can become saturated and may yield insignificant gradients, failing to adapt to the downstream task. This hinders the ability of the model to train effectively, and is commonly referred to as loss of neural plasticity. Loss of plasticity may prevent the model from fully adapting to the target domain, especially when the downstream dataset is atypical in nature. While this issue has been widely explored in continual learning, it remains relatively understudied in the context of transfer learning. In this work, we propose the use of a targeted weight re-initialization strategy to restore neural plasticity prior to fine-tuning. Our experiments show that both convolutional neural networks (CNNs) and vision transformers (ViTs) benefit from this approach, yielding higher test accuracy with faster convergence on several image classification benchmarks. Our method introduces negligible computational overhead and is compatible with common transfer learning pipelines.

Method: Combines perceptually optimized autoencoder, conditional latent diffusion restoration in compact latent space, and lightweight Timestep-Adaptive Frequency-Gated (TAFG) conditioning. TAFG decomposes condition features into low/high-frequency components, predicts timestep-adaptive gates from current denoising feature and timestep embedding, and progressively releases high-frequency guidance in later denoising stages.

Result: TAFG-MAN achieves favorable quality-efficiency trade-off against baselines. Compared to base variant without TAFG, it improves detail preservation and perceptual quality while maintaining same inference cost. Ablation confirms effectiveness of proposed conditioning mechanism.

Conclusion: The TAFG-MAN framework effectively addresses the LDCT denoising challenge by balancing noise suppression with anatomical detail preservation through timestep-adaptive frequency-gated conditioning in latent diffusion models.

Abstract: Low-dose computed tomography (LDCT) reduces radiation exposure but also introduces substantial noise and structural degradation, making it difficult to suppress noise without erasing subtle anatomical details. In this paper, we present TAFG-MAN, a latent diffusion framework for efficient and high-quality LDCT image denoising. The framework combines a perceptually optimized autoencoder, conditional latent diffusion restoration in a compact latent space, and a lightweight Timestep-Adaptive Frequency-Gated (TAFG) conditioning design. TAFG decomposes condition features into low- and high-frequency components, predicts timestep-adaptive gates from the current denoising feature and timestep embedding, and progressively releases high-frequency guidance in later denoising stages before cross-attention. In this way, the model relies more on stable structural guidance at early reverse steps and introduces fine details more cautiously as denoising proceeds, improving the balance between noise suppression and detail preservation. Experiments show that TAFG-MAN achieves a favorable quality-efficiency trade-off against representative baselines. Compared with its base variant without TAFG, it further improves detail preservation and perceptual quality while maintaining essentially the same inference cost, and ablation results confirm the effectiveness of the proposed conditioning mechanism.

Method: Proposes SG-FSCFormer with two key components: 1) Prompt-guided Temporal Graph Former that captures user intent via adaptive prompt adaptor, and 2) Fine-grained Mask-linguistic Decoder that collaboratively predicts caption-mask pairs using Multi-entity Contrastive loss for fine-grained alignment.

Result: Comprehensive experiments on two benchmark datasets show SG-FSCFormer achieves remarkable performance, effectively capturing user intent and generating precise multimodal outputs tailored to user specifications.

Conclusion: The proposed Controllable Video Segmentation and Captioning task and SG-FSCFormer framework successfully enable interactive video understanding with precise user intent alignment, advancing video multimodal interpretation capabilities.

Abstract: Recent advancements in multimodal large models have significantly bridged the representation gap between diverse modalities, catalyzing the evolution of video multimodal interpretation, which enhances users’ understanding of video content by generating correlated modalities. However, most existing video multimodal interpretation methods primarily concentrate on global comprehension with limited user interaction. To address this, we propose a novel task, Controllable Video Segmentation and Captioning (SegCaptioning), which empowers users to provide specific prompts, such as a bounding box around an object of interest, to simultaneously generate correlated masks and captions that precisely embody user intent. An innovative framework Scene Graph-guided Fine-grained SegCaptioning Transformer (SG-FSCFormer) is designed that integrates a Prompt-guided Temporal Graph Former to effectively captures and represents user intent through an adaptive prompt adaptor, ensuring that the generated content well aligns with the user’s requirements. Furthermore, our model introduces a Fine-grained Mask-linguistic Decoder to collaboratively predict high-quality caption-mask pairs using a Multi-entity Contrastive loss, as well as provide fine-grained alignment between each mask and its corresponding caption tokens, thereby enhancing users’ comprehension of videos. Comprehensive experiments conducted on two benchmark datasets demonstrate that SG-FSCFormer achieves remarkable performance, effectively capturing user intent and generating precise multimodal outputs tailored to user specifications. Our code is available at https://github.com/XuZhang1211/SG-FSCFormer.

Method: GraPHFormer uses a multimodal architecture with two branches: 1) Vision branch processes three-channel persistence images (unweighted, persistence-weighted, radius-weighted topological densities) via DINOv2-ViT-S, 2) TreeLSTM encoder captures geometric and radial attributes from skeleton graphs. Both project to a shared embedding space trained with symmetric InfoNCE loss, augmented by persistence-space transformations that preserve topological semantics.

Result: Achieves state-of-the-art performance on five out of six benchmarks (BIL-6, ACT-4, JML-4, N7, M1-Cell, M1-REG) spanning self-supervised and supervised settings, significantly outperforming topology-only, graph-only, and morphometrics baselines. Demonstrates practical utility in discriminating glial morphologies across cortical regions and species, and detecting signatures of developmental and degenerative processes.

Conclusion: GraPHFormer successfully unifies topological and graph-based analysis of neuronal morphology through multimodal contrastive learning, providing a powerful framework for comprehensive morphological analysis with applications in neuroscience research.

Abstract: Neuronal morphology encodes critical information about circuit function, development, and disease, yet current methods analyze topology or graph structure in isolation. We introduce GraPHFormer, a multimodal architecture that unifies these complementary views through CLIP-style contrastive learning. Our vision branch processes a novel three-channel persistence image encoding unweighted, persistence-weighted, and radius-weighted topological densities via DINOv2-ViT-S. In parallel, a TreeLSTM encoder captures geometric and radial attributes from skeleton graphs. Both project to a shared embedding space trained with symmetric InfoNCE loss, augmented by persistence-space transformations that preserve topological semantics. Evaluated on six benchmarks (BIL-6, ACT-4, JML-4, N7, M1-Cell, M1-REG) spanning self-supervised and supervised settings, GraPHFormer achieves state-of-the-art performance on five benchmarks, significantly outperforming topology-only, graph-only, and morphometrics baselines. We demonstrate practical utility by discriminating glial morphologies across cortical regions and species, and detecting signatures of developmental and degenerative processes. Code: https://github.com/Uzshah/GraPHFormer

Method: Introduces a four-quadrant safety taxonomy evaluating both consistency (stable predictions across paraphrased prompts) and image reliance (predictions that change when image is removed). Evaluates five medical VLM configurations across two chest X-ray datasets (MIMIC-CXR, PadChest) using this framework.

Result: LoRA fine-tuning dramatically reduces flip rates but shifts majority of samples into Dangerous quadrant (consistent but not image-reliant). LLaVA-Rad Base achieves 1.5% flip rate while 98.5% of samples are Dangerous. Dangerous samples exhibit high accuracy (up to 99.6%) and low entropy, making them invisible to standard confidence-based screening.

Conclusion: Consistency alone is insufficient for reliability assessment. Deployment evaluations must pair consistency checks with text-only baselines to expose the false reliability trap where models ignore images while appearing consistent and accurate.

Abstract: Consistency under paraphrase, the property that semantically equivalent prompts yield identical predictions, is increasingly used as a proxy for reliability when deploying medical vision-language models (VLMs). We show this proxy is fundamentally flawed: a model can achieve perfect consistency by relying on text patterns rather than the input image. We introduce a four-quadrant per-sample safety taxonomy that jointly evaluates consistency (stable predictions across paraphrased prompts) and image reliance (predictions that change when the image is removed). Samples are classified as Ideal (consistent and image-reliant), Fragile (inconsistent but image-reliant), Dangerous (consistent but not image-reliant), or Worst (inconsistent and not image-reliant). Evaluating five medical VLM configurations across two chest X-ray datasets (MIMIC-CXR, PadChest), we find that LoRA fine-tuning dramatically reduces flip rates but shifts a majority of samples into the Dangerous quadrant: LLaVA-Rad Base achieves a 1.5% flip rate on PadChest while 98.5% of its samples are Dangerous. Critically, Dangerous samples exhibit high accuracy (up to 99.6%) and low entropy, making them invisible to standard confidence-based screening. We observe a negative correlation between flip rate and Dangerous fraction (r = -0.89, n=10) and recommend that deployment evaluations always pair consistency checks with a text-only baseline: a single additional forward pass that exposes the false reliability trap.

Method: Proposes SkinCLIP-VL framework with frozen CLIP encoder integrated with lightweight, quantized Qwen2.5-VL via LoRA. Introduces Consistency-aware Focal Alignment (CFA) Loss to align visual regions with clinical semantics under long-tailed distributions, combining focal re-weighting, semantic alignment, and calibration.

Result: On ISIC and Derm7pt benchmarks, SkinCLIP-VL surpasses 13B-parameter baselines by 4.3-6.2% in accuracy with 43% fewer parameters. Blinded expert evaluation and out-of-distribution testing show visually grounded rationales significantly enhance clinical trust compared to traditional saliency maps.

Conclusion: SkinCLIP-VL provides a resource-efficient, interpretable solution for medical vision-language tasks, demonstrating that smaller, well-designed models can outperform larger baselines while improving clinical trust through better interpretability.

Abstract: The deployment of vision-language models (VLMs) in dermatology is hindered by the trilemma of high computational costs, extreme data scarcity, and the black-box nature of deep learning. To address these challenges, we present SkinCLIP-VL, a resource-efficient framework that adapts foundation models for trustworthy skin cancer diagnosis. Adopting a frozen perception, adaptive reasoning paradigm, we integrate a frozen CLIP encoder with a lightweight, quantized Qwen2.5-VL via low-rank adaptation (LoRA). To strictly align visual regions with clinical semantics under long-tailed distributions, we propose the Consistency-aware Focal Alignment (CFA) Loss. This objective synergizes focal re-weighting, semantic alignment, and calibration. On ISIC and Derm7pt benchmarks, SkinCLIP-VL surpasses 13B-parameter baselines by 4.3-6.2% in accuracy with 43% fewer parameters. Crucially, blinded expert evaluation and out-of-distribution testing confirm that our visually grounded rationales significantly enhance clinical trust compared to traditional saliency maps.

Method: 1) Derives closed-form analytical solution for optimal intermediate noise in residual-shifting diffusion paradigm; 2) Designs LR-guided multi-input-aware noise predictor to replace random Gaussian noise, embedding LR structural priors while preserving efficient residual-shifting mechanism; 3) Uses high-quality pre-upsampling network to mitigate initial bias; 4) Optimizes end-to-end with compact 4-step trajectory.

Result: Extensive experiments show LPNSR achieves state-of-the-art perceptual performance on both synthetic and real-world datasets without relying on large-scale text-to-image priors.

Conclusion: LPNSR successfully addresses the efficiency-quality trade-off in diffusion-based SR by optimizing intermediate noise and initialization, achieving efficient high-quality super-resolution without external priors.

Abstract: Diffusion-based image super-resolution (SR), which aims to reconstruct high-resolution (HR) images from corresponding low-resolution (LR) observations, faces a fundamental trade-off between inference efficiency and reconstruction quality. The state-of-the-art residual-shifting diffusion framework achieves efficient 4-step inference, yet suffers from severe performance degradation in compact sampling trajectories. This is mainly attributed to two core limitations: the inherent suboptimality of unconstrained random Gaussian noise in intermediate steps, which leads to error accumulation and insufficient LR prior guidance, and the initialization bias caused by naive bicubic upsampling. In this paper, we propose LPNSR, a prior-enhanced efficient diffusion framework to address these issues. We first mathematically derive the closed-form analytical solution of the optimal intermediate noise for the residual-shifting diffusion paradigm, and accordingly design an LR-guided multi-input-aware noise predictor to replace random Gaussian noise, embedding LR structural priors into the reverse process while fully preserving the framework’s core efficient residual-shifting mechanism. We further mitigate initial bias with a high-quality pre-upsampling network to optimize the diffusion starting point. With a compact 4-step trajectory, LPNSR can be optimized in an end-to-end manner. Extensive experiments demonstrate that LPNSR achieves state-of-the-art perceptual performance on both synthetic and real-world datasets, without relying on any large-scale text-to-image priors. The source code of our method can be found at https://github.com/Faze-Hsw/LPNSR.

Method: Proposes a geometry-guided spatial representation framework with two key modules: 1) geometric prior injection module that injects global structural cues into 2D semantic tokens, and 2) geometry-aware reparameterization module that aligns 2D semantic tokens with 3D geometric tokens through cross-modal attention with gated residual fusion.

Result: Consistently outperforms state-of-the-art UAV VLN baselines across both seen and unseen environments, reducing navigation error by 4.03m and improving success rate by 1.27% over strongest baseline on unseen Full split. Produces trajectories with better path alignment and smoother, more stable motion.

Conclusion: SpatialFly effectively bridges the 2D-3D representation gap for UAV visual-language navigation through geometry-guided spatial representation, enabling better spatial reasoning without explicit 3D reconstruction.

Abstract: UAVs play an important role in applications such as autonomous exploration, disaster response, and infrastructure inspection. However, UAV VLN in complex 3D environments remains challenging. A key difficulty is the structural representation mismatch between 2D visual perception and the 3D trajectory decision space, which limits spatial reasoning. To this end, we propose SpatialFly, a geometry-guided spatial representation framework for UAV VLN. Operating on RGB observations without explicit 3D reconstruction, SpatialFly introduces a geometry-guided 2D representation alignment mechanism. Specifically, the geometric prior injection module injects global structural cues into 2D semantic tokens to provide scene-level geometric guidance. The geometry-aware reparameterization module then aligns 2D semantic tokens with 3D geometric tokens through cross-modal attention, followed by gated residual fusion to preserve semantic discrimination. Experimental results show that SpatialFly consistently outperforms state-of-the-art UAV VLN baselines across both seen and unseen environments, reducing NE by 4.03m and improving SR by 1.27% over the strongest baseline on the unseen Full split. Additional trajectory-level analysis shows that SpatialFly produces trajectories with better path alignment and smoother, more stable motion.

Method: Proposes a scalable adversarial evaluation framework using a large language model to generate clinically plausible adversarial prompt variants via “humanized” rewrites and minimal edits mimicking routine clinical communication. Evaluates state-of-the-art medical VLMs on ultrasound multiple-choice question answering benchmarks.

Result: Systematically assesses vulnerability of SOTA Med-VLMs to these attacks, examines how attacker LLM capacity influences attack success, analyzes relationship between attack success and model confidence, and identifies consistent failure patterns across models.

Conclusion: Results highlight realistic robustness gaps that must be addressed for safe clinical translation of medical vision-language models in ultrasound applications.

Abstract: Ultrasound is widely used in clinical practice due to its portability, cost-effectiveness, safety, and real-time imaging capabilities. However, image acquisition and interpretation remain highly operator dependent, motivating the development of robust AI-assisted analysis methods. Vision-language models (VLMs) have recently demonstrated strong multimodal reasoning capabilities and competitive performance in medical image analysis, including ultrasound. However, emerging evidence highlights significant concerns about their trustworthiness. In particular, adversarial robustness is critical because Med-VLMs operate via natural-language instructions, rendering prompt formulation a realistic and practically exploitable point of vulnerability. Small variations (typos, shorthand, underspecified requests, or ambiguous wording) can meaningfully shift model outputs. We propose a scalable adversarial evaluation framework that leverages a large language model (LLM) to generate clinically plausible adversarial prompt variants via “humanized” rewrites and minimal edits that mimic routine clinical communication. Using ultrasound multiple-choice question answering benchmarks, we systematically assess the vulnerability of SOTA Med-VLMs to these attacks, examine how attacker LLM capacity influences attack success, analyze the relationship between attack success and model confidence, and identify consistent failure patterns across models. Our results highlight realistic robustness gaps that must be addressed for safe clinical translation. Code will be released publicly following the review process.

Method: Two-stage pipeline: 1) VideoMAE-based feature extraction using ViT-Giant or ViT-based backbones, 2) Augmented Self-Mask Attention detector enhanced with Spatial Pyramid Pooling-Fast module to capture multi-scale temporal features.

Result: ViT-Giant achieves 88.09% Top-1 accuracy but with high computational cost (1584.06 GFLOPs/segment), while ViT-based variant achieves 82.55% accuracy with much lower cost (101.85 GFLOPs/segment). SPPF integration consistently improves performance, with ViT-Giant + SPPF achieving 92.67% mAP.

Conclusion: The framework demonstrates a clear trade-off between model capacity and efficiency, offering both high-performance (ViT-Giant + SPPF) and lightweight (ViT-based) options suitable for different driver monitoring scenarios like safety checkpoints or fleet management.

Abstract: The identification of hazardous driving behaviors from in-cabin video streams is essential for enhancing road safety and supporting the detection of traffic violations and unsafe driver actions. However, current temporal action localization techniques often struggle to balance accuracy with computational efficiency. In this work, we develop and evaluate a temporal action localization framework tailored for driver monitoring scenarios, particularly suitable for periodic inspection settings such as transportation safety checkpoints or fleet management assessment systems. Our approach follows a two-stage pipeline that combines VideoMAE-based feature extraction with an Augmented Self-Mask Attention (AMA) detector, enhanced by a Spatial Pyramid Pooling-Fast (SPPF) module to capture multi-scale temporal features. Experimental results reveal a distinct trade-off between model capacity and efficiency. At the feature extraction stage, the ViT-Giant backbone delivers higher representations with 88.09% Top-1 test accuracy, while the ViT-based variant proves to be a practical alternative, achieving 82.55% accuracy with significantly lower computational fine-tuning costs (101.85 GFLOPs/segment compared to 1584.06 GFLOPs/segment for Giant). In the downstream localization task, the integration of SPPF consistently improves performance across all configurations. Notably, the ViT-Giant + SPPF model achieves a peak mAP of 92.67%, while the lightweight ViT-based configuration maintains robust results.

Method: 1) Use pixel-aligned Gaussians that can adjust positions along their rays to maximize rendering quality while maintaining scalability. 2) Model depth distribution around each pixel as a Gaussian distribution to enable fast frame-to-scene alignment for tracking. 3) Optimize the system for runtime and storage efficiency.

Result: The method shows advantages over state-of-the-art methods in view rendering quality, camera tracking accuracy, runtime efficiency, and storage complexity on widely used benchmarks.

Conclusion: SGAD-SLAM successfully addresses the limitations of existing 3DGS approaches by introducing adjustable pixel-aligned Gaussians and Gaussian depth distributions, achieving better performance across multiple metrics in RGBD SLAM applications.

Abstract: 3D Gaussian Splatting (3DGS) has made remarkable progress in RGBD SLAM. Current methods usually use 3D Gaussians or view-tied 3D Gaussians to represent radiance fields in tracking and mapping. However, these Gaussians are either too flexible or too limited in movements, resulting in slow convergence or limited rendering quality. To resolve this issue, we adopt pixel-aligned Gaussians but allow each Gaussian to adjust its position along its ray to maximize the rendering quality, even if Gaussians are simplified to improve system scalability. To speed up the tracking, we model the depth distribution around each pixel as a Gaussian distribution, and then use these distributions to align each frame to the 3D scene quickly. We report our evaluations on widely used benchmarks, justify our designs, and show advantages over the latest methods in view rendering, camera tracking, runtime, and storage complexity. Please see our project page for code and videos at https://machineperceptionlab.github.io/SGAD-SLAM-Project .

Method: The system uses single-view camera video to interpret the environment, combining end-to-end learning efficiency with local mapping detail. It processes monocular images into a five-channel tensor (RGB + road segmentation + depth estimation) augmented with object detection and trajectory prediction in a unified framework.

Result: Achieves real-time processing at 29 FPS on a single GPU, representing a 555% speedup over the fastest mapping-based approach, with strong performance in object tracking, prediction, road segmentation, and depth estimation.

Conclusion: LRHPerception successfully demonstrates that real-time monocular perception for autonomous driving can be achieved without sacrificing performance, offering a computationally efficient alternative to existing approaches.

Abstract: Amidst the rapid advancement of camera-based autonomous driving technology, effectiveness is often prioritized with limited attention to computational efficiency. To address this issue, this paper introduces LRHPerception, a real-time monocular perception package for autonomous driving that uses single-view camera video to interpret the surrounding environment. The proposed system combines the computational efficiency of end-to-end learning with the rich representational detail of local mapping methodologies. With significant improvements in object tracking and prediction, road segmentation, and depth estimation integrated into a unified framework, LRHPerception processes monocular image data into a five-channel tensor consisting of RGB, road segmentation, and pixel-level depth estimation, augmented with object detection and trajectory prediction. Experimental results demonstrate strong performance, achieving real-time processing at 29 FPS on a single GPU, representing a 555% speedup over the fastest mapping-based approach.

Method: A two-expert design: 1) Geometry expert predicts camera poses from uncalibrated multi-view images, 2) Appearance expert uses these poses to synthesize 3D Gaussian representations. The modular approach explicitly separates geometry estimation from Gaussian generation.

Result: The framework achieves performance on par with state-of-the-art posed methods and substantially outperforms prior pose-free feed-forward 3DGS approaches, requiring fewer than 5K training iterations.

Conclusion: The modular two-expert design proves highly effective, challenging the prevailing unified paradigm and suggesting advantages of separation for complex 3D geometric estimation and appearance synthesis tasks.

Abstract: Pose-free feed-forward 3D Gaussian Splatting (3DGS) has opened a new frontier for rapid 3D modeling, enabling high-quality Gaussian representations to be generated from uncalibrated multi-view images in a single forward pass. The dominant approach in this space adopts unified monolithic architectures, often built on geometry-centric 3D foundation models, to jointly estimate camera poses and synthesize 3DGS representations within a single network. While architecturally streamlined, such “all-in-one” designs may be suboptimal for high-fidelity 3DGS generation, as they entangle geometric reasoning and appearance modeling within a shared representation. In this work, we introduce 2Xplat, a pose-free feed-forward 3DGS framework based on a two-expert design that explicitly separates geometry estimation from Gaussian generation. A dedicated geometry expert first predicts camera poses, which are then explicitly passed to a powerful appearance expert that synthesizes 3D Gaussians. Despite its conceptual simplicity, being largely underexplored in prior works, the proposed approach proves highly effective. In fewer than 5K training iterations, the proposed two-experts pipeline substantially outperforms prior pose-free feed-forward 3DGS approaches and achieves performance on par with state-of-the-art posed methods. These results challenge the prevailing unified paradigm and suggest the potential advantages of modular design principles for complex 3D geometric estimation and appearance synthesis tasks.

Method: Proposes NoOVD framework with two key components: 1) K-FPN leverages pretrained knowledge from frozen vision-language models to guide novel-category object discovery and enable self-distillation without additional data, preventing forced alignment of novel objects with background. 2) R-RPN adjusts confidence scores of proposals during inference to improve recall of novel-category objects.

Result: Cross-dataset evaluations on OV-LVIS, OV-COCO, and Objects365 demonstrate consistent superior performance across multiple metrics compared to existing approaches.

Conclusion: The proposed NoOVD framework effectively bridges the training-testing gap in open-vocabulary object detection by leveraging frozen VLMs for knowledge distillation and adjusting inference mechanisms, leading to improved novel-category detection performance.

Abstract: Despite the remarkable progress in open-vocabulary object detection (OVD), a significant gap remains between the training and testing phases. During training, the RPN and RoI heads often misclassify unlabeled novel-category objects as background, causing some proposals to be prematurely filtered out by the RPN while others are further misclassified by the RoI head. During testing, these proposals again receive low scores and are removed in post-processing, leading to a significant drop in recall and ultimately weakening novel-category detection performance.To address these issues, we propose a novel training framework-NoOVD-which innovatively integrates a self-distillation mechanism grounded in the knowledge of frozen vision-language models (VLMs). Specifically, we design K-FPN, which leverages the pretrained knowledge of VLMs to guide the model in discovering novel-category objects and facilitates knowledge distillation-without requiring additional data-thus preventing forced alignment of novel objects with background.Additionally, we introduce R-RPN, which adjusts the confidence scores of proposals during inference to improve the recall of novel-category objects. Cross-dataset evaluations on OV-LVIS, OV-COCO, and Objects365 demonstrate that our approach consistently achieves superior performance across multiple metrics.

Method: Introduces a multi-teacher collaborative mechanism with one general teacher and multiple sonar-specific teachers, using multi-teacher alternating guidance and a cross-teacher reliability assessment mechanism to quantify pseudo-label reliability based on consistency across multiple views and teachers.

Result: Achieves 5.08% improvement in mIoU compared to state-of-the-art approaches on the FLSMD dataset when only 2% of data is labeled.

Conclusion: The proposed collaborative teacher framework effectively addresses sonar image challenges by combining general semantic learning with sonar-specific characteristics and mitigating noisy pseudo-labels through reliability assessment.

Abstract: As one of the most important underwater sensing technologies, forward-looking sonar exhibits unique imaging characteristics. Sonar images are often affected by severe speckle noise, low texture contrast, acoustic shadows, and geometric distortions. These factors make it difficult for traditional teacher-student frameworks to achieve satisfactory performance in sonar semantic segmentation tasks under extremely limited labeled data conditions. To address this issue, we propose a Collaborative Teacher Semantic Segmentation Framework for forward-looking sonar images. This framework introduces a multi-teacher collaborative mechanism composed of one general teacher and multiple sonar-specific teachers. By adopting a multi-teacher alternating guidance strategy, the student model can learn general semantic representations while simultaneously capturing the unique characteristics of sonar images, thereby achieving more comprehensive and robust feature modeling. Considering the challenges of sonar images, which can lead teachers to generate a large number of noisy pseudo-labels, we further design a cross-teacher reliability assessment mechanism. This mechanism dynamically quantifies the reliability of pseudo-labels by evaluating the consistency and stability of predictions across multiple views and multiple teachers, thereby mitigating the negative impact caused by noisy pseudo-labels. Notably, on the FLSMD dataset, when only 2% of the data is labeled, our method achieves a 5.08% improvement in mIoU compared to other state-of-the-art approaches.

Method: Proposes Context-aware Visual Fine-tuning (CoVFT) with Context Vector Extraction (CVE) and Contextual Mixture-of-Experts (CoMoE) modules to incorporate multimodal context into visual adaptation, decomposing conflicting optimization signals.

Result: CoVFT achieves state-of-the-art performance on 12 multimodal benchmarks with superior stability. Fine-tuning a 7B MLLM with CoVFT surpasses average performance of 13B counterparts, revealing untapped potential in visual encoder optimization.

Conclusion: Explicitly incorporating multimodal context into visual fine-tuning resolves visual preference conflicts and enables stable, context-sensitive visual updates in MLLMs, unlocking substantial performance gains.

Abstract: Multimodal large language models (MLLMs) achieve remarkable progress in cross-modal perception and reasoning, yet a fundamental question remains unresolved: should the vision encoder be fine-tuned or frozen? Despite the success of models such as LLaVA and Qwen-VL, inconsistent design choices and heterogeneous training setups hinder a unified understanding of visual fine-tuning (VFT) in MLLMs. Through a configuration-aligned benchmark, we find that existing VFT methods fail to consistently outperform the frozen baseline across multimodal tasks. Our analysis suggests that this instability arises from visual preference conflicts, where the context-agnostic nature of vision encoders induces divergent parameter updates under diverse multimodal context. To address this issue, we propose the Context-aware Visual Fine-tuning (CoVFT) framework, which explicitly incorporates multimodal context into visual adaptation. By integrating a Context Vector Extraction (CVE) and a Contextual Mixture-of-Experts (CoMoE) module, CoVFT decomposes conflicting optimization signals and enables stable, context-sensitive visual updates. Extensive experiments on 12 multimodal benchmarks demonstrate that CoVFT achieves state-of-the-art performance with superior stability. Notably, fine-tuning a 7B MLLM with CoVFT surpasses the average performance of its 13B counterpart, revealing substantial untapped potential in visual encoder optimization within MLLMs.

Method: Three novel components: (1) text-modulated soft cascade decoder predicting WT→TC→ET in coarse-to-fine anatomical hierarchy, (2) sub-region-aware prompt tuning with learnable soft prompts and LoRA-adapted BioBERT encoder for specialized text representations per sub-region, (3) text-semantic channel modulators converting representations into channel-wise refinement signals to emphasize clinically described patterns.

Result: Experiments on TextBraTS dataset show consistent improvements across all sub-regions against state-of-the-art methods by 1.7% on Dice and 6% on HD95 metrics.

Conclusion: TextCSP effectively integrates radiological text descriptions with imaging through hierarchical text guidance, addressing limitations of global text embeddings and improving brain tumor segmentation performance.

Abstract: Brain tumor segmentation remains challenging because the three standard sub-regions, i.e., whole tumor (WT), tumor core (TC), and enhancing tumor (ET), often exhibit ambiguous visual boundaries. Integrating radiological description texts with imaging has shown promise. However, most multimodal approaches typically compress a report into a single global text embedding shared across all sub-regions, overlooking their distinct clinical characteristics. We propose TextCSP (text-modulated soft cascade architecture), a hierarchical text-guided framework that builds on the TextBraTS baseline with three novel components: (1) a text-modulated soft cascade decoder that predicts WT->TC->ET in a coarse-to-fine manner consistent with their anatomical containment hierarchy. (2) sub-region-aware prompt tuning, which uses learnable soft prompts with a LoRA-adapted BioBERT encoder to generate specialized text representations tailored for each sub-region; (3) text-semantic channel modulators that convert the aforementioned representations into channel-wise refinement signals, enabling the decoder to emphasize features aligned with clinically described patterns. Experiments on the TextBraTS dataset demonstrate consistent improvements across all sub-regions against state-of-the-art methods by 1.7% and 6% on the main metrics Dice and HD95.

Method: Introduces Variance Expansion loss to counteract variance collapse in latent spaces, using adversarial interplay between reconstruction and variance expansion to balance fidelity and robustness to stochastic sampling.

Result: Extensive experiments show the approach consistently enhances generation quality across different latent diffusion architectures, confirming robustness in latent space is crucial for stable diffusion sampling.

Conclusion: Robustness to sampling perturbations is a key missing ingredient in latent diffusion models; the proposed method improves generation quality by addressing variance collapse while maintaining reconstruction fidelity.

Abstract: Latent diffusion models have emerged as the dominant framework for high-fidelity and efficient image generation, owing to their ability to learn diffusion processes in compact latent spaces. However, while previous research has focused primarily on reconstruction accuracy and semantic alignment of the latent space, we observe that another critical factor, robustness to sampling perturbations, also plays a crucial role in determining generation quality. Through empirical and theoretical analyses, we show that the commonly used $β$-VAE-based tokenizers in latent diffusion models, tend to produce overly compact latent manifolds that are highly sensitive to stochastic perturbations during diffusion sampling, leading to visual degradation. To address this issue, we propose a simple yet effective solution that constructs a latent space robust to sampling perturbations while maintaining strong reconstruction fidelity. This is achieved by introducing a Variance Expansion loss that counteracts variance collapse and leverages the adversarial interplay between reconstruction and variance expansion to achieve an adaptive balance that preserves reconstruction accuracy while improving robustness to stochastic sampling. Extensive experiments demonstrate that our approach consistently enhances generation quality across different latent diffusion architectures, confirming that robustness in latent space is a key missing ingredient for stable and faithful diffusion sampling.

Method: Proposes Disagreement-Guided Refinement Network (DGRNet) with: 1) Four lightweight view-specific adapters attached to shared encoder-decoder for diverse predictions and uncertainty quantification in single forward pass, 2) Disagreement maps to identify high uncertainty regions, 3) Text-conditioned refinement using clinical reports, 4) Diversity-preserving training with pairwise similarity penalties and gradient isolation to prevent view collapse.

Result: On TextBraTS dataset, DGRNet improves state-of-the-art segmentation accuracy by 2.4% in Dice and 11% in HD95 metrics, while providing meaningful uncertainty estimates.

Conclusion: DGRNet effectively addresses uncertainty quantification and leverages textual clinical information to improve brain tumor segmentation accuracy, making it suitable for clinical deployment.

Abstract: Accurate brain tumor segmentation from MRI scans is critical for diagnosis and treatment planning. Despite the strong performance of recent deep learning approaches, two fundamental limitations remain: (1) the lack of reliable uncertainty quantification in single-model predictions, which is essential for clinical deployment because the level of uncertainty may impact treatment decision-making, and (2) the under-utilization of rich information in radiology reports that can guide segmentation in ambiguous regions. In this paper, we propose the Disagreement-Guided Refinement Network (DGRNet), a novel framework that addresses both limitations through multi-view disagreement-based uncertainty estimation and text-conditioned refinement. DGRNet generates diverse predictions via four lightweight view-specific adapters attached to a shared encoder-decoder, enabling efficient uncertainty quantification within a single forward pass. Afterward, we build disagreement maps to identify regions of high segmentation uncertainty, which are then selectively refined according to clinical reports. Moreover, we introduce a diversity-preserving training strategy that combines pairwise similarity penalties and gradient isolation to prevent view collapse. The experimental results on the TextBraTS dataset show that DGRNet favorably improves state-of-the-art segmentation accuracy by 2.4% and 11% in main metrics Dice and HD95, respectively, while providing meaningful uncertainty estimates.

Method: Proposes a clinically guided multitask framework that jointly predicts nodule masks and TI-RADS categories. Uses TI-RADS-aligned radiomics targets to ground risk prediction in clinically meaningful evidence. Introduces RLAR (representation-level adversarial gradient regularizer) to handle gradient competition between tasks by penalizing excessive angular alignment between task-specific adversarial directions in latent space.

Result: On a public TI-RADS dataset, the clinically guided multitask model with RLAR consistently improves risk stratification while maintaining segmentation quality compared to single-task training and conventional multitask baselines.

Conclusion: The proposed framework effectively addresses clinical workflow needs by jointly handling segmentation and risk classification, with RLAR mitigating gradient competition issues under annotator variability.

Abstract: Thyroid ultrasound is the first-line exam for assessing thyroid nodules and determining whether biopsy is warranted. In routine reporting, radiologists produce two coupled outputs: a nodule contour for measurement and a TI-RADS risk category based on sonographic criteria. Yet both contouring style and risk grading vary across readers, creating inconsistent supervision that can degrade standard learning pipelines. In this paper, we address this workflow with a clinically guided multitask framework that jointly predicts the nodule mask and TI-RADS category within a single model. To ground risk prediction in clinically meaningful evidence, we guide the classification embedding using a compact TI-RADS aligned radiomics target during training, while preserving complementary deep features for discriminative performance. However, under annotator variability, naive multitask optimization often fails not because the tasks are unrelated, but because their gradients compete within the shared representation. To make this competition explicit and controllable, we introduce RLAR, a representation-level adversarial gradient regularizer. Rather than performing parameter-level gradient surgery, RLAR uses each task’s normalized adversarial direction in latent space as a geometric probe of task sensitivity and penalizes excessive angular alignment between task-specific adversarial directions. On a public TI-RADS dataset, our clinically guided multitask model with RLAR consistently improves risk stratification while maintaining segmentation quality compared to single-task training and conventional multitask baselines. Code and pretrained models will be released.

Method: Progressive adaptation framework with three adapters: modality-dependent (enhances intra-modal features via frequency decomposition), modality-entangled (enables cross-modal interactions via cross-attention), and task-level (adapts prediction head to fused information).

Result: Extensive experiments on RGB+Thermal, RGB+Depth, and RGB+Event tracking tasks demonstrate impressive performance against state-of-the-art methods.

Conclusion: PATrack effectively bridges RGB pre-trained networks to multi-modal data through progressive adaptation, integrating intra-modal, inter-modal, and task-level adaptations in a unified framework.

Abstract: Due to the limited availability of paired multi-modal data, multi-modal trackers are typically built by adopting pre-trained RGB models with parameter-efficient fine-tuning modules. However, these fine-tuning methods overlook advanced adaptations for applying RGB pre-trained models and fail to modulate a single specific modality, cross-modal interactions, and the prediction head. To address the issues, we propose to perform Progressive Adaptation for Multi-Modal Tracking (PATrack). This innovative approach incorporates modality-dependent, modality-entangled, and task-level adapters, effectively bridging the gap in adapting RGB pre-trained networks to multi-modal data through a progressive strategy. Specifically, modality-specific information is enhanced through the modality-dependent adapter, decomposing the high- and low-frequency components, which ensures a more robust feature representation within each modality. The inter-modal interactions are introduced in the modality-entangled adapter, which implements a cross-attention operation guided by inter-modal shared information, ensuring the reliability of features conveyed between modalities. Additionally, recognising that the strong inductive bias of the prediction head does not adapt to the fused information, a task-level adapter specific to the prediction head is introduced. In summary, our design integrates intra-modal, inter-modal, and task-level adapters into a unified framework. Extensive experiments on RGB+Thermal, RGB+Depth, and RGB+Event tracking tasks demonstrate that our method shows impressive performance against state-of-the-art methods. Code is available at https://github.com/ouha1998/Learning-Progressive-Adaptation-for-Multi-Modal-Tracking.

Method: Introduces Sine-AWB layers that combine low-rank factors and convolutional priors into a single kernel, modulated elementwise by sinusoidal transformations to produce task-specialized weights. A lightweight Clock Net produces bounded frequencies to stabilize modulation during training.

Result: Achieves state-of-the-art performance-efficiency trade-offs on dense prediction benchmarks, with up to +5.39% improvement over single-task fine-tuning using only 6.53M trainable parameters.

Conclusion: Free Sinewich offers a compact and scalable paradigm for multi-task learning based on frequency-based parameter sharing, enabling efficient adaptation to multiple tasks with minimal parameter overhead.

Abstract: Multi-task learning (MTL) aims to enable a single model to solve multiple tasks efficiently; however, current parameter-efficient fine-tuning (PEFT) methods remain largely limited to single-task adaptation. We introduce \textbf{Free Sinewich}, a parameter-efficient multi-task learning framework that enables near-zero-cost weight modulation via frequency switching (\textbf{Free}). Specifically, a \textbf{Sine-AWB (Sinewich)} layer combines low-rank factors and convolutional priors into a single kernel, which is then modulated elementwise by a sinusoidal transformation to produce task-specialized weights. A lightweight Clock Net is introduced to produce bounded frequencies that stabilize this modulation during training. Theoretically, sine modulation enhances the rank of low-rank adapters, while frequency separation decorrelates the weights of different tasks. On dense prediction benchmarks, Free Sinewich achieves state-of-the-art performance-efficiency trade-offs (e.g., up to +5.39% improvement over single-task fine-tuning with only 6.53M trainable parameters), offering a compact and scalable paradigm based on frequency-based parameter sharing. Project page: \href{https://casperliuliuliu.github.io/projects/Free-Sinewich/}{https://casperliuliuliu.github.io/projects/Free-Sinewich}.

Method: Created a controlled diagnostic benchmark with 100 synthetic scenes and 6,000 relational queries to evaluate relational consistency under hypothetical camera orbit transformations without re-rendering images. Measured viewpoint consistency, 360° cycle agreement, and relational stability over sequential transformations.

Result: Despite high single-view accuracy, state-of-the-art MLLMs exhibit systematic degradation under counterfactual viewpoint changes, with frequent violations of cycle consistency and rapid decay in relational stability. Increasing representational structure (from visual input to textual bounding boxes to structured scene graphs) improves stability.

Conclusion: Single-view spatial accuracy overestimates the robustness of induced spatial representations, and representation structure plays a critical role in counterfactual spatial reasoning for MLLMs.

Abstract: Multimodal large language models (MLLMs) achieve strong performance on single-view spatial reasoning tasks, yet it remains unclear whether they maintain stable spatial state representations under counterfactual viewpoint changes. We introduce a controlled diagnostic benchmark that evaluates relational consistency under hypothetical camera orbit transformations without re-rendering images. Across 100 synthetic scenes and 6,000 relational queries, we measure viewpoint consistency, 360° cycle agreement, and relational stability over sequential transformations. Despite high single-view accuracy, state-of-the-art MLLMs exhibit systematic degradation under counterfactual viewpoint changes, with frequent violations of cycle consistency and rapid decay in relational stability. We further evaluate multiple input representations, visual input, textual bounding boxes, and structured scene graphs, and show that increasing representational structure improves stability. Our results suggest that single-view spatial accuracy overestimates the robustness of induced spatial representations and that representation structure plays a critical role in counterfactual spatial reasoning.

Method: Proposes LiFR-Seg framework with uncertainty-aware warping process guided by event-driven motion field and explicit confidence learning. Includes temporal memory attention module for coherence in dynamic scenes.

Result: Achieves 73.82% mIoU on DSEC dataset, statistically indistinguishable from HFR upper-bound (within 0.09%). Validated on DSEC and new synthetic benchmark SHF-DSEC.

Conclusion: Presents efficient paradigm for robust high-frame-rate perception with low-frame-rate hardware through event-guided feature propagation.

Abstract: Dense semantic segmentation in dynamic environments is fundamentally limited by the low-frame-rate (LFR) nature of standard cameras, which creates critical perceptual gaps between frames. To solve this, we introduce Anytime Interframe Semantic Segmentation: a new task for predicting segmentation at any arbitrary time using only a single past RGB frame and a stream of asynchronous event data. This task presents a core challenge: how to robustly propagate dense semantic features using a motion field derived from sparse and often noisy event data, all while mitigating feature degradation in highly dynamic scenes. We propose LiFR-Seg, a novel framework that directly addresses these challenges by propagating deep semantic features through time. The core of our method is an uncertainty-aware warping process, guided by an event-driven motion field and its learned, explicit confidence. A temporal memory attention module further ensures coherence in dynamic scenarios. We validate our method on the DSEC dataset and a new high-frequency synthetic benchmark (SHF-DSEC) we contribute. Remarkably, our LFR system achieves performance (73.82% mIoU on DSEC) that is statistically indistinguishable from an HFR upper-bound (within 0.09%) that has full access to the target frame. This work presents a new, efficient paradigm for achieving robust, high-frame-rate perception with low-frame-rate hardware. Project Page: https://candy-crusher.github.io/LiFR_Seg_Proj/#; Code: https://github.com/Candy-Crusher/LiFR-Seg.git.

Method: Constructs basic diffusion architectures to achieve end-to-end rotation equivariance, with rigorous theoretical analysis of intrinsic equivariance error to validate the embedding of equivariance structures.

Result: Comprehensive evaluation against various MFIF methods across four datasets (Lytro, MFFW, MFI-WHU, Road-MF) shows competitive performance with improvements of 0.28-6.64% across six evaluation metrics.

Conclusion: ReDiffuse successfully addresses the structural preservation challenge in diffusion-based multi-focus image fusion through rotation-equivariant design, achieving state-of-the-art performance while maintaining geometric consistency.

Abstract: Diffusion models have achieved impressive performance on multi-focus image fusion (MFIF). However, a key challenge in applying diffusion models to the ill-posed MFIF problem is that defocus blur can make common symmetric geometric structures (e.g., textures and edges) appear warped and deformed, often leading to unexpected artifacts in the fused images. Therefore, embedding rotation equivariance into diffusion networks is essential, as it enables the fusion results to faithfully preserve the original orientation and structural consistency of geometric patterns underlying the input images. Motivated by this, we propose ReDiffuse, a rotation-equivariant diffusion model for MFIF. Specifically, we carefully construct the basic diffusion architectures to achieve end-to-end rotation equivariance. We also provide a rigorous theoretical analysis to evaluate its intrinsic equivariance error, demonstrating the validity of embedding equivariance structures. ReDiffuse is comprehensively evaluated against various MFIF methods across four datasets (Lytro, MFFW, MFI-WHU, and Road-MF). Results demonstrate that ReDiffuse achieves competitive performance, with improvements of 0.28-6.64% across six evaluation metrics. The code is available at https://github.com/MorvanLi/ReDiffuse.

Method: Multi-Cluster Memory (MCM) organizes stored samples into multiple clusters using pixel-level statistical descriptors, with three mechanisms: descriptor-based cluster assignment, Adjacent Cluster Consolidation to bound memory, and Uniform Cluster Retrieval for balanced supervision.

Result: MCM achieves consistent improvements across 12 configurations on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and DomainNet, with gains up to 5.00% on ImageNet-C and 12.13% on DomainNet, scaling with distributional complexity.

Conclusion: Memory organization is a key design axis for practical test-time adaptation, and multi-cluster approaches better match the inherent multi-modality of real-world test streams.

Abstract: Test-time adaptation (TTA) adapts pre-trained models to distribution shifts at inference using only unlabeled test data. Under the Practical TTA (PTTA) setting, where test streams are temporally correlated and non-i.i.d., memory has become an indispensable component for stable adaptation, yet existing methods universally store amples in a single unstructured pool. We show that this single-cluster design is fundamentally mismatched to PTTA: a stream clusterability analysis reveals that test streams are inherently multi-modal, with the optimal number of mixture components consistently far exceeding one. To close this structural gap, we propose Multi-Cluster Memory (MCM), a plug-and-play framework that organizes stored samples into multiple clusters using lightweight pixel-level statistical descriptors. MCM introduces three complementary mechanisms: descriptor-based cluster assignment to capture distinct distributional modes, Adjacent Cluster Consolidation (ACC) to bound memory usage by merging the most similar temporally adjacent clusters, and Uniform Cluster Retrieval (UCR) to ensure balanced supervision across all modes during adaptation. Integrated with three contemporary TTA methods on CIFAR-10-C, CIFAR-100-C, ImageNet-C, and DomainNet, MCM achieves consistent improvements across all 12 configurations, with gains up to 5.00% on ImageNet-C and 12.13% on DomainNet. Notably, these gains scale with distributional complexity: larger label spaces with greater multi-modality benefit most from multi-cluster organization. GMM-based memory diagnostics further confirm that MCM maintains near-optimal distributional balance, entropy, and mode coverage, whereas single-cluster memory exhibits persistent imbalance and progressive mode loss. These results establish memory organization as a key design axis for practical test-time adaptation.

Method: Introduces MS-CustomNet framework for zero-shot integration of multiple user-provided objects with explicit hierarchical arrangement and spatial placement control. Also presents MSI dataset derived from COCO for training on complex multi-subject compositions.

Result: Achieves DINO-I score of 0.61 for identity preservation and YOLO-L score of 0.94 for positional control in multi-subject customization tasks, demonstrating superior capability in generating high-fidelity images with precise user-directed compositions.

Conclusion: MS-CustomNet offers enhanced fine-grained control over multi-subject image generation, addressing the challenge of maintaining subject identity while enabling explicit user control over compositional structure and spatial relationships.

Abstract: Diffusion-based text-to-image generation has advanced significantly, yet customizing scenes with multiple distinct subjects while maintaining fine-grained control over their interactions remains challenging. Existing methods often struggle to provide explicit user-defined control over the compositional structure and precise spatial relationships between subjects. To address this, we introduce MS-CustomNet, a novel framework for multi-subject customization. MS-CustomNet allows zero-shot integration of multiple user-provided objects and, crucially, empowers users to explicitly define these hierarchical arrangements and spatial placements within the generated image. Our approach ensures individual subject identity preservation while learning and enacting these user-specified inter-subject compositions. We also present the MSI dataset, derived from COCO, to facilitate training on such complex multi-subject compositions. MS-CustomNet offers enhanced, fine-grained control over multi-subject image generation. Our method achieves a DINO-I score of 0.61 for identity preservation and a YOLO-L score of 0.94 for positional control in multi-subject customization tasks, demonstrating its superior capability in generating high-fidelity images with precise, user-directed multi-subject compositions and spatial control.

Method: Uses reinforcement learning with outcome-based rewards to evolve the generative model, incorporates class-wise visual cues to align synthesized features with visual prototypes, and employs a novel cold-start training strategy.

Result: Achieves state-of-the-art results on three ZSL benchmarks with a 4.7% performance gain.

Conclusion: RLVC effectively addresses limitations of generative ZSL by making synthesized features more task-relevant and visually aligned, demonstrating the value of reinforcement learning for feature generation.

Abstract: Recent advances in zero-shot learning (ZSL) have demonstrated the potential of generative models. Typically, generative ZSL synthesizes visual features conditioned on semantic prototypes to model the data distribution of unseen classes, followed by training a classifier on the synthesized data. However, the synthesized features often remain task-agnostic, leading to degraded performance. Moreover, inferring a faithful distribution from semantic prototypes alone is insufficient for classes that are semantically similar but visually distinct. To address these and advance ZSL, we propose RLVC, an outcome-reward reinforcement learning RL framework with visual cues for generative ZSL. At its core, RL empowers the generative model to self-evolve, implicitly enhancing its generation capability. In particular, RLVC updates the generative model using an outcome-based reward, encouraging the synthesis of task-relevant features. Furthermore, we introduce class-wise visual cues that (i) align synthesized features with visual prototypes and (ii) stabilize the RL training updates. For the training process, we present a novel cold-start strategy. Comprehensive experiments and analyses on three prevalent ZSL benchmarks demonstrate that RLVC achieves state-of-the-art results with a 4.7% gain.

Method: Three-stage pipeline: 1) Robust point cloud reconstruction with warping-based anomaly removal, 2) Warping-guided 2D-to-3D instance lifting for consistent instance-aware representation, 3) Novel rendering via point cloud projection refined with 3D-aware diffusion model.

Result: Achieves high-fidelity 3D reconstruction from sparse unposed images, supports object-level scene editing (instance removal) by modifying point cloud, and generates consistent edited views without retraining.

Conclusion: Establishes a new direction for efficient, editable 3D content generation without scene-specific optimization, leveraging diffusion models to compensate for sparse geometry.

Abstract: We introduce a novel, training-free system for reconstructing, understanding, and rendering 3D indoor scenes from a sparse set of unposed RGB images. Unlike traditional radiance field approaches that require dense views and per-scene optimization, our pipeline achieves high-fidelity results without any training or pose preprocessing. The system integrates three key innovations: (1) A robust point cloud reconstruction module that filters unreliable geometry using a warping-based anomaly removal strategy; (2) A warping-guided 2D-to-3D instance lifting mechanism that propagates 2D segmentation masks into a consistent, instance-aware 3D representation; and (3) A novel rendering approach that projects the point cloud into new views and refines the renderings with a 3D-aware diffusion model. Our method leverages the generative power of diffusion to compensate for missing geometry and enhances realism, especially under sparse input conditions. We further demonstrate that object-level scene editing such as instance removal can be naturally supported in our pipeline by modifying only the point cloud, enabling the synthesis of consistent, edited views without retraining. Our results establish a new direction for efficient, editable 3D content generation without relying on scene-specific optimization. Project page: https://jiatongxia.github.io/TID3R/

Method: GIDE introduces a unified framework with Discrete Noise Inversion to capture latent noise patterns in discrete token space, decomposing editing into grounding, inversion, and refinement stages to support various editing instructions while preserving unedited background.

Result: GIDE significantly outperforms prior training-free methods on GIDE-Bench (805 compositional editing scenarios), improving Semantic Correctness by 51.83% and Perceptual Quality by 50.39%, and shows consistent gains over trained baselines on ImgEdit-Bench.

Conclusion: GIDE successfully bridges the gap between DLLMs and precise image editing through discrete noise inversion, enabling training-free editing with various instructions while maintaining background integrity, validated by comprehensive benchmarks.

Abstract: While Diffusion Large Language Models (DLLMs) have demonstrated remarkable capabilities in multi-modal generation, performing precise, training-free image editing remains an open challenge. Unlike continuous diffusion models, the discrete tokenization inherent in DLLMs hinders the application of standard noise inversion techniques, often leading to structural degradation during editing. In this paper, we introduce GIDE (Grounded Inversion for DLLM Image Editing), a unified framework designed to bridge this gap. GIDE incorporates a novel Discrete Noise Inversion mechanism that accurately captures latent noise patterns within the discrete token space, ensuring high-fidelity reconstruction. We then decompose the editing pipeline into grounding, inversion, and refinement stages. This design enables GIDE supporting various editing instructions (text, point and box) and operations while strictly preserving the unedited background. Furthermore, to overcome the limitations of existing single-step evaluation protocols, we introduce GIDE-Bench, a rigorous benchmark comprising 805 compositional editing scenarios guided by diverse multi-modal inputs. Extensive experiments on GIDE-Bench demonstrate that GIDE significantly outperforms prior training-free methods, improving Semantic Correctness by 51.83% and Perceptual Quality by 50.39%. Additional evaluations on ImgEdit-Bench confirm its broad applicability, demonstrating consistent gains over trained baselines and yielding photorealistic consistency on par with leading models.

Method: Proposes a prompt-free, boundary-aware instance segmentation framework that predicts signed distance functions (SDFs) instead of binary masks, enabling smooth and geometry-consistent modeling of cell contours. Uses learned sigmoid mapping to convert SDFs into probability maps for sharp boundary localization. Training is guided by a unified Modified Hausdorff Distance (MHD) loss integrating region- and boundary-based terms.

Result: Evaluations on both public and private high-throughput microscopy datasets demonstrate improved boundary accuracy and instance-level performance compared to recent SAM-based and foundation-model approaches.

Conclusion: The proposed framework effectively addresses the challenge of separating touching cell instances in dense microscopy scenes without requiring extensive prompting, offering improved boundary accuracy and instance separation compared to existing methods.

Abstract: Accurate delineation of individual cells in microscopy videos is essential for studying cellular dynamics, yet separating touching or overlapping instances remains a persistent challenge. Although foundation-model for segmentation such as SAM have broadened the accessibility of image segmentation, they still struggle to separate nearby cell instances in dense microscopy scenes without extensive prompting. We propose a prompt-free, boundary-aware instance segmentation framework that predicts signed distance functions (SDFs) instead of binary masks, enabling smooth and geometry-consistent modeling of cell contours. A learned sigmoid mapping converts SDFs into probability maps, yielding sharp boundary localization and robust separation of adjacent instances. Training is guided by a unified Modified Hausdorff Distance (MHD) loss that integrates region- and boundary-based terms. Evaluations on both public and private high-throughput microscopy datasets demonstrate improved boundary accuracy and instance-level performance compared to recent SAM-based and foundation-model approaches. Source code is available at: https://github.com/ThomasMendelson/BAISeg.git

Method: JANUS formulates jailbreak as optimizing a structured prompt distribution under black-box, end-to-end rewards from T2I systems and safety filters. It replaces high-capacity generators with a low-dimensional mixing policy over two semantically anchored prompt distributions, enabling efficient exploration while preserving target semantics.

Result: JANUS outperforms state-of-the-art jailbreak methods on modern T2I models, improving ASR-8 from 25.30% to 43.15% on Stable Diffusion 3.5 Large Turbo with consistently higher CLIP and NSFW scores. It succeeds across both open-source and commercial models.

Conclusion: The findings expose structural weaknesses in current T2I safety pipelines and motivate stronger, distribution-aware defenses. The paper warns that it contains potentially offensive model outputs.

Abstract: Text-to-image (T2I) models such as Stable Diffusion and DALLE remain susceptible to generating harmful or Not-Safe-For-Work (NSFW) content under jailbreak attacks despite deployed safety filters. Existing jailbreak attacks either rely on proxy-loss optimization instead of the true end-to-end objective, or depend on large-scale and costly RL-trained generators. Motivated by these limitations, we propose JANUS , a lightweight framework that formulates jailbreak as optimizing a structured prompt distribution under a black-box, end-to-end reward from the T2I system and its safety filters. JANUS replaces a high-capacity generator with a low-dimensional mixing policy over two semantically anchored prompt distributions, enabling efficient exploration while preserving the target semantics. On modern T2I models, we outperform state-of-the-art jailbreak methods, improving ASR-8 from 25.30% to 43.15% on Stable Diffusion 3.5 Large Turbo with consistently higher CLIP and NSFW scores. JANUS succeeds across both open-source and commercial models. These findings expose structural weaknesses in current T2I safety pipelines and motivate stronger, distribution-aware defenses. Warning: This paper contains model outputs that may be offensive.

Method: Extends Seg-CFT by subdividing each anatomical structure into regional segments and deriving independent measurements per region, enabling spatially localized and anatomically coherent counterfactuals.

Result: Experiments on coronary CT angiography show realistic, region-specific modifications with finer spatial control for modeling disease progression.

Conclusion: Positional Seg-CFT provides a practical approach for generating localized counterfactuals with anatomical coherence, advancing capabilities for disease modeling and data augmentation.

Abstract: Counterfactual image generation enables controlled data augmentation, bias mitigation, and disease modeling. However, existing methods guided by external classifiers or regressors are limited to subject-level factors (e.g., age) and fail to produce localized structural changes, often resulting in global artifacts. Pixel-level guidance using segmentation masks has been explored, but requires user-defined counterfactual masks, which are tedious and impractical. Segmentor-guided Counterfactual Fine-Tuning (Seg-CFT) addressed this by using segmentation-derived measurements to supervise structure-specific variables, yet it remains restricted to global interventions. We propose Positional Seg-CFT, which subdivides each structure into regional segments and derives independent measurements per region, enabling spatially localized and anatomically coherent counterfactuals. Experiments on coronary CT angiography show that Pos-Seg-CFT generates realistic, region-specific modifications, providing finer spatial control for modeling disease progression.

Method: Proposes a continual learning inspired framework with two key modules: 1) Grouped Knowledge Preservation module to memorize group-specific convergence parameters and promote convergence toward shared solutions without inefficient per-class parameter preservation, and 2) Grouped Sharpness Aware module to explicitly address loss landscape geometry and seek flatter minima. The framework requires no external training samples or pre-trained models.

Result: Extensive experiments on four benchmarks demonstrate significant performance gains over state-of-the-art methods.

Conclusion: The proposed framework effectively addresses tail performance degradation in long-tail classification by promoting convergence to shared, flat minima through grouped knowledge preservation and sharpness-aware optimization, achieving superior performance without requiring additional data or pre-trained models.

Abstract: Balancing performance trade-off on long-tail (LT) data distributions remains a long-standing challenge. In this paper, we posit that this dilemma stems from a phenomenon called “tail performance degradation” (the model tends to severely overfit on head classes while quickly forgetting tail classes) and pose a solution from a loss landscape perspective. We observe that different classes possess divergent convergence points in the loss landscape. Besides, this divergence is aggravated when the model settles into sharp and non-robust minima, rather than a shared and flat solution that is beneficial for all classes. In light of this, we propose a continual learning inspired framework to prevent “tail performance degradation”. To avoid inefficient per-class parameter preservation, a Grouped Knowledge Preservation module is proposed to memorize group-specific convergence parameters, promoting convergence towards a shared solution. Concurrently, our framework integrates a Grouped Sharpness Aware module to seek flatter minima by explicitly addressing the geometry of the loss landscape. Notably, our framework requires neither external training samples nor pre-trained models, facilitating the broad applicability. Extensive experiments on four benchmarks demonstrate significant performance gains over state-of-the-art methods. The code is available at:https://gkp-gsa.github.io/.

Method: Vision-language-geometry framework combining frequency-sensitive cues, multi-scale context, skeleton-segment hierarchy inference with geometric descriptors and geometry-aware textual prompts queried by vision-language models

Result: RoadReasoner surpasses state-of-the-art baselines with 72.6% OA, 64.2% F1 score, and 60.6% SegAcc on SYSU-HiRoads and CHN6-CUG datasets

Conclusion: Framework enables accurate, semantically consistent road hierarchy mapping with public dataset release to support infrastructure applications

Abstract: In this work, we present SYSU-HiRoads, a large-scale hierarchical road dataset, and RoadReasoner, a vision-language-geometry framework for automatic multi-grade road mapping from remote sensing imagery. SYSU-HiRoads is built from GF-2 imagery covering 3631 km2 in Henan Province, China, and contains 1079 image tiles at 0.8 m spatial resolution. Each tile is annotated with dense road masks, vectorized centerlines, and three-level hierarchy labels, enabling the joint training and evaluation of segmentation, topology reconstruction, and hierarchy classification. Building on this dataset, RoadReasoner is designed to generate robust road surface masks, topology-preserving road networks, and semantically coherent hierarchy assignments. We strengthen road feature representation and network connectivity by explicitly enhancing frequency-sensitive cues and multi-scale context. Moreover, we perform hierarchy inference at the skeleton-segment level with geometric descriptors and geometry-aware textual prompts, queried by vision-language models to obtain linguistically interpretable grade decisions. Experiments on SYSU-HiRoads and the CHN6-CUG dataset show that RoadReasoner surpasses state-of-the-art road extraction baselines and produces accurate and semantically consistent road hierarchy maps with 72.6% OA, 64.2% F1 score, and 60.6% SegAcc. The dataset and code will be publicly released to support automated transport infrastructure mapping, road inventory updating, and broader infrastructure management applications.

Method: Created TPC-268 dataset with 10,000 images and 678,050 point annotations covering 268 countable plant categories across 242 species, with Linnaean taxonomy labels (kingdom to species) and organ categories. Provides taxonomy-consistent, scale-aware data splits and benchmarks state-of-the-art regression- and detection-based class-agnostic counting approaches.

Result: Dataset includes diverse plant and fungi categories across multiple observation scales (canopy-level remote sensing to tissue-level microscopy). Provides biologically grounded testbed for evaluating fine-grained counting methods with hierarchical reasoning capabilities.

Conclusion: TPC-268 fills the gap in taxonomy-aware plant counting benchmarks, enabling hierarchical reasoning and species-aware evaluation to advance fine-grained class-agnostic counting in computer vision.

Abstract: Visually cataloging and quantifying the natural world requires pushing the boundaries of both detailed visual classification and counting at scale. Despite significant progress, particularly in crowd and traffic analysis, the fine-grained, taxonomy-aware plant counting remains underexplored in vision. In contrast to crowds, plants exhibit nonrigid morphologies and physical appearance variations across growth stages and environments. To fill this gap, we present TPC-268, the first plant counting benchmark incorporating plant taxonomy. Our dataset couples instance-level point annotations with Linnaean labels (kingdom -> species) and organ categories, enabling hierarchical reasoning and species-aware evaluation. The dataset features 10,000 images with 678,050 point annotations, includes 268 countable plant categories over 242 plant species in Plantae and Fungi, and spans observation scales from canopy-level remote sensing imagery to tissue-level microscopy. We follow the problem setting of class-agnostic counting (CAC), provide taxonomy-consistent, scale-aware data splits, and benchmark state-of-the-art regression- and detection-based CAC approaches. By capturing the biodiversity, hierarchical structure, and multi-scale nature of botanical and mycological taxa, TPC-268 provides a biologically grounded testbed to advance fine-grained class-agnostic counting. Dataset and code are available at https://github.com/tiny-smart/TPC-268.

Method: Proposes Query Guided Mixture-of-Projector (QMoP) with three branches: pooling-based for global semantics, resampler for high-level semantics, and pruning-based for fine-grained detail. Uses Query Guided Router to dynamically select/weight outputs based on visual input and textual queries, with MoE-style fusion.

Result: QMoP outperforms strong baselines, delivers significant savings in memory, computation, and inference time. Also introduces VTCBench benchmark for evaluating visual token compression.

Conclusion: QMoP provides an effective adaptive framework for visual token compression in multimodal LLMs, addressing computational bottlenecks while maintaining performance.

Abstract: Multimodal large language models suffer from severe computational and memory bottlenecks, as the number of visual tokens far exceeds that of textual tokens. While recent methods employ projector modules to align and compress visual tokens into text-aligned features, they typically depend on fixed heuristics that limit adaptability across diverse scenarios. In this paper, we first propose Query Guided Mixture-of-Projector (QMoP), a novel and flexible framework that adaptively compresses visual tokens via three collaborative branches: (1) a pooling-based branch for coarse-grained global semantics, (2) a resampler branch for extracting high-level semantic representations, and (3) a pruning-based branch for fine-grained token selection to preserve critical visual detail. To adaptively coordinate these branches, we introduce the Query Guided Router (QGR), which dynamically selects and weights the outputs from different branches based on both visual input and textual queries. A Mixture-of-Experts-style fusion mechanism is designed to aggregate the outputs, harnessing the strengths of each strategy while suppressing noise. To systematically evaluate the effects of Visual Token Compression, we also develop VTCBench, a dedicated benchmark for evaluating the information loss induced by visual token compression. Extensive experiments demonstrate that despite relying on fundamental compression modules, QMoP outperforms strong baselines and delivers significant savings in memory, computation, and inference time.

Method: Converts high-bit depth maps to 3-channel images using multiwavelength depth encoding, quantizes to 4 bits per channel, then compresses with a learned codec combining convolutional and Transformer layers.

Result: Achieves 0.307 bpp while preserving 99.38% accuracy on Middlebury 2014, with 41.48 ms encoder and 47.45 ms decoder inference times on ScanNet++ iPhone RGB-D subset.

Conclusion: DepthTCM provides efficient depth compression with high fidelity, demonstrating the effectiveness of physics-aware encoding and Transformer-CNN hybrid architectures.

Abstract: We propose DepthTCM, a physics-aware end-to-end framework for depth map compression. In our framework of DepthTCM, the high-bit depth map is first converted to a conventional 3-channel image representation losslessly using a method inspired by a physical sinusoidal fringe pattern based profiliometry system, then the 3-channel color image is encoded and decoded by a recently developed Transformer-CNN mixed neural network architecture. Specifically, DepthTCM maps depth to a smooth 3-channel using multiwavelength depth (MWD) encoding, then globally quantized the MWD encoded representation to 4 bits per channel to reduce entropy, and finally is compressed using a learned codec that combines convolutional and Transformer layers. Experiment results demonstrate the advantage of our proposed method. On Middlebury 2014, DepthTCM reaches 0.307 bpp while preserving 99.38% accuracy, a level of fidelity commensurate with lossless PNG. We additionally demonstrate practical efficiency and scalability, reporting average end-to-end inference times of 41.48 ms (encoder) and 47.45 ms (decoder) on the ScanNet++ iPhone RGB-D subset. Ablations validate our design choices: relative to 8-bit quantization, 4-bit quantization reduces bitrate by 66% while maintaining comparable reconstruction quality, with only a marginal 0.68 dB PSNR change and a 0.04% accuracy difference. In addition, Transformer–CNN blocks further improve PSNR by up to 0.75 dB over CNN-only architectures.

Method: Proposes a deep learning framework combining Vision Transformers (ViT) with colormap-based feature representation to capture long-range dependencies while emphasizing important structural and intensity variations in MRI scans.

Result: Achieves 98.90% classification accuracy and 99.97% AUC on the BRISC2025 dataset (glioma, meningioma, pituitary tumor, non-tumor), outperforming ResNet50, ResNet101, and EfficientNetB2 baselines.

Conclusion: The combination of Vision Transformers with colormap-based feature enhancement provides accurate and robust brain tumor classification with strong potential for clinical decision support applications.

Abstract: Accurate classification of brain tumors from magnetic resonance imaging (MRI) plays a critical role in early diagnosis and effective treatment planning. In this study, we propose a deep learning framework based on Vision Transformers (ViT) enhanced with colormap-based feature representation to improve multi-class brain tumor classification performance. The proposed approach leverages the ability of transformer architectures to capture long-range dependencies while incorporating color mapping techniques to emphasize important structural and intensity variations within MRI scans. Experiments are conducted on the BRISC2025 dataset, which includes four classes: glioma, meningioma, pituitary tumor, and non-tumor cases. The model is trained and evaluated using standard performance metrics such as accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). The proposed method achieves a classification accuracy of 98.90%, outperforming baseline convolutional neural network models including ResNet50, ResNet101, and EfficientNetB2. In addition, the model demonstrates strong generalization capability with an AUC of 99.97%, indicating high discriminative performance across all classes. These results highlight the effectiveness of combining Vision Transformers with colormap-based feature enhancement for accurate and robust brain tumor classification and suggest strong potential for clinical decision support applications.

Method: Retrospective collection of 1,454 eyes from 744 patients (889 normal eyes, 565 keratoconus cases) with four corneal maps per eye (axial curvature, anterior elevation, posterior elevation, pachymetry) and structured tabular clinical data. Data were anonymized and pre-processed into standardized PNG and CSV formats.

Result: Created a comprehensive multimodal dataset with 1,454 eyes, including four types of corneal topography images and clinical parameters, openly available at Zenodo for AI research purposes.

Conclusion: CornOrb represents one of the first large-scale Orbscan-based multimodal datasets from Africa, specifically designed to enable AI-driven detection and analysis of keratoconus using both image and tabular data.

Abstract: In this paper, we present CornOrb, a publicly accessible multimodal dataset of Orbscan corneal topography images and clinical annotations collected from patients in Algeria. The dataset comprises 1,454 eyes from 744 patients, including 889 normal eyes and 565 keratoconus cases. For each eye, four corneal maps are provided (axial curvature, anterior elevation, posterior elevation, and pachymetry), together with structured tabular data including demographic information and key clinical parameters such as astigmatism, maximum keratometry (Kmax), central and thinnest pachymetry, and anterior/posterior asphericity. All data were retrospectively acquired, fully anonymized, and pre-processed into standardized PNG and CSV formats to ensure direct usability for artificial intelligence research. This dataset represents one of the first large-scale Orbscan-based resources from Africa, specifically built to enable robust AI-driven detection and analysis of keratoconus using multimodal data. The data are openly available at Zenodo.

Method: Proposes FoB (Focus on Background), a background-centric prompt generator that treats segmentation as prompt localization. It generates category-agnostic background prompts from support images and localizes them in query images, modeling contextual information for spatial dependencies and using structural patterns of background prompts as constraints for progressive refinement.

Result: FoB outperforms other baselines by large margins on three diverse medical image datasets, achieving state-of-the-art performance on few-shot medical image segmentation and exhibiting strong cross-domain generalization.

Conclusion: Reformulating SAM-based few-shot segmentation as prompt localization with background-centric constraints effectively addresses over-segmentation in medical images, enabling better clinical applicability through improved performance and generalization.

Abstract: Conventional few-shot medical image segmentation (FSMIS) approaches face performance bottlenecks that hinder broader clinical applicability. Although the Segment Anything Model (SAM) exhibits strong category-agnostic segmentation capabilities, its direct application to medical images often leads to over-segmentation due to ambiguous anatomical boundaries. In this paper, we reformulate SAM-based FSMIS as a prompt localization task and propose FoB (Focus on Background), a background-centric prompt generator that provides accurate background prompts to constrain SAM’s over-segmentation. Specifically, FoB bridges the gap between segmentation and prompt localization by category-agnostic generation of support background prompts and localizing them directly in the query image. To address the challenge of prompt localization for novel categories, FoB models rich contextual information to capture foreground-background spatial dependencies. Moreover, inspired by the inherent structural patterns of background prompts in medical images, FoB models this structure as a constraint to progressively refine background prompt predictions. Experiments on three diverse medical image datasets demonstrate that FoB outperforms other baselines by large margins, achieving state-of-the-art performance on FSMIS, and exhibiting strong cross-domain generalization. Our code is available at https://github.com/primebo1/FoB_SAM.

Method: Proposes Group Relative Policy Optimization (GRPO): samples multiple reasoning trajectories, models within-group structure, uses Actor’s self-consistency as training prior, introduces bounded Judge-based modulation to reweight trajectories, converts absolute scores to relative advantages within groups.

Result: Consistently improves reasoning performance and generalization on five mathematical reasoning benchmarks using only unlabeled data.

Conclusion: Offers scalable path toward self-evolving multimodal models through unsupervised self-evolution training framework without human annotations or external reward models.

Abstract: Recent progress in multimodal large language models has led to strong performance on reasoning tasks, but these improvements largely rely on high-quality annotated data or teacher-model distillation, both of which are costly and difficult to scale.To address this, we propose an unsupervised self-evolution training framework for multimodal reasoning that achieves stable performance improvements without using human-annotated answers or external reward models. For each input, we sample multiple reasoning trajectories and jointly model their within group structure.We use the Actor’s self-consistency signal as a training prior, and introduce a bounded Judge based modulation to continuously reweight trajectories of different quality.We further model the modulated scores as a group level distribution and convert absolute scores into relative advantages within each group, enabling more robust policy updates. Trained with Group Relative Policy Optimization (GRPO) on unlabeled data, our method consistently improves reasoning performance and generalization on five mathematical reasoning benchmarks, offering a scalable path toward self-evolving multimodal models.The code are available at https://dingwu1021.github.io/SelfJudge/.

Method: Collected 87,561 handwritten Burmese digit samples from over 150 contributors of different ages and backgrounds, organized into MNIST-compatible 28x28 grayscale images with 60,000 training and 27,561 test samples. Analyzed class distribution, pixel statistics, and morphological variation. Evaluated with simple baselines: MLP, two-layer CNN, and improved CNN with batch normalization and data augmentation.

Result: Baseline models achieved 99.40% (MLP), 99.75% (two-layer CNN), and 99.83% (improved CNN with batch normalization and augmentation) test accuracy. Identified digit pairs that are easily confused due to the round shapes of Myanmar script.

Conclusion: BHDD provides a valuable resource for Burmese handwritten digit recognition research, with baseline results demonstrating the dataset’s quality and the challenges posed by Myanmar script’s characteristics.

Abstract: We introduce the Burmese Handwritten Digit Dataset (BHDD), a collection of 87,561 grayscale images of handwritten Burmese digits in ten classes. Each image is 28x28 pixels, following the MNIST format. The training set has 60,000 samples split evenly across classes; the test set has 27,561 samples with class frequencies as they arose during collection. Over 150 people of different ages and backgrounds contributed samples. We analyze the dataset’s class distribution, pixel statistics, and morphological variation, and identify digit pairs that are easily confused due to the round shapes of the Myanmar script. Simple baselines (an MLP, a two-layer CNN, and an improved CNN with batch normalization and augmentation) reach 99.40%, 99.75%, and 99.83% test accuracy respectively. BHDD is available under CC BY-SA 4.0 at https://github.com/baseresearch/BHDD

Method: TIGON uses a minimalist dual-branch architecture with separate image- and text-conditioned backbones, plus lightweight cross-modal fusion for joint reasoning over visual exemplars and textual specifications.

Result: Text-image conditioning consistently outperforms single-modality methods, demonstrating strong cross-modal complementarity and improved 3D generation quality.

Conclusion: Combining vision and language guidance is a promising direction for future 3D generation research, addressing limitations of current single-modality approaches.

Abstract: High-quality 3D assets are essential for VR/AR, industrial design, and entertainment, motivating growing interest in generative models that create 3D content from user prompts. Most existing 3D generators, however, rely on a single conditioning modality: image-conditioned models achieve high visual fidelity by exploiting pixel-aligned cues but suffer from viewpoint bias when the input view is limited or ambiguous, while text-conditioned models provide broad semantic guidance yet lack low-level visual detail. This limits how users can express intent and raises a natural question: can these two modalities be combined for more flexible and faithful 3D generation? Our diagnostic study shows that even simple late fusion of text- and image-conditioned predictions outperforms single-modality models, revealing strong cross-modal complementarity. We therefore formalize Text-Image Conditioned 3D Generation, which requires joint reasoning over a visual exemplar and a textual specification. To address this task, we introduce TIGON, a minimalist dual-branch baseline with separate image- and text-conditioned backbones and lightweight cross-modal fusion. Extensive experiments show that text-image conditioning consistently improves over single-modality methods, highlighting complementary vision-language guidance as a promising direction for future 3D generation research. Project page: https://jumpat.github.io/tigon-page

Method: Proposes Mv²ID framework with: 1) Region-masking training strategy to prevent shortcut learning and extract essential identity features by aggregating complementary identity cues across views; 2) Reference decoupled-RoPE mechanism assigning distinct positional encoding to video and conditioning tokens for better modeling of heterogeneous properties; 3) Construction of large-scale dataset with diverse facial-angle variations.

Result: Extensive experiments demonstrate significant improvement in identity consistency while maintaining motion naturalness, outperforming existing approaches trained with cross-paired data. Dedicated evaluation metrics for identity consistency and motion naturalness were also proposed.

Conclusion: Mv²ID effectively balances identity consistency and motion naturalness in multi-view reference-to-video generation, addressing the view-dependent copy-paste problem without requiring costly cross-paired data.

Abstract: Single-view reference-to-video methods often struggle to preserve identity consistency under large facial-angle variations. This limitation naturally motivates the incorporation of multi-view facial references. However, simply introducing additional reference images exacerbates the \textit{copy-paste} problem, particularly the \textbf{\textit{view-dependent copy-paste}} artifact, which reduces facial motion naturalness. Although cross-paired data can alleviate this issue, collecting such data is costly. To balance the consistency and naturalness, we propose $\mathrm{Mv}^2\mathrm{ID}$, a multi-view conditioned framework under in-paired supervision. We introduce a region-masking training strategy to prevent shortcut learning and extract essential identity features by encouraging the model to aggregate complementary identity cues across views. In addition, we design a reference decoupled-RoPE mechanism that assigns distinct positional encoding to video and conditioning tokens for better modeling of their heterogeneous properties. Furthermore, we construct a large-scale dataset with diverse facial-angle variations and propose dedicated evaluation metrics for identity consistency and motion naturalness. Extensive experiments demonstrate that our method significantly improves identity consistency while maintaining motion naturalness, outperforming existing approaches trained with cross-paired data.

Method: F4Splat uses a densification-score-guided allocation strategy that predicts per-region densification scores to estimate required Gaussian density based on spatial complexity and multi-view overlap, allowing explicit control over final Gaussian budget without retraining.

Result: The model achieves superior novel-view synthesis performance compared to prior uncalibrated feed-forward methods while using significantly fewer Gaussians, reducing redundancy in simple regions and minimizing duplicate Gaussians across overlapping views.

Conclusion: F4Splat provides an effective solution for compact yet high-quality 3D representations through spatially adaptive Gaussian allocation, addressing limitations of rigid allocation pipelines in feed-forward 3D Gaussian Splatting.

Abstract: Feed-forward 3D Gaussian Splatting methods enable single-pass reconstruction and real-time rendering. However, they typically adopt rigid pixel-to-Gaussian or voxel-to-Gaussian pipelines that uniformly allocate Gaussians, leading to redundant Gaussians across views. Moreover, they lack an effective mechanism to control the total number of Gaussians while maintaining reconstruction fidelity. To address these limitations, we present F4Splat, which performs Feed-Forward predictive densification for Feed-Forward 3D Gaussian Splatting, introducing a densification-score-guided allocation strategy that adaptively distributes Gaussians according to spatial complexity and multi-view overlap. Our model predicts per-region densification scores to estimate the required Gaussian density and allows explicit control over the final Gaussian budget without retraining. This spatially adaptive allocation reduces redundancy in simple regions and minimizes duplicate Gaussians across overlapping views, producing compact yet high-quality 3D representations. Extensive experiments demonstrate that our model achieves superior novel-view synthesis performance compared to prior uncalibrated feed-forward methods, while using significantly fewer Gaussians.

Method: Proposes FedDP-STECAR framework that selectively fine-tunes and perturbs only a small subset of task-relevant layers under Differential Privacy (DP), reducing information leakage while preserving temporal coherence in video features. Only transmits tuned layers during aggregation.

Result: Achieves up to 70.2% higher accuracy under strict privacy (ε=0.65) in centralized settings and 48% faster training with 73.1% accuracy in federated setups. Communication traffic reduced by over 99% compared to full-model updates.

Conclusion: FedDP-STECAR enables scalable and privacy-preserving video action recognition by addressing both privacy leakage and communication efficiency challenges in federated learning for video understanding.

Abstract: Federated video action recognition enables collaborative model training without sharing raw video data, yet remains vulnerable to two key challenges: \textit{model exposure} and \textit{communication overhead}. Gradients exchanged between clients and the server can leak private motion patterns, while full-model synchronization of high-dimensional video networks causes significant bandwidth and communication costs. To address these issues, we propose \textit{Federated Differential Privacy with Selective Tuning and Efficient Communication for Action Recognition}, namely \textit{FedDP-STECAR}. Our \textit{FedDP-STECAR} framework selectively fine-tunes and perturbs only a small subset of task-relevant layers under Differential Privacy (DP), reducing the surface of information leakage while preserving temporal coherence in video features. By transmitting only the tuned layers during aggregation, communication traffic is reduced by over 99% compared to full-model updates. Experiments on the UCF-101 dataset using the MViT-B-16x4 transformer show that \textit{FedDP-STECAR} achieves up to \textbf{70.2% higher accuracy} under strict privacy ($ε=0.65$) in centralized settings and \textbf{48% faster training} with \textbf{73.1% accuracy} in federated setups, enabling scalable and privacy-preserving video action recognition. Code available at https://github.com/izakariyya/mvit-federated-videodp

Method: Proposes TTA-CaP with three coordinated caches: personalized source cache (source prototypes), positive target cache (reliable subject samples), and negative target cache (low-confidence cases). Uses tri-gate mechanism (temporal stability, confidence, consistency) for cache updates and fusion of embeddings for refined predictions.

Result: Outperforms state-of-the-art TTA methods on three challenging video FER datasets (BioVid, StressID, BAH) under subject-specific and environmental shifts while maintaining low computational and memory overhead.

Conclusion: TTA-CaP enables cost-effective personalization of VLMs for video FER through cache-based adaptation, addressing computational limitations of traditional TTA methods while improving performance under distribution shifts.

Abstract: Facial expression recognition (FER) in videos requires model personalization to capture the considerable variations across subjects. Vision-language models (VLMs) offer strong transfer to downstream tasks through image-text alignment, but their performance can still degrade under inter-subject distribution shifts. Personalizing models using test-time adaptation (TTA) methods can mitigate this challenge. However, most state-of-the-art TTA methods rely on unsupervised parameter optimization, introducing computational overhead that is impractical in many real-world applications. This paper introduces TTA through Cache Personalization (TTA-CaP), a cache-based TTA method that enables cost-effective (gradient-free) personalization of VLMs for video FER. Prior cache-based TTA methods rely solely on dynamic memories that store test samples, which can accumulate errors and drift due to noisy pseudo-labels. TTA-CaP leverages three coordinated caches: a personalized source cache that stores source-domain prototypes, a positive target cache that accumulates reliable subject-specific samples, and a negative target cache that stores low-confidence cases as negative samples to reduce the impact of noisy pseudo-labels. Cache updates and replacement are controlled by a tri-gate mechanism based on temporal stability, confidence, and consistency with the personalized cache. Finally, TTA-CaP refines predictions through fusion of embeddings, yielding refined representations that support temporally stable video-level predictions. Our experiments on three challenging video FER datasets, BioVid, StressID, and BAH, indicate that TTA-CaP can outperform state-of-the-art TTA methods under subject-specific and environmental shifts, while maintaining low computational and memory overhead for real-world deployment.

Method: KHMP applies an adaptive Kalman filter in the DCT domain, treating high-frequency DCT coefficients as noisy signals. The filter’s noise parameters are dynamically adjusted based on estimated SNR, enabling aggressive denoising for jittery predictions and conservative filtering for clean motions. This is complemented by training-time physical constraints (temporal smoothness and joint angle limits) that encode biomechanical principles.

Result: Experiments on Human3.6M and HumanEva-I datasets show KHMP achieves state-of-the-art accuracy and effectively mitigates jitter artifacts to produce smooth and physically plausible motions.

Conclusion: KHMP establishes a new paradigm integrating adaptive signal processing with physics-informed learning for high-fidelity human motion prediction, addressing key challenges of jitter and temporal discontinuities.

Abstract: Stochastic human motion prediction aims to generate diverse, plausible futures from observed sequences. Despite advances in generative modeling, existing methods often produce predictions corrupted by high-frequency jitter and temporal discontinuities. To address these challenges, we introduce KHMP, a novel framework featuring an adaptiveKalman filter applied in the DCT domain to generate high-fidelity human motion predictions. By treating high-frequency DCT coefficients as a frequency-indexed noisy signal, the Kalman filter recursively suppresses noise while preserving motion details. Notably, its noise parameters are dynamically adjusted based on estimated Signal-to-Noise Ratio (SNR), enabling aggressive denoising for jittery predictions and conservative filtering for clean motions. This refinement is complemented by training-time physical constraints (temporal smoothness and joint angle limits) that encode biomechanical principles into the generative model. Together, these innovations establish a new paradigm integrating adaptive signal processing with physics-informed learning. Experiments on the Human3.6M and HumanEva-I datasets demonstrate that KHMP achieves state-of-the-art accuracy, effectively mitigating jitter artifacts to produce smooth and physically plausible motions.

Method: Uses Pretrain-and-Adapt paradigm with structured FLAME parameter space instead of direct 3D Gaussian deformation. Introduces Gated Residual Motion Network (GRMN) to capture emotional prosody from audio while supplementing head pose and upper-face cues.

Result: Achieves state-of-the-art performance in emotional expressiveness, lip synchronization, visual realism, and motion stability through extensive experiments.

Conclusion: EmoTaG provides an effective framework for few-shot emotion-aware 3D talking head synthesis with improved geometric stability and emotional coherence.

Abstract: Audio-driven 3D talking head synthesis has advanced rapidly with Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). By leveraging rich pre-trained priors, few-shot methods enable instant personalization from just a few seconds of video. However, under expressive facial motion, existing few-shot approaches often suffer from geometric instability and audio-emotion mismatch, highlighting the need for more effective emotion-aware motion modeling. In this work, we present EmoTaG, a few-shot emotion-aware 3D talking head synthesis framework built on the Pretrain-and-Adapt paradigm. Our key insight is to reformulate motion prediction in a structured FLAME parameter space rather than directly deforming 3D Gaussians, thereby introducing explicit geometric priors that improve motion stability. Building upon this, we propose a Gated Residual Motion Network (GRMN), which captures emotional prosody from audio while supplementing head pose and upper-face cues absent from audio, enabling expressive and coherent motion generation. Extensive experiments demonstrate that EmoTaG achieves state-of-the-art performance in emotional expressiveness, lip synchronization, visual realism, and motion stability.

Method: Proposes a dual-temporal pathway framework: 1) dense JEPA branch for fine-grained motion and interaction cues, and 2) uniformly sampled VLM “thinker” branch with larger temporal stride for knowledge-rich guidance. Uses hierarchical pyramid representation extraction to aggregate multi-layer VLM representations into guidance features compatible with latent prediction.

Result: Outperforms both VLM-only and JEPA-predictor baselines on hand-manipulation trajectory prediction, yielding more robust long-horizon rollout behavior.

Conclusion: The VLM-guided JEPA framework successfully combines the strengths of dense dynamics modeling and semantic guidance, demonstrating improved performance in video prediction tasks requiring both fine-grained motion understanding and long-horizon reasoning.

Abstract: Recent progress in latent world models (e.g., V-JEPA2) has shown promising capability in forecasting future world states from video observations. Nevertheless, dense prediction from a short observation window limits temporal context and can bias predictors toward local, low-level extrapolation, making it difficult to capture long-horizon semantics and reducing downstream utility. Vision–language models (VLMs), in contrast, provide strong semantic grounding and general knowledge by reasoning over uniformly sampled frames, but they are not ideal as standalone dense predictors due to compute-driven sparse sampling, a language-output bottleneck that compresses fine-grained interaction states into text-oriented representations, and a data-regime mismatch when adapting to small action-conditioned datasets. We propose a VLM-guided JEPA-style latent world modeling framework that combines dense-frame dynamics modeling with long-horizon semantic guidance via a dual-temporal pathway: a dense JEPA branch for fine-grained motion and interaction cues, and a uniformly sampled VLM \emph{thinker} branch with a larger temporal stride for knowledge-rich guidance. To transfer the VLM’s progressive reasoning signals effectively, we introduce a hierarchical pyramid representation extraction module that aggregates multi-layer VLM representations into guidance features compatible with latent prediction. Experiments on hand-manipulation trajectory prediction show that our method outperforms both a strong VLM-only baseline and a JEPA-predictor baseline, and yields more robust long-horizon rollout behavior.

Method: 1) Coarse-to-Fine Denoising (C2F) reduces computation during early-stage coarse feature generation where images are indistinguishable. 2) Time Step Sequence Redistribution (TRD) efficiently adjusts sampling trajectories with less than 10 minutes search time.

Result: Achieves near-lossless performance with 80% to 90% reduction in computation on CIFAR10 and LSUN-Church datasets.

Conclusion: The proposed methods effectively reduce computational overhead in diffusion models while maintaining generation quality, making them suitable for edge device deployment.

Abstract: Recently, diffusion models (DMs) have made significant strides in high-quality image generation. However, the multi-step denoising process often results in considerable computational overhead, impeding deployment on resource-constrained edge devices. Existing methods mitigate this issue by compressing models and adjusting the time step sequence. However, they overlook input redundancy and require lengthy search times. In this paper, we propose Coarse-to-Fine Diffusion Models with Time Step Sequence Redistribution. Recognizing indistinguishable early-stage generated images, we introduce Coarse-to-Fine Denoising (C2F) to reduce computation during coarse feature generation. Furthermore, we design Time Step Sequence Redistribution (TRD) for efficient sampling trajectory adjustment, requiring less than 10 minutes for search. Experimental results demonstrate that the proposed methods achieve near-lossless performance with an 80% to 90% reduction in computation on CIFAR10 and LSUN-Church.

Method: WorldCache introduces a Perception-Constrained Dynamical Caching framework with: 1) motion-adaptive thresholds, 2) saliency-weighted drift estimation, 3) optimal approximation via blending and warping, and 4) phase-aware threshold scheduling across diffusion steps.

Result: On Cosmos-Predict2.5-2B evaluated on PAI-Bench, WorldCache achieves 2.3× inference speedup while preserving 99.4% of baseline quality, substantially outperforming prior training-free caching approaches.

Conclusion: WorldCache enables adaptive, motion-consistent feature reuse without retraining, significantly accelerating video generation inference while maintaining high quality.

Abstract: Diffusion Transformers (DiTs) power high-fidelity video world models but remain computationally expensive due to sequential denoising and costly spatio-temporal attention. Training-free feature caching accelerates inference by reusing intermediate activations across denoising steps; however, existing methods largely rely on a Zero-Order Hold assumption i.e., reusing cached features as static snapshots when global drift is small. This often leads to ghosting artifacts, blur, and motion inconsistencies in dynamic scenes. We propose \textbf{WorldCache}, a Perception-Constrained Dynamical Caching framework that improves both when and how to reuse features. WorldCache introduces motion-adaptive thresholds, saliency-weighted drift estimation, optimal approximation via blending and warping, and phase-aware threshold scheduling across diffusion steps. Our cohesive approach enables adaptive, motion-consistent feature reuse without retraining. On Cosmos-Predict2.5-2B evaluated on PAI-Bench, WorldCache achieves \textbf{2.3$\times$} inference speedup while preserving \textbf{99.4%} of baseline quality, substantially outperforming prior training-free caching approaches. Our code can be accessed on \href{https://umair1221.github.io/World-Cache/}{World-Cache}.

Method: Collected videos of healthy volunteers recovering after strenuous exercise, used natural recovery to create labeled dataset, split videos into clips with earlier clips corresponding to more shortness of breath, designed temporal ordering challenge, used ViViT encoder with Lie Relative Encodings (LieRE) and Motion Guided Masking, combined with embedding-based comparison strategy

Result: Achieved F1 score of 0.81 on the respiratory distress recognition task, suggesting modern video transformers can recognize subtle changes in respiratory mechanics

Conclusion: Video transformers show promise for recognizing respiratory distress from video, potentially enabling AI systems to assist in clinical monitoring and early intervention

Abstract: Recognition of respiratory distress through visual inspection is a life saving clinical skill. Clinicians can detect early signs of respiratory deterioration, creating a valuable window for earlier intervention. In this study, we evaluate whether recent advances in video transformers can enable Artificial Intelligence systems to recognize the signs of respiratory distress from video. We collected videos of healthy volunteers recovering after strenuous exercise and used the natural recovery of each participants respiratory status to create a labeled dataset for respiratory distress. Splitting the video into short clips, with earlier clips corresponding to more shortness of breath, we designed a temporal ordering challenge to assess whether an AI system can detect respiratory distress. We found a ViViT encoder augmented with Lie Relative Encodings (LieRE) and Motion Guided Masking, combined with an embedding based comparison strategy, can achieve an F1 score of 0.81 on this task. Our findings suggest that modern video transformers can recognize subtle changes in respiratory mechanics.

Method: Proposes a plug-and-play method that couples geometry reconstruction with fluid-structure interactions, using simulation-based uncertainty metrics and active learning to prioritize views that improve both visual and physical fidelity.

Result: Achieves up to +8.6% visual PSNR improvement and -62.3% velocity divergence during fluid simulations on NeRF Synthetic, Mip-NeRF 360, and DrivAerNet++ datasets.

Conclusion: FluidGaussian successfully integrates physical interaction awareness into 3D reconstruction, improving both visual quality and physical plausibility of reconstructed objects.

Abstract: Real objects that inhabit the physical world follow physical laws and thus behave plausibly during interaction with other physical objects. However, current methods that perform 3D reconstructions of real-world scenes from multi-view 2D images optimize primarily for visual fidelity, i.e., they train with photometric losses and reason about uncertainty in the image or representation space. This appearance-centric view overlooks body contacts and couplings, conflates function-critical regions (e.g., aerodynamic or hydrodynamic surfaces) with ornamentation, and reconstructs structures suboptimally, even when physical regularizers are added. All these can lead to unphysical and implausible interactions. To address this, we consider the question: How can 3D reconstruction become aware of real-world interactions and underlying object functionality, beyond visual cues? To answer this question, we propose FluidGaussian, a plug-and-play method that tightly couples geometry reconstruction with ubiquitous fluid-structure interactions to assess surface quality at high granularity. We define a simulation-based uncertainty metric induced by fluid simulations and integrate it with active learning to prioritize views that improve both visual and physical fidelity. In an empirical evaluation on NeRF Synthetic (Blender), Mip-NeRF 360, and DrivAerNet++, our FluidGaussian method yields up to +8.6% visual PSNR (Peak Signal-to-Noise Ratio) and -62.3% velocity divergence during fluid simulations. Our code is available at https://github.com/delta-lab-ai/FluidGaussian.

Method: Relax Forcing decomposes temporal context into three functional roles: Sink for global stability, Tail for short-term continuity, and dynamically selected History for structural motion guidance. This structured approach selectively incorporates only the most relevant past information instead of attending to dense generated history.

Result: Experiments on VBench-Long demonstrate that Relax Forcing improves motion dynamics and overall temporal consistency while reducing attention overhead. The approach mitigates error accumulation during extrapolation while preserving motion evolution.

Conclusion: Structured temporal memory is essential for scalable long video generation, complementing existing forcing-based training strategies. The work shows that how temporal memory is utilized, not just its quantity, is critical for long-horizon video synthesis.

Abstract: Autoregressive (AR) video diffusion has recently emerged as a promising paradigm for long video generation, enabling causal synthesis beyond the limits of bidirectional models. To address training-inference mismatch, a series of self-forcing strategies have been proposed to improve rollout stability by conditioning the model on its own predictions during training. While these approaches substantially mitigate exposure bias, extending generation to minute-scale horizons remains challenging due to progressive temporal degradation. In this work, we show that this limitation is not primarily caused by insufficient memory, but by how temporal memory is utilised during inference. Through empirical analysis, we find that increasing memory does not consistently improve long-horizon generation, and that the temporal placement of historical context significantly influences motion dynamics while leaving visual quality largely unchanged. These findings suggest that temporal memory should not be treated as a homogeneous buffer. Motivated by this insight, we introduce Relax Forcing, a structured temporal memory mechanism for AR diffusion. Instead of attending to the dense generated history, Relax Forcing decomposes temporal context into three functional roles: Sink for global stability, Tail for short-term continuity, and dynamically selected History for structural motion guidance, and selectively incorporates only the most relevant past information. This design mitigates error accumulation during extrapolation while preserving motion evolution. Experiments on VBench-Long demonstrate that Relax Forcing improves motion dynamics and overall temporal consistency while reducing attention overhead. Our results suggest that structured temporal memory is essential for scalable long video generation, complementing existing forcing-based training strategies.

Method: Uses phase-space decomposition of oscillator dynamics to yield three representations: position (feature content), momentum (spatial gradients for boundaries/texture), and energy (parameter-free saliency). For segmentation (HamSeg), energy gates skip connections while momentum injects boundary info. For classification (HamCls), representations are pooled and concatenated into phase-space feature vector.

Result: Achieves SOTA Dice scores on ISIC 2018 (89.38%), ISIC 2017 (88.40%), TN3K (87.05%), ACDC (92.40%) for segmentation. SOTA accuracy on BloodMNIST (98.85%) and PathMNIST (96.65%) for classification, with competitive results on other MedMNIST datasets. Only 8.57M parameters.

Conclusion: The damped harmonic oscillator provides effective structured inductive bias for medical image analysis, yielding interpretable representations that emerge from Hamiltonian dynamics and achieve strong performance across diverse imaging modalities.

Abstract: We present HamVision, a framework for medical image analysis that uses the damped harmonic oscillator, a fundamental building block of signal processing, as a structured inductive bias for both segmentation and classification tasks. The oscillator’s phase-space decomposition yields three functionally distinct representations: position~$q$ (feature content), momentum~$p$ (spatial gradients that encode boundary and texture information), and energy $H = \tfrac{1}{2}|z|^2$ (a parameter-free saliency map). These representations emerge from the dynamics, not from supervision, and can be exploited by different task-specific heads without any modification to the oscillator itself. For segmentation, energy gates the skip connections while momentum injects boundary information at every decoder level (HamSeg). For classification, the three representations are globally pooled and concatenated into a phase-space feature vector (HamCls). We evaluate HamVision across ten medical imaging benchmarks spanning five imaging modalities. On segmentation, HamSeg achieves state-of-the-art Dice scores on ISIC,2018 (89.38%), ISIC,2017 (88.40%), TN3K (87.05%), and ACDC (92.40%), outperforming most baselines with only 8.57M parameters. On classification, HamCls achieves state-of-the-art accuracy on BloodMNIST (98.85%) and PathMNIST (96.65%), and competitive results on the remaining MedMNIST datasets against MedMamba and MedViT. Diagnostic analysis confirms that the oscillator’s momentum consistently encodes an interior$,{>},$boundary$,{>},$exterior gradient for segmentation and that the energy map correlates with discriminative regions for classification, properties that emerge entirely from the Hamiltonian dynamics. Code is available at https://github.com/Minds-R-Lab/hamvision.

Method: Proposes a phase unwrapping framework based on a diffusion model architecture designed to process large-scale interferograms and address phase discontinuities caused by deformation. The diffusion model can recover physically consistent unwrapped phase fields even with fault-related phase jumps.

Result: Experimental results on both synthetic and real datasets show the method effectively addresses discontinuities associated with near-surface deformation and scales well to large InSAR images.

Conclusion: The diffusion model-based approach offers a practical alternative to manual unwrapping in challenging scenarios involving complex deformation patterns and large-scale interferograms.

Abstract: Phase unwrapping remains a critical and challenging problem in InSAR processing, particularly in scenarios involving complex deformation patterns. In earthquake-related deformation, shallow sources can generate surface-breaking faults and abrupt displacement discontinuities, which severely disrupt phase continuity and often cause conventional unwrapping algorithms to fail. Another limitation of existing learning-based unwrapping methods is their reliance on fixed and relatively small input sizes, while real InSAR interferograms are typically large-scale and spatially heterogeneous. This mismatch restricts the applicability of many neural network approaches to real-world data. In this work, we present a phase unwrapping framework based on a diffusion model, developed to process large-scale interferograms and to address phase discontinuities caused by deformation. By leveraging a diffusion model architecture, the proposed method can recover physically consistent unwrapped phase fields even in the presence of fault-related phase jumps. Experimental results on both synthetic and real datasets demonstrate that the method effectively addresses discontinuities associated with near-surface deformation and scales well to large InSAR images, offering a practical alternative to manual unwrapping in challenging scenarios.

Method: OVRCOAT introduces two modular components: (1) CLIP-conditioned objectness adjustment (COAT) that updates background/foreground probabilities to preserve high-quality masks for out-of-vocabulary objects, and (2) open-vocabulary mask-to-text refinement (OVR) that strengthens CLIP’s region-level alignment to improve classification of both seen and unseen classes with lower memory cost than prior fine-tuning approaches.

Result: OVRCOAT achieves state-of-the-art performance on ADE20K (+5.5% PQ) and delivers significant gains on Mapillary Vistas (+7.1% PQ) and Cityscapes (+3% PQ). The framework improves both objectness estimation and mask recognition while maintaining simplicity and modularity.

Conclusion: The proposed OVRCOAT framework effectively addresses the coupled issues of mask selection bias and limited regional understanding in vision-language models, enabling consistent improvements in open-vocabulary panoptic segmentation across multiple benchmarks with a simple, modular approach.

Abstract: Open-vocabulary panoptic segmentation remains hindered by two coupled issues: (i) mask selection bias, where objectness heads trained on closed vocabularies suppress masks of categories not observed in training, and (ii) limited regional understanding in vision-language models such as CLIP, which were optimized for global image classification rather than localized segmentation. We introduce OVRCOAT, a simple, modular framework that tackles both. First, a CLIP-conditioned objectness adjustment (COAT) updates background/foreground probabilities, preserving high-quality masks for out-of-vocabulary objects. Second, an open-vocabulary mask-to-text refinement (OVR) strengthens CLIP’s region-level alignment to improve classification of both seen and unseen classes with markedly lower memory cost than prior fine-tuning schemes. The two components combine to jointly improve objectness estimation and mask recognition, yielding consistent panoptic gains. Despite its simplicity, OVRCOAT sets a new state of the art on ADE20K (+5.5% PQ) and delivers clear gains on Mapillary Vistas and Cityscapes (+7.1% and +3% PQ, respectively). The code is available at: https://github.com/nickormushev/OVRCOAT

Method: Proposes a two-stage framework: 1) Train identity-suppression network using intra- and inter-video knowledge priors from real-world videos without identity labels to anonymize identity while preserving expressive cues. 2) Denoising module restores expression-related information. Also introduces falsification-based validation method using recognition priors to evaluate privacy robustness without annotated identity labels.

Result: Experiments on three video datasets demonstrate the method effectively protects privacy while maintaining FER accuracy comparable to identity-supervised baselines.

Conclusion: The framework enables privacy-preserving facial expression recognition in challenging open-set video settings without requiring identity labels at any stage.

Abstract: Facial expression recognition relies on facial data that inherently expose identity and thus raise significant privacy concerns. Current privacy-preserving methods typically fail in realistic open-set video settings where identities are unknown, and identity labels are unavailable. We propose a two-stage framework for video-based privacy-preserving FER in challenging open-set settings that requires no identity labels at any stage. To decouple privacy and utility, we first train an identity-suppression network using intra- and inter-video knowledge priors derived from real-world videos without identity labels. This network anonymizes identity while preserving expressive cues. A subsequent denoising module restores expression-related information and helps recover FER performance. Furthermore, we introduce a falsification-based validation method that uses recognition priors to rigorously evaluate privacy robustness without requiring annotated identity labels. Experiments on three video datasets demonstrate that our method effectively protects privacy while maintaining FER accuracy comparable to identity-supervised baselines.

Method: Proposes Beta-weighted Knowledge Distillation (Beta-KD) that formulates teacher-student learning from a Bayesian perspective, interpreting teacher supervision as a Gibbs prior over student activations. This yields a closed-form, uncertainty-aware weighting mechanism that supports arbitrary distillation objectives and their combinations.

Result: Extensive experiments on multimodal VQA benchmarks show that distilling student Vision-Language Models from large teacher VLMs consistently improves performance. Beta-KD outperforms existing knowledge distillation methods.

Conclusion: Beta-KD provides an effective uncertainty-aware framework for adaptive knowledge distillation that balances data and teacher supervision, demonstrating superior performance on multimodal VQA tasks.

Abstract: Knowledge distillation establishes a learning paradigm that leverages both data supervision and teacher guidance. However, determining the optimal balance between learning from data and learning from the teacher is challenging, as some samples may be noisy while others are subject to teacher uncertainty. This motivates the need for adaptively balancing data and teacher supervision. We propose Beta-weighted Knowledge Distillation (Beta-KD), an uncertainty-aware distillation framework that adaptively modulates how much the student relies on teacher guidance. Specifically, we formulate teacher–student learning from a unified Bayesian perspective and interpret teacher supervision as a Gibbs prior over student activations. This yields a closed-form, uncertainty-aware weighting mechanism and supports arbitrary distillation objectives and their combinations. Extensive experiments on multimodal VQA benchmarks demonstrate that distilling student Vision-Language Models from a large teacher VLM consistently improves performance. The results show that Beta-KD outperforms existing knowledge distillation methods. The code is available at https://github.com/Jingchensun/beta-kd.

Method: Uses computer vision and statistical learning techniques for detection and visual interpretation of structural elements in drawings, implemented in the PhotoBeamSolver program.

Result: Development of a documented program capable of solving idealized beam models from drawings, with analysis of challenges and limitations in computer vision integration for structural analysis.

Conclusion: The system demonstrates potential for applying computer vision in civil engineering for structural analysis, infrastructure inspection, and decision-support systems, though challenges remain for reliable field application.

Abstract: This paper presents the development of a documented program capable of solving idealized beam models, such as those commonly used in textbooks and academic exercises, from drawings made by a person. The system is based on computer vision and statistical learning techniques for the detection and visual interpretation of structural elements. Likewise, the main challenges and limitations associated with the integration of computer vision into structural analysis are analyzed, as well as the requirements necessary for its reliable application in the field of civil engineering. In this context, the implementation of the PhotoBeamSolver program is explored, and the current state of computer vision in civil engineering is discussed, particularly in relation to structural analysis, infrastructure inspection, and engineering decision-support systems.

Method: Introduces motion-aware update mechanism using inter-frame pose variation and image frequency cues to estimate frame importance; trajectory-consistent training with relative pose constraints and acceleration regularization; lightweight online stabilization module.

Result: Significantly improves trajectory accuracy, depth estimation, and point cloud reconstruction quality in long video sequences while maintaining competitive performance on shorter sequences across multiple benchmarks.

Conclusion: PAS3R effectively addresses the stability-adaptation dilemma through pose-adaptive streaming, enabling more robust and accurate online 3D reconstruction from monocular video.

Abstract: Online monocular 3D reconstruction enables dense scene recovery from streaming video but remains fundamentally limited by the stability-adaptation dilemma: the reconstruction model must rapidly incorporate novel viewpoints while preserving previously accumulated scene structure. Existing streaming approaches rely on uniform or attention-based update mechanisms that often fail to account for abrupt viewpoint transitions, leading to trajectory drift and geometric inconsistencies over long sequences. We introduce PAS3R, a pose-adaptive streaming reconstruction framework that dynamically modulates state updates according to camera motion and scene structure. Our key insight is that frames contributing significant geometric novelty should exert stronger influence on the reconstruction state, while frames with minor viewpoint variation should prioritize preserving historical context. PAS3R operationalizes this principle through a motion-aware update mechanism that jointly leverages inter-frame pose variation and image frequency cues to estimate frame importance. To further stabilize long-horizon reconstruction, we introduce trajectory-consistent training objectives that incorporate relative pose constraints and acceleration regularization. A lightweight online stabilization module further suppresses high-frequency trajectory jitter and geometric artifacts without increasing memory consumption. Extensive experiments across multiple benchmarks demonstrate that PAS3R significantly improves trajectory accuracy, depth estimation, and point cloud reconstruction quality in long video sequences while maintaining competitive performance on shorter sequences.

Method: EpiMask incorporates three key components: 1) patch-wise affine approximations to camera modeling geometry, 2) epipolar distance-based attention mask to restrict cross-attention to geometrically plausible regions, and 3) fine-tuning of a foundational pretrained image encoder for robust feature extraction.

Result: Experiments on the SatDepth dataset demonstrate up to 30% improvement in matching accuracy compared to re-trained ground-based models.

Conclusion: EpiMask effectively addresses the limitations of ground-based image matching networks for satellite imagery by incorporating domain-specific geometric constraints and attention mechanisms, significantly improving matching performance for satellite images.

Abstract: The deep-learning based image matching networks can now handle significantly larger variations in viewpoints and illuminations while providing matched pairs of pixels with sub-pixel precision. These networks have been trained with ground-based image datasets and, implicitly, their performance is optimized for the pinhole camera geometry. Consequently, you get suboptimal performance when such networks are used to match satellite images since those images are synthesized as a moving satellite camera records one line at a time of the points on the ground. In this paper, we present EpiMask, a semi-dense image matching network for satellite images that (1) Incorporates patch-wise affine approximations to the camera modeling geometry; (2) Uses an epipolar distance-based attention mask to restrict cross-attention to geometrically plausible regions; and (3) That fine-tunes a foundational pretrained image encoder for robust feature extraction. Experiments on the SatDepth dataset demonstrate up to 30% improvement in matching accuracy compared to re-trained ground-based models.

Method: Proposes ALADIN (Attribute-Language Distillation Network) that distills knowledge from frozen CLIP teacher to lightweight ReID student. Key components: 1) Fine-grained attribute-local alignment for adaptive text-visual correspondence, 2) Scene-Aware Prompt Generator for image-specific soft prompts, 3) Attribute-local distillation for consistency between textual attributes and local visual features, 4) Cross-modal contrastive and relation distillation to preserve attribute relationships, 5) Uses Multimodal LLMs to generate structured attribute descriptions converted to localized attention maps via CLIP.

Result: Experiments on Market-1501, DukeMTMC-reID, and MSMT17 show improvements over CNN-, Transformer-, and CLIP-based methods. Demonstrates better generalization and interpretability, with enhanced robustness under occlusions.

Conclusion: ALADIN effectively bridges the gap between global CLIP features and fine-grained ReID requirements through attribute-language distillation, achieving state-of-the-art performance with improved robustness and interpretability.

Abstract: Recent vision-language models such as CLIP provide strong cross-modal alignment, but current CLIP-guided ReID pipelines rely on global features and fixed prompts. This limits their ability to capture fine-grained attribute cues and adapt to diverse appearances. We propose ALADIN, an attribute-language distillation network that distills knowledge from a frozen CLIP teacher to a lightweight ReID student. ALADIN introduces fine-grained attribute-local alignment to establish adaptive text-visual correspondence and robust representation learning. A Scene-Aware Prompt Generator produces image-specific soft prompts to facilitate adaptive alignment. Attribute-local distillation enforces consistency between textual attributes and local visual features, significantly enhancing robustness under occlusions. Furthermore, we employ cross-modal contrastive and relation distillation to preserve the inherent structural relationships among attributes. To provide precise supervision, we leverage Multimodal LLMs to generate structured attribute descriptions, which are then converted into localized attention maps via CLIP. At inference, only the student is used. Experiments on Market-1501, DukeMTMC-reID, and MSMT17 show improvements over CNN-, Transformer-, and CLIP-based methods, with better generalization and interpretability.

Method: Proposes a concept modulator to identify visual-linguistic concept combinations for forget categories, and a mixture of refusal experts (refusers) specialized for concept-aligned refusal generation. Uses multimodal concept-driven routing to reuse refusers for similar concepts and adapt underutilized ones for novel concepts.

Result: Extensive experiments on vision-language benchmarks show the framework outperforms existing methods by generating concept-grounded refusal responses and preserving general utility across unlearning sequences.

Conclusion: The proposed continual unlearning framework effectively addresses the challenge of selective refusal in vision-language models by grounding refusal behavior in fine-grained concepts, enabling precise identification of refusal targets while maintaining model utility.

Abstract: Continual unlearning poses the challenge of enabling large vision-language models to selectively refuse specific image-instruction pairs in response to sequential deletion requests, while preserving general utility. However, sequential unlearning updates distort shared representations, creating spurious associations between vision-language pairs and refusal behaviors that hinder precise identification of refusal targets, resulting in inappropriate refusals. To address this challenge, we propose a novel continual unlearning framework that grounds refusal behavior in fine-grained descriptions of visual and textual concepts decomposed from deletion targets. We first identify which visual-linguistic concept combinations characterize each forget category through a concept modulator, then determine how to generate appropriate refusal responses via a mixture of refusal experts, termed refusers, each specialized for concept-aligned refusal generation. To generate concept-specific refusal responses across sequential tasks, we introduce a multimodal, concept-driven routing scheme that reuses refusers for tasks sharing similar concepts and adapts underutilized ones for novel concepts. Extensive experiments on vision-language benchmarks demonstrate that the proposed framework outperforms existing methods by generating concept-grounded refusal responses and preserving the general utility across unlearning sequences.

Method: Proposes bidirectional text-trajectory alignment where MLLMs handle both grounding-intended (text-to-trajectory) and captioning-intended (trajectory-to-text) instructions. Uses frame-level content integration to adapt trajectory tokens to frame-specific information and a unified mask decoder for end-to-end training.

Result: Outperforms all existing video reasoning segmentation methods on all metrics across referring and reasoning video segmentation datasets.

Conclusion: TrajSeg demonstrates that bidirectional alignment enhances trajectory perception in MLLMs for video reasoning segmentation, achieving state-of-the-art performance with a simplified, end-to-end trainable framework.

Abstract: The prosperity of Multimodal Large Language Models (MLLMs) has stimulated the demand for video reasoning segmentation, which aims to segment video objects based on human instructions. Previous studies rely on unidirectional and implicit text-trajectory alignment, which struggles with trajectory perception when faced with severe video dynamics. In this work, we propose TrajSeg, a simple and unified framework built upon MLLMs. Concretely, we introduce bidirectional text-trajectory alignment, where MLLMs accept grounding-intended (text-to-trajectory) and captioning-intended (trajectory-to-text) instructions. This way, MLLMs can benefit from enhanced correspondence and better perceive object trajectories in videos. The mask generation from trajectories is achieved via a frame-level content integration (FCI) module and a unified mask decoder. The former adapts the MLLM-parsed trajectory-level token to frame-specific information. The latter unifies segmentation for all frames into a single structure, enabling the proposed framework to be simplified and end-to-end trainable. Extensive experiments on referring and reasoning video segmentation datasets demonstrate the effectiveness of TrajSeg, which outperforms all video reasoning segmentation methods on all metrics. The code will be publicly available at https://github.com/haodi19/TrajSeg.

Method: 1) Parameter-efficient prompt tuning via scaling/shifting features in text encoder. 2) WSI representation learning with soft hierarchical textual guidance (no hard instance filtering), leveraging both VLM knowledge and WSI hierarchical structure.

Result: Consistent improvements up to 10.9%, 7.8%, and 13.8% on breast, lung, and ovarian cancer datasets over SOTA. Reduces trainable parameters by 18.1% (breast/lung) and 5.8% (ovarian). Excels at weakly-supervised tumor localization.

Conclusion: Proposed method effectively addresses computational efficiency and information loss issues in FSWC, achieving superior performance with fewer parameters while maintaining localization capabilities.

Abstract: Whole Slide Images (WSIs) are giga-pixel in scale and are typically partitioned into small instances in WSI classification pipelines for computational feasibility. However, obtaining extensive instance level annotations is costly, making few-shot weakly supervised WSI classification (FSWC) crucial for learning from limited slide-level labels. Recently, pre-trained vision-language models (VLMs) have been adopted in FSWC, yet they exhibit several limitations. Existing prompt tuning methods in FSWC substantially increase both the number of trainable parameters and inference overhead. Moreover, current methods discard instances with low alignment to text embeddings from VLMs, potentially leading to information loss. To address these challenges, we propose two key contributions. First, we introduce a new parameter efficient prompt tuning method by scaling and shifting features in text encoder, which significantly reduces the computational cost. Second, to leverage not only the pre-trained knowledge of VLMs, but also the inherent hierarchical structure of WSIs, we introduce a WSI representation learning approach with a soft hierarchical textual guidance strategy without utilizing hard instance filtering. Comprehensive evaluations on pathology datasets covering breast, lung, and ovarian cancer types demonstrate consistent improvements up-to 10.9%, 7.8%, and 13.8% respectively, over the state-of-the-art methods in FSWC. Our method reduces the number of trainable parameters by 18.1% on both breast and lung cancer datasets, and 5.8% on the ovarian cancer dataset, while also excelling at weakly-supervised tumor localization. Code at https://github.com/Jayanie/HIPSS.

Method: Proposes BTP framework that aligns multi-granularity patch features from 3D point clouds with textual representations for localized anomaly detection. Incorporates geometric descriptors to enhance sensitivity to structural anomalies. Uses joint representation learning with auxiliary point cloud data to improve robustness and enrich anomaly semantics.

Result: Extensive experiments on Real3D-AD and Anomaly-ShapeNet datasets demonstrate superior performance in zero-shot 3D anomaly detection compared to existing methods.

Conclusion: BTP effectively leverages pre-trained Point-Language Models for zero-shot 3D anomaly detection by aligning 3D point cloud and textual embeddings while preserving geometric details, outperforming 2D rendering approaches.

Abstract: Zero-shot (ZS) 3D anomaly detection is crucial for reliable industrial inspection, as it enables detecting and localizing defects without requiring any target-category training data. Existing approaches render 3D point clouds into 2D images and leverage pre-trained Vision-Language Models (VLMs) for anomaly detection. However, such strategies inevitably discard geometric details and exhibit limited sensitivity to local anomalies. In this paper, we revisit intrinsic 3D representations and explore the potential of pre-trained Point-Language Models (PLMs) for ZS 3D anomaly detection. We propose BTP (Back To Point), a novel framework that effectively aligns 3D point cloud and textual embeddings. Specifically, BTP aligns multi-granularity patch features with textual representations for localized anomaly detection, while incorporating geometric descriptors to enhance sensitivity to structural anomalies. Furthermore, we introduce a joint representation learning strategy that leverages auxiliary point cloud data to improve robustness and enrich anomaly semantics. Extensive experiments on Real3D-AD and Anomaly-ShapeNet demonstrate that BTP achieves superior performance in ZS 3D anomaly detection. Code will be available at \href{https://github.com/wistful-8029/BTP-3DAD}{https://github.com/wistful-8029/BTP-3DAD}.

Method: VIGIL uses a part-centric structured forensic framework with: 1) Plan-then-examine pipeline: first plans which facial parts to inspect based on global visual cues, then examines each part with independently sourced forensic evidence; 2) Stage-gated injection mechanism: delivers part-level forensic evidence only during examination to prevent bias; 3) Progressive three-stage training with reinforcement learning using part-aware rewards for anatomical validity and evidence-conclusion coherence; 4) OmniFake benchmark: hierarchical 5-Level benchmark for rigorous generalizability evaluation.

Result: Extensive experiments on OmniFake and cross-dataset evaluations show VIGIL consistently outperforms both expert detectors and concurrent MLLM-based methods across all generalizability levels. The model trained on only three foundational generators demonstrates strong performance even on in-the-wild social-media data.

Conclusion: VIGIL provides a more reliable and interpretable approach to deepfake detection by structuring the forensic process into distinct planning and examination stages, preventing evidence contamination and hallucination while maintaining strong generalization capabilities across diverse manipulation scenarios.

Abstract: Multimodal large language models (MLLMs) offer a promising path toward interpretable deepfake detection by generating textual explanations. However, the reasoning process of current MLLM-based methods combines evidence generation and manipulation localization into a unified step. This combination blurs the boundary between faithful observations and hallucinated explanations, leading to unreliable conclusions. Building on this, we present VIGIL, a part-centric structured forensic framework inspired by expert forensic practice through a plan-then-examine pipeline: the model first plans which facial parts warrant inspection based on global visual cues, then examines each part with independently sourced forensic evidence. A stage-gated injection mechanism delivers part-level forensic evidence only during examination, ensuring that part selection remains driven by the model’s own perception rather than biased by external signals. We further propose a progressive three-stage training paradigm whose reinforcement learning stage employs part-aware rewards to enforce anatomical validity and evidence–conclusion coherence. To enable rigorous generalizability evaluation, we construct OmniFake, a hierarchical 5-Level benchmark where the model, trained on only three foundational generators, is progressively tested up to in-the-wild social-media data. Extensive experiments on OmniFake and cross-dataset evaluations demonstrate that VIGIL consistently outperforms both expert detectors and concurrent MLLM-based methods across all generalizability levels.

Method: PEARL follows an align-then-propagate principle: 1) Procrustes alignment performs orthogonal projection in the last self-attention block, rotating keys toward query subspace via stable polar iteration; 2) Text-aware Laplacian propagation refines per-pixel logits on a small grid through confidence-weighted, text-guided graph solve where text provides data-trust signals and neighbor gating, while image gradients preserve boundaries.

Result: PEARL sets new state-of-the-art in training-free OVSS without extra data or auxiliary backbones across standard benchmarks, achieving superior performance under both with-background and without-background protocols.

Conclusion: PEARL provides a compact, efficient training-free solution for open-vocabulary semantic segmentation that effectively leverages cross-modal geometry while maintaining simplicity and low latency.

Abstract: Training-free open-vocabulary semantic segmentation (OVSS) promises rapid adaptation to new label sets without retraining. Yet, many methods rely on heavy post-processing or handle text and vision in isolation, leaving cross-modal geometry underutilized. Others introduce auxiliary vision backbones or multi-model pipelines, which increase complexity and latency while compromising design simplicity. We present PEARL, \textbf{\underline{P}}rocrust\textbf{\underline{e}}s \textbf{\underline{a}}lignment with text-awa\textbf{\underline{r}}e \textbf{\underline{L}}aplacian propagation, a compact two-step inference that follows an align-then-propagate principle. The Procrustes alignment step performs an orthogonal projection inside the last self-attention block, rotating keys toward the query subspace via a stable polar iteration. The text-aware Laplacian propagation then refines per-pixel logits on a small grid through a confidence-weighted, text-guided graph solve: text provides both a data-trust signal and neighbor gating, while image gradients preserve boundaries. In this work, our method is fully training-free, plug-and-play, and uses only fixed constants, adding minimal latency with a small per-head projection and a few conjugate-gradient steps. Our approach, PEARL, sets a new state-of-the-art in training-free OVSS without extra data or auxiliary backbones across standard benchmarks, achieving superior performance under both with-background and without-background protocols.

Method: PROBE optimizes lightweight pseudo-token embeddings through denoising reconstruction with latent alignment constraints that anchor recovery to spatiotemporal structure. It uses multi-level evaluation including classifier detection, semantic similarity, temporal reactivation analysis, and human validation across three T2V architectures, three concept categories, and three erasure strategies.

Result: All tested erasure methods leave measurable residual capacity whose robustness correlates with intervention depth. The paper identifies temporal re-emergence as a video-specific failure mode where suppressed concepts progressively resurface across frames, invisible to frame-level metrics.

Conclusion: Current concept erasure methods achieve only output-level suppression rather than true representational removal. The PROBE protocol provides a reproducible safety auditing framework for T2V models.

Abstract: Concept erasure techniques for text-to-video (T2V) diffusion models report substantial suppression of sensitive content, yet current evaluation is limited to checking whether the target concept is absent from generated frames, treating output-level suppression as evidence of representational removal. We introduce PROBE, a diagnostic protocol that quantifies the \textit{reactivation potential} of erased concepts in T2V models. With all model parameters frozen, PROBE optimizes a lightweight pseudo-token embedding through a denoising reconstruction objective combined with a novel latent alignment constraint that anchors recovery to the spatiotemporal structure of the original concept. We make three contributions: (1) a multi-level evaluation framework spanning classifier-based detection, semantic similarity, temporal reactivation analysis, and human validation; (2) systematic experiments across three T2V architectures, three concept categories, and three erasure strategies revealing that all tested methods leave measurable residual capacity whose robustness correlates with intervention depth; and (3) the identification of temporal re-emergence, a video-specific failure mode where suppressed concepts progressively resurface across frames, invisible to frame-level metrics. These findings suggest that current erasure methods achieve output-level suppression rather than representational removal. We release our protocol to support reproducible safety auditing. Our code is available at https://github.com/YiweiXie/PRObingBasedEvaluation.

Method: Proposes a part-to-whole 3D generative world model with adaptive slot-gating mechanism that dynamically determines slot activation probabilities and consolidates redundant slots. Uses a class-agnostic prototype bank for cross-category shape sharing, and a lightweight 3D denoiser with unified diffusion objectives.

Result: Experiments show consistent gains in cross-category transfer and part-count extrapolation, with ablations confirming benefits of the prototype bank for shape-prior sharing and slot-gating for structural adaptation.

Conclusion: The approach successfully rethinks single-image 3D generation as learning adaptive part-whole hierarchies in flexible 3D latent space, enabling better generalization across diverse categories and structural complexity.

Abstract: Single-image 3D generation lies at the core of vision-to-graphics models in the real world. However, it remains a fundamental challenge to achieve reliable generalization across diverse semantic categories and highly variable structural complexity under sparse supervision. Existing approaches typically model objects in a monolithic manner or rely on a fixed number of parts, including recent part-aware models such as PartCrafter, which still require a labor-intensive user-specified part count. Such designs easily lead to overfitting, fragmented or missing structural components, and limited compositional generalization when encountering novel object layouts. To this end, this paper rethinks single-image 3D generation as learning an adaptive part-whole hierarchy in the flexible 3D latent space. We present a novel part-to-whole 3D generative world model that autonomously discovers latent structural slots by inferring soft and compositional masks directly from image tokens. Specifically, an adaptive slot-gating mechanism dynamically determines the slot-wise activation probabilities and smoothly consolidates redundant slots within different objects, ensuring that the emergent structure remains compact yet expressive across categories. Each distilled slot is then aligned to a learnable, class-agnostic prototype bank, enabling powerful cross-category shape sharing and denoising through universal geometric prototypes in the real world. Furthermore, a lightweight 3D denoiser is introduced to reconstruct geometry and appearance via unified diffusion objectives. Experiments show consistent gains in cross-category transfer and part-count extrapolation, and ablations confirm complementary benefits of the prototype bank for shape-prior sharing as well as slot-gating for structural adaptation.

Method: Proposes Pair Affinity Learning and Scoring (PALS) with learnable pair affinity to estimate interaction likelihood, Pair Affinity Modulation (PAM) for contextual reasoning, and Relation-Aware Matching (RAM) using vision-language grounding for cleaner supervision in pseudo-label generation.

Result: Extensive experiments on Action Genome dataset show substantial improvements across different baselines and backbones, achieving state-of-the-art performance in weakly-supervised video scene graph generation.

Conclusion: The proposed approach effectively addresses the fundamental discrepancy between weakly-supervised and fully-supervised pipelines by filtering non-interactive object pairs, leading to significant performance gains in video scene graph generation with reduced annotation costs.

Abstract: Weakly-supervised video scene graph generation (WS-VSGG) aims to parse video content into structured relational triplets without bounding box annotations and with only sparse temporal labeling, significantly reducing annotation costs. Without ground-truth bounding boxes, these methods rely on off-the-shelf detectors to generate object proposals, yet largely overlook a fundamental discrepancy from fullysupervised pipelines. Fully-supervised detectors implicitly filter out noninteractive objects, while off-the-shelf detectors indiscriminately detect all visible objects, overwhelming relation models with noisy pairs.We address this by introducing a learnable pair affinity that estimates the likelihood of interaction between subject-object pairs. Through Pair Affinity Learning and Scoring (PALS), pair affinity is incorporated into inferencetime ranking and further integrated into contextual reasoning through Pair Affinity Modulation (PAM), enabling the model to suppress noninteractive pairs and focus on relationally meaningful ones. To provide cleaner supervision for pair affinity learning, we further propose Relation- Aware Matching (RAM), which leverages vision-language grounding to resolve class-level ambiguity in pseudo-label generation. Extensive experiments on Action Genome demonstrate that our approach consistently yields substantial improvements across different baselines and backbones, achieving state-of-the-art WS-VSGG performance.

Method: Introduces Continual Multimodal Prompt Memory Bank (CMPMB) to distill prototypical normal patterns from both visual and textual domains across tasks, and a Defect-Semantic-Guided Adaptive Fusion Mechanism (DSG-AFM) with Adaptive Normalization Module and Dynamic Fusion Strategy for enhanced accuracy and robustness.

Result: Achieves state-of-the-art performance on MVTec AD and VisA datasets for both image-level AUROC and pixel-level AUPR metrics.

Conclusion: Multimodal prompting effectively enhances unsupervised continual anomaly detection by leveraging both visual and textual information to better represent normality across sequential learning tasks.

Abstract: Unsupervised Continuous Anomaly Detection (UCAD) is gaining attention for effectively addressing the catastrophic forgetting and heavy computational burden issues in traditional Unsupervised Anomaly Detection (UAD). However, existing UCAD approaches that rely solely on visual information are insufficient to capture the manifold of normality in complex scenes, thereby impeding further gains in anomaly detection accuracy. To overcome this limitation, we propose an unsupervised continual anomaly detection framework grounded in multimodal prompting. Specifically, we introduce a Continual Multimodal Prompt Memory Bank (CMPMB) that progressively distills and retains prototypical normal patterns from both visual and textual domains across consecutive tasks, yielding a richer representation of normality. Furthermore, we devise a Defect-Semantic-Guided Adaptive Fusion Mechanism (DSG-AFM) that integrates an Adaptive Normalization Module (ANM) with a Dynamic Fusion Strategy (DFS) to jointly enhance detection accuracy and adversarial robustness. Benchmark experiments on MVTec AD and VisA datasets show that our approach achieves state-of-the-art (SOTA) performance on image-level AUROC and pixel-level AUPR metrics.

Method: Proposes FSCE framework with frequency-spatial shallow feature adaptive enhancement (DSAF) module using spatial multi-scale convolution and frequency-domain wavelet convolution. Uses teacher-student learning with online knowledge distillation to guide focus on target regions and enhance noise robustness.

Result: DSAFNet-L achieves competitive/superior performance on MSTAR, FUSARShip and OpenSARShip datasets; DSAFNet-M reduces model complexity while maintaining comparable accuracy. Framework shows strong cross-model generalization.

Conclusion: The FSCE framework effectively enhances SAR target recognition under noisy conditions through collaborative optimization of attention transfer and noise-resilient representation learning.

Abstract: Synthetic aperture radar automatic target recognition (SAR ATR) is of considerable importance in marine navigation and disaster monitoring. However, the coherent speckle noise inherent in SAR imagery often obscures salient target features, leading to degraded recognition accuracy and limited model generalization. To address this issue, this paper proposes a target-aware frequency-spatial enhancement framework with noise-resilient knowledge guidance (FSCE) for SAR target recognition. The proposed framework incorporates a frequency-spatial shallow feature adaptive enhancement (DSAF) module, which processes shallow features through spatial multi-scale convolution and frequency-domain wavelet convolution. In addition, a teacher-student learning paradigm combined with an online knowledge distillation method (KD) is employed to guide the student network to focus more effectively on target regions, thereby enhancing its robustness to high-noise backgrounds. Through the collaborative optimization of attention transfer and noise-resilient representation learning, the proposed approach significantly improves the stability of target recognition under noisy conditions. Based on the FSCE framework, two network architectures with different performance emphases are developed: lightweight DSAFNet-M and high-precision DSAFNet-L. Extensive experiments are conducted on the MSTAR, FUSARShip and OpenSARShip datasets. The results show that DSAFNet-L achieves competitive or superior performance compared with various methods on three datasets; DSAFNet-M significantly reduces the model complexity while maintaining comparable accuracy. These results indicate that the proposed FSCE framework exhibits strong cross-model generalization.

Method: Domain adaptation of Meta’s Segment Anything Model 2 (SAM-2) for cataract surgery videos, plus an interactive annotation framework combining sparse prompts with video-based mask propagation to accelerate dataset creation.

Result: Achieves real-time semantic segmentation with high accuracy for cataract surgery videos, significantly reduces annotation time, demonstrates strong zero-shot generalization to glaucoma trabeculectomy procedures, and releases model/toolkit as open-source.

Conclusion: CataractSAM-2 establishes a foundation for expanding anterior ophthalmic surgical datasets and advancing real-time AI-driven solutions in medical robotics and surgical video understanding through domain adaptation and efficient annotation tools.

Abstract: We present CataractSAM-2, a domain-adapted extension of Meta’s Segment Anything Model 2, designed for real-time semantic segmentation of cataract ophthalmic surgery videos with high accuracy. Positioned at the intersection of computer vision and medical robotics, CataractSAM-2 enables precise intraoperative perception crucial for robotic-assisted and computer-guided surgical systems. Furthermore, to alleviate the burden of manual labeling, we introduce an interactive annotation framework that combines sparse prompts with video-based mask propagation. This tool significantly reduces annotation time and facilitates the scalable creation of high-quality ground-truth masks, accelerating dataset development for ocular anterior segment surgeries. We also demonstrate the model’s strong zero-shot generalization to glaucoma trabeculectomy procedures, confirming its cross-procedural utility and potential for broader surgical applications. The trained model and annotation toolkit are released as open-source resources, establishing CataractSAM-2 as a foundation for expanding anterior ophthalmic surgical datasets and advancing real-time AI-driven solutions in medical robotics, as well as surgical video understanding.

Method: Introduces Compositional Privacy Risk Taxonomy (CPRT), a regulation-aware framework organizing visual attributes by standalone identifiability and compositional harm potential. Defines four graded severity levels with interpretable scoring function. Constructs taxonomy-aligned dataset of 6.7K images with ground-truth compositional risk scores. Evaluates frontier and open-weight VLMs, and introduces deployable 8B supervised fine-tuned model for compositional privacy assessment.

Result: Frontier models align well with compositional severity when given structured guidance but systematically underestimate composition-driven risks. Smaller models struggle with graded privacy reasoning. The introduced 8B SFT model closely matches frontier-level performance on compositional privacy assessment.

Conclusion: Privacy assessment requires compositional reasoning beyond binary classification. The CPRT framework provides a more nuanced approach to visual privacy evaluation, and the 8B SFT model demonstrates that deployable models can achieve frontier-level performance on compositional privacy assessment tasks.

Abstract: Existing visual privacy benchmarks largely treat privacy as a binary property, labeling images as private or non-private based on visible sensitive content. We argue that privacy is fundamentally compositional. Attributes that are benign in isolation may combine to produce severe privacy violations. We introduce the Compositional Privacy Risk Taxonomy (CPRT), a regulation-aware framework that organizes visual attributes according to standalone identifiability and compositional harm potential. CPRT defines four graded severity levels and is paired with an interpretable scoring function that assigns continuous privacy severity scores. We further construct a taxonomy-aligned dataset of 6.7K images and derive ground-truth compositional risk scores. By evaluating frontier and open-weight VLMs we find that frontier models align well with compositional severity when provided structured guidance, but systematically underestimate composition-driven risks. Smaller models struggle to internalize graded privacy reasoning. To bridge this gap, we introduce a deployable 8B supervised fine-tuned (SFT) model that closely matches frontier-level performance on compositional privacy assessment.

Method: Proposes hardness-aware curriculum learning that dynamically selects pseudo-labeled samples based on difficulty (easy to complex). Includes multi-stage and adaptive curriculum strategies to replace fixed-threshold filtering. Also introduces structured data augmentation tailored for rotation estimation, assembling composite images from augmented patches while preserving geometric integrity.

Result: Comprehensive experiments on PASCAL3D+ and ObjectNet3D show the method outperforms existing supervised and semi-supervised baselines, particularly in low-data regimes.

Conclusion: The hardness-aware curriculum learning framework and structured augmentation approach are effective for semi-supervised 3D rotation regression from limited 2D images.

Abstract: Regressing 3D rotations of objects from 2D images is a crucial yet challenging task, with broad applications in autonomous driving, virtual reality, and robotic control. Existing rotation regression models often rely on large amounts of labeled data for training or require additional information beyond 2D images, such as point clouds or CAD models. Therefore, exploring semi-supervised rotation regression using only a limited number of labeled 2D images is highly valuable. While recent work FisherMatch introduces semi-supervised learning to rotation regression, it suffers from rigid entropy-based pseudo-label filtering that fails to effectively distinguish between reliable and unreliable unlabeled samples. To address this limitation, we propose a hardness-aware curriculum learning framework that dynamically selects pseudo-labeled samples based on their difficulty, progressing from easy to complex examples. We introduce both multi-stage and adaptive curriculum strategies to replace fixed-threshold filtering with more flexible, hardness-aware mechanisms. Additionally, we present a novel structured data augmentation strategy specifically tailored for rotation estimation, which assembles composite images from augmented patches to introduce feature diversity while preserving critical geometric integrity. Comprehensive experiments on PASCAL3D+ and ObjectNet3D demonstrate that our method outperforms existing supervised and semi-supervised baselines, particularly in low-data regimes, validating the effectiveness of our curriculum learning framework and structured augmentation approach.

Method: SARe uses a two-stage approach: 1) SARe-Gen jointly predicts fracture-surface token probabilities and an inter-fragment contact graph using query-point-based conditioning and aligned local geometric tokens from a frozen geometry encoder; 2) SARe-Refine performs inference-time refinement by verifying candidate contacts with geometric-consistency checks, selecting reliable substructures, and resampling uncertain regions.

Result: State-of-the-art performance across synthetic fractures, simulated fractures from scanned real objects, and real physically fractured scans, with more graceful degradation and higher success rates as fragment count increases in large-scale reassembly.

Conclusion: SARe addresses key challenges in many-fragment 3D reassembly through explicit contact modeling and inference-time refinement, demonstrating robust performance across diverse fracture scenarios.

Abstract: 3D fragment reassembly aims to recover the rigid poses of unordered fragment point clouds or meshes in a common object coordinate system to reconstruct the complete shape. The problem becomes particularly challenging as the number of fragments grows, since the target shape is unknown and fragments provide weak semantic cues. Existing end-to-end approaches are prone to cascading failures due to unreliable contact reasoning, most notably inaccurate fragment adjacencies. To address this, we propose Structure-Aware Reassembly (SARe), a generative framework with SARe-Gen for Euclidean-space assembly generation and SARe-Refine for inference-time refinement, with explicit contact modeling. SARe-Gen jointly predicts fracture-surface token probabilities and an inter-fragment contact graph to localize contact regions and infer candidate adjacencies. It adopts a query-point-based conditioning scheme and extracts aligned local geometric tokens at query locations from a frozen geometry encoder, yielding queryable structural representations without additional structural pretraining. We further introduce an inference-time refinement stage, SARe-Refine. By verifying candidate contact edges with geometric-consistency checks, it selects reliable substructures and resamples the remaining uncertain regions while keeping verified parts fixed, leading to more stable and consistent assemblies in the many-fragment regime. We evaluate SARe across three settings, including synthetic fractures, simulated fractures from scanned real objects, and real physically fractured scans. The results demonstrate state-of-the-art performance, with more graceful degradation and higher success rates as the fragment count increases in challenging large-scale reassembly.

Method: Two innovations: 1) Progressive Injection Schedule with continuous decay functions (sigmoid, cosine, linear) instead of binary cutoffs, 2) Channel-Selective Latent Perturbation that estimates per-channel importance and applies differentiated perturbation strengths.

Result: On PIE-Bench benchmark (700 images, 10 editing types), AdaEdit achieves 8.7% reduction in LPIPS, 2.6% improvement in SSIM, and 2.3% improvement in PSNR over baselines while maintaining competitive CLIP similarity.

Conclusion: AdaEdit effectively resolves the injection dilemma in flow matching models through adaptive strategies, enabling better balance between source preservation and target generation without training.

Abstract: Inversion-based image editing in flow matching models has emerged as a powerful paradigm for training-free, text-guided image manipulation. A central challenge in this paradigm is the injection dilemma: injecting source features during denoising preserves the background of the original image but simultaneously suppresses the model’s ability to synthesize edited content. Existing methods address this with fixed injection strategies – binary on/off temporal schedules, uniform spatial mixing ratios, and channel-agnostic latent perturbation – that ignore the inherently heterogeneous nature of injection demand across both the temporal and channel dimensions. In this paper, we present AdaEdit, a training-free adaptive editing framework that resolves this dilemma through two complementary innovations. First, we propose a Progressive Injection Schedule that replaces hard binary cutoffs with continuous decay functions (sigmoid, cosine, or linear), enabling a smooth transition from source-feature preservation to target-feature generation and eliminating feature discontinuity artifacts. Second, we introduce Channel-Selective Latent Perturbation, which estimates per-channel importance based on the distributional gap between the inverted and random latents and applies differentiated perturbation strengths accordingly – strongly perturbing edit-relevant channels while preserving structure-encoding channels. Extensive experiments on the PIE-Bench benchmark (700 images, 10 editing types) demonstrate that AdaEdit achieves an 8.7% reduction in LPIPS, a 2.6% improvement in SSIM, and a 2.3% improvement in PSNR over strong baselines, while maintaining competitive CLIP similarity. AdaEdit is fully plug-and-play and compatible with multiple ODE solvers including Euler, RF-Solver, and FireFlow. Code is available at https://github.com/leeguandong/AdaEdit

Method: Proposes 4DGS360 with AnchorTAP3D tracker that uses confident 2D track points as anchors to produce reinforced 3D point trajectories, suppressing drift and providing reliable 3D-native initialization that preserves geometry in occluded regions.

Result: Achieves state-of-the-art performance on iPhone360 (new benchmark), iPhone, and DAVIS datasets, both qualitatively and quantitatively, enabling coherent 360° 4D reconstructions.

Conclusion: 4DGS360 successfully addresses the challenge of 360° dynamic object reconstruction from monocular video through advanced 3D-native initialization, outperforming existing methods and introducing a new benchmark for comprehensive evaluation.

Abstract: We introduce 4DGS360, a diffusion-free framework for 360$^{\circ}$ dynamic object reconstruction from casual monocular video. Existing methods often fail to reconstruct consistent 360$^{\circ}$ geometry, as their heavy reliance on 2D-native priors causes initial points to overfit to visible surface in each training view. 4DGS360 addresses this challenge through a advanced 3D-native initialization that mitigates the geometric ambiguity of occluded regions. Our proposed 3D tracker, AnchorTAP3D, produces reinforced 3D point trajectories by leveraging confident 2D track points as anchors, suppressing drift and providing reliable initialization that preserves geometry in occluded regions. This initialization, combined with optimization, yields coherent 360$^{\circ}$ 4D reconstructions. We further present iPhone360, a new benchmark where test cameras are placed up to 135$^{\circ}$ apart from training views, enabling 360$^{\circ}$ evaluation that existing datasets cannot provide. Experiments show that 4DGS360 achieves state-of-the-art performance on the iPhone360, iPhone, and DAVIS datasets, both qualitatively and quantitatively.

Method: Proposes a training-free detection method that measures representation sensitivity to structured frequency perturbations. The approach uses Fourier transforms to generate perturbations and analyzes how image representations respond to these frequency-based modifications to detect minute manipulations characteristic of AI-generated content.

Result: Achieves 1-2 orders of magnitude faster inference than most training-free detectors. On the OpenFake benchmark, improves AUC by nearly 10% compared to state-of-the-art methods while maintaining substantially lower computational cost.

Conclusion: The proposed frequency perturbation sensitivity method provides an effective, computationally lightweight solution for AI-generated image detection that outperforms existing approaches in both accuracy and efficiency.

Abstract: The rapid progress of text-to-image models has made AI-generated images increasingly realistic, posing significant challenges for accurate detection of generated content. While training-based detectors often suffer from limited generalization to unseen images, training-free approaches offer better robustness, yet struggle to capture subtle discrepancies between real and synthetic images. In this work, we propose a training-free AI-generated image detection method that measures representation sensitivity to structured frequency perturbations, enabling detection of minute manipulations. The proposed method is computationally lightweight, as perturbation generation requires only a single Fourier transform for an input image. As a result, it achieves one to two orders of magnitude faster inference than most training-free detectors.Extensive experiments on challenging benchmarks demonstrate the efficacy of our method over state-of-the-art (SoTA). In particular, on OpenFake benchmark, our method improves AUC by nearly $10%$ compared to SoTA, while maintaining substantially lower computational cost.

Method: Proposes PGR-Net with: 1) Data-driven spatial prior set capturing tumor distribution/scale characteristics, 2) Hierarchical Top-K ROI decision mechanism for progressive lesion candidate selection, 3) WinGS-ROI module using multi-window Gaussian templates with spatial decay for center-enhanced guidance maps, 4) Windowed RetNet backbone for enhanced localization reliability.

Result: Outperforms existing approaches on BraTS-2019/2023 and MSD Task01 with only 8.64M parameters, achieving Dice scores of 89.02%, 91.82%, and 89.67% on Whole Tumor region.

Conclusion: PGR-Net effectively addresses spatial sparsity in brain tumor segmentation by incorporating spatial priors and ROI reasoning, providing stable and precise segmentation with computational efficiency.

Abstract: Brain tumor MRI segmentation is essential for clinical diagnosis and treatment planning, enabling accurate lesion detection and radiotherapy target delineation. However, tumor lesions occupy only a small fraction of the volumetric space, resulting in severe spatial sparsity, while existing segmentation networks often overlook clinically observed spatial priors of tumor occurrence, leading to redundant feature computation over extensive background regions. To address this issue, we propose PGR-Net (Prior-Guided ROI Reasoning Network) - an explicit ROI-aware framework that incorporates a data-driven spatial prior set to capture the distribution and scale characteristics of tumor lesions, providing global guidance for more stable segmentation. Leveraging these priors, PGR-Net introduces a hierarchical Top-K ROI decision mechanism that progressively selects the most confident lesion candidate regions across encoder layers to improve localization precision. We further develop the WinGS-ROI (Windowed Gaussian-Spatial Decay ROI) module, which uses multi-window Gaussian templates with a spatial decay function to produce center-enhanced guidance maps, thus directing feature learning throughout the network. With these ROI features, a windowed RetNet backbone is adopted to enhance localization reliability. Experiments on BraTS-2019/2023 and MSD Task01 show that PGR-Net consistently outperforms existing approaches while using only 8.64M Params, achieving Dice scores of 89.02%, 91.82%, and 89.67% on the Whole Tumor region. Code is available at https://github.com/CNU-MedAI-Lab/PGR-Net.

Method: Proposes a Test-time Calibration from Experience and Intuition (TCEI) framework with two systems: 1) Intuitive system uses transient memory to recall recently observed objects for rapid predictions, 2) Experiential system leverages accumulated experience from prior test videos to reassess and calibrate intuitive predictions. Both confident and uncertain objects are exploited as historical priors and reflective cases.

Result: Extensive experiments show TCEI consistently achieves superior performance across multiple benchmark datasets and significantly enhances model adaptability under distribution shifts.

Conclusion: TCEI effectively addresses distribution shifts in MOT by combining intuitive and experiential systems, outperforming existing test-time adaptation methods through better temporal consistency and identity association handling.

Abstract: Multiple Object Tracking (MOT) has long been a fundamental task in computer vision, with broad applications in various real-world scenarios. However, due to distribution shifts in appearance, motion pattern, and catagory between the training and testing data, model performance degrades considerably during online inference in MOT. Test-Time Adaptation (TTA) has emerged as a promising paradigm to alleviate such distribution shifts. However, existing TTA methods often fail to deliver satisfactory results in MOT, as they primarily focus solely on frame-level adaptation while neglecting temporal consistency and identity association across frames and videos. Inspired by human decision-making process, this paper propose a Test-time Calibration from Experience and Intuition (TCEI) framework. In this framework, the Intuitive system utilizes transient memory to recall recently observed objects for rapid predictions, while the Experiential system leverages the accumulated experience from prior test videos to reassess and calibrate these intuitive predictions. Furthermore, both confident and uncertain objects during online testing are exploited as historical priors and reflective cases, respectively, enabling the model to adapt to the testing environment and alleviate performance degradation. Extensive experiments demonstrate that the proposed TCEI framework consistently achieves superior performance across multiple benchmark datasets and significantly enhances the model’s adaptability under distribution shifts. The code will be released at https://github.com/1941Zpf/TCEI.

Method: Proposes SparseVoxelDet, a fully sparse object detector where backbone feature extraction, feature pyramid fusion, and detection head operate exclusively on occupied voxel positions using 3D sparse convolutions. No dense feature tensors are created at any stage, processing only active voxels in the scene.

Result: Achieves 83.38% mAP at IoU 0.50 on FRED benchmark (629,832 frames), processing only 14,900 active voxels per frame (0.23% of grid). Shows 858× GPU memory compression and 3,670× storage reduction vs dense 3D voxel tensor. Accuracy gap to dense baseline (87.68% mAP) is mostly due to box regression precision rather than detection capability.

Conclusion: Native sparse processing is viable for event-camera object detection, exploiting structural sparsity without requiring neuromorphic hardware. The framework’s representation cost scales with scene activity rather than pixel count, making it increasingly valuable as event cameras scale to higher resolutions.

Abstract: Event cameras produce asynchronous, high-dynamic-range streams well suited for detecting small, fast-moving drones, yet most event-based detectors convert the sparse event stream into dense tensors, discarding the representational efficiency of neuromorphic sensing. We propose SparseVoxelDet, to our knowledge the first fully sparse object detector for event cameras, in which backbone feature extraction, feature pyramid fusion, and the detection head all operate exclusively on occupied voxel positions through 3D sparse convolutions; no dense feature tensor is instantiated at any stage of the pipeline. On the FRED benchmark (629,832 annotated frames), SparseVoxelDet achieves 83.38% mAP at 50 while processing only 14,900 active voxels per frame (0.23% of the T.H.W grid), compared to 409,600 pixels for the dense YOLOv11 baseline (87.68% mAP at 50). Relaxing the IoU threshold from 0.50 to 0.40 recovers mAP to 89.26%, indicating that the remaining accuracy gap is dominated by box regression precision rather than detection capability. The sparse representation yields 858 times GPU memory compression and 3,670 times storage reduction relative to the equivalent dense 3D voxel tensor, with data-structure size that scales with scene dynamics rather than sensor resolution. Error forensics across 119,459 test frames confirms that 71 percent of failures are localization near-misses rather than missed targets. These results demonstrate that native sparse processing is a viable paradigm for event-camera object detection, exploiting the structural sparsity of neuromorphic sensor data without requiring neuromorphic computing hardware, and providing a framework whose representation cost is governed by scene activity rather than pixel count, a property that becomes increasingly valuable as event cameras scale to higher resolutions.

Method: FedCVU proposes three components: VS-Norm preserves view-specific normalization parameters to handle different camera statistics; CV-Align uses lightweight contrastive regularization to improve cross-view representation alignment; and SLA (Selective Layer Aggregation) reduces communication by selectively aggregating only critical layers without sacrificing accuracy.

Result: Extensive experiments on action understanding and person re-identification tasks under cross-view protocols show FedCVU consistently boosts unseen-view accuracy while maintaining strong seen-view performance, outperforming state-of-the-art FL baselines and demonstrating robustness to domain heterogeneity and communication constraints.

Conclusion: FedCVU effectively addresses key challenges in federated cross-view video understanding by handling view heterogeneity, aligning representations across cameras, and reducing communication overhead, making it a practical solution for privacy-preserving multi-camera video analysis.

Abstract: Federated learning (FL) has emerged as a promising paradigm for privacy-preserving multi-camera video understanding. However, applying FL to cross-view scenarios faces three major challenges: (i) heterogeneous viewpoints and backgrounds lead to highly non-IID client distributions and overfitting to view-specific patterns, (ii) local distribution biases cause misaligned representations that hinder consistent cross-view semantics, and (iii) large video architectures incur prohibitive communication overhead. To address these issues, we propose FedCVU, a federated framework with three components: VS-Norm, which preserves normalization parameters to handle view-specific statistics; CV-Align, a lightweight contrastive regularization module to improve cross-view representation alignment; and SLA, a selective layer aggregation strategy that reduces communication without sacrificing accuracy. Extensive experiments on action understanding and person re-identification tasks under a cross-view protocol demonstrate that FedCVU consistently boosts unseen-view accuracy while maintaining strong seen-view performance, outperforming state-of-the-art FL baselines and showing robustness to domain heterogeneity and communication constraints.

Method: OmniFM leverages frequency-domain insights that low-frequency spectral components exhibit cross-modality consistency and encode modality-invariant anatomical structures. It integrates: (1) Global Spectral Knowledge Retrieval to inject global frequency priors, (2) Embedding-wise Cross-Attention Fusion to align representations, (3) Prefix-Suffix Spectral Prompting to jointly condition global and personalized cues, all regularized by a Spectral-Proximal Alignment objective that stabilizes aggregation.

Result: Experiments on real-world datasets show OmniFM consistently surpasses state-of-the-art FL baselines across intra- and cross-modality heterogeneity, achieving superior results under both fine-tuning and training-from-scratch setups.

Conclusion: OmniFM provides a unified federated learning framework that effectively handles modality heterogeneity and diverse medical imaging tasks through frequency-domain regularization and cross-modality alignment techniques.

Abstract: Federated learning (FL) has become a promising paradigm for collaborative medical image analysis, yet existing frameworks remain tightly coupled to task-specific backbones and are fragile under heterogeneous imaging modalities. Such constraints hinder real-world deployment, where institutions vary widely in modality distributions and must support diverse downstream tasks. To address this limitation, we propose OmniFM, a modality- and task-agnostic FL framework that unifies training across classification, segmentation, super-resolution, visual question answering, and multimodal fusion without re-engineering the optimization pipeline. OmniFM builds on a key frequency-domain insight: low-frequency spectral components exhibit strong cross-modality consistency and encode modality-invariant anatomical structures. Accordingly, OmniFM integrates (i) Global Spectral Knowledge Retrieval to inject global frequency priors, (ii) Embedding-wise Cross-Attention Fusion to align representations, and (iii) Prefix-Suffix Spectral Prompting to jointly condition global and personalized cues, together regularized by a Spectral-Proximal Alignment objective that stabilizes aggregation. Experiments on real-world datasets show that OmniFM consistently surpasses state-of-the-art FL baselines across intra- and cross-modality heterogeneity, achieving superior results under both fine-tuning and training-from-scratch setups.

Method: Proposes HumanOmni-Speaker with Visual Delta Encoder that samples raw video at 25 fps and compresses inter-frame motion residuals into just 6 tokens per frame to capture fine-grained visemes and speaker trajectories without token explosion.

Result: Demonstrates strong multimodal synergy, enabling end-to-end lip-reading and high-precision spatial localization without intrusive cropping, achieving superior performance across speaker-centric tasks.

Conclusion: The approach overcomes architectural perception gaps in multimodal LLMs by providing true spatio-temporal identity binding for complex conversational dynamics through efficient high-frequency visual processing.

Abstract: While Omni-modal Large Language Models have made strides in joint sensory processing, they fundamentally struggle with a cornerstone of human interaction: deciphering complex, multi-person conversational dynamics to accurately answer Who said what and when.'' Current models suffer from an illusion of competence’’ – they exploit visual biases in conventional benchmarks to bypass genuine cross-modal alignment, while relying on sparse, low-frame-rate visual sampling that destroys crucial high-frequency dynamics like lip movements. To shatter this illusion, we introduce Visual-Registered Speaker Diarization and Recognition (VR-SDR) and the HumanOmni-Speaker Benchmark. By strictly eliminating visual shortcuts, this rigorous paradigm demands true end-to-end spatio-temporal identity binding using only natural language queries. To overcome the underlying architectural perception gap, we propose HumanOmni-Speaker, powered by a Visual Delta Encoder. By sampling raw video at 25 fps and explicitly compressing inter-frame motion residuals into just 6 tokens per frame, it captures fine-grained visemes and speaker trajectories without triggering a catastrophic token explosion. Ultimately, HumanOmni-Speaker demonstrates strong multimodal synergy, natively enabling end-to-end lip-reading and high-precision spatial localization without intrusive cropping, and achieving superior performance across a wide spectrum of speaker-centric tasks.

Method: The method decouples the refractive boundary from target objects: uses a neural height field to model the refractive water surface geometry, and a 3D Gaussian field for the underlying scene. It implements refraction-aware Gaussian ray tracing using Snell’s law to compute non-linear ray trajectories.

Result: RefracGS outperforms prior refractive methods in visual quality on both synthetic and real-world scenes with complex waves, achieves 15x faster training, and enables real-time rendering at 200 FPS.

Conclusion: The framework successfully addresses the challenge of novel view synthesis through refractive surfaces by jointly optimizing refractive surface and scene representations with physically accurate ray tracing, enabling high-fidelity rendering and view-consistent surface recovery.

Abstract: Novel view synthesis (NVS) through non-planar refractive surfaces presents fundamental challenges due to severe, spatially varying optical distortions. While recent representations like NeRF and 3D Gaussian Splatting (3DGS) excel at NVS, their assumption of straight-line ray propagation fails under these conditions, leading to significant artifacts. To overcome this limitation, we introduce RefracGS, a framework that jointly reconstructs the refractive water surface and the scene beneath the interface. Our key insight is to explicitly decouple the refractive boundary from the target objects: the refractive surface is modeled via a neural height field, capturing wave geometry, while the underlying scene is represented as a 3D Gaussian field. We formulate a refraction-aware Gaussian ray tracing approach that accurately computes non-linear ray trajectories using Snell’s law and efficiently renders the underlying Gaussian field while backpropagating the loss gradients to the parameterized refractive surface. Through end-to-end joint optimization of both representations, our method ensures high-fidelity NVS and view-consistent surface recovery. Experiments on both synthetic and real-world scenes with complex waves demonstrate that RefracGS outperforms prior refractive methods in visual quality, while achieving 15x faster training and real-time rendering at 200 FPS. The project page for RefracGS is available at https://yimgshao.github.io/refracgs/.

Method: PPGL-Swarm uses an agentic system that decomposes diagnosis into micro-tasks assigned to specialized agents. It includes automated GAPP scoring with quantified cellularity and Ki-67, genotype risk alerts, and multimodal report generation. The system employs knowledge enhancement for gene and table agents to interpret genotype and laboratory findings, and uses reinforcement learning during training to refine tool selection and task assignment.

Result: The system generates comprehensive diagnostic reports including automated GAPP scoring, genotype risk alerts, and integrated multimodal evidence. It provides an auditable reasoning trail through the decomposition of diagnosis into micro-tasks handled by specialized agents.

Conclusion: PPGL-Swarm addresses limitations of current PPGL diagnostic approaches by automating complex scoring, incorporating genotype information, and providing traceable reasoning through an agentic architecture with specialized agents and knowledge enhancement.

Abstract: Pheochromocytomas and paragangliomas (PPGLs) are rare neuroendocrine tumors, of which 15-25% develop metastatic disease with 5-year survival rates reported as low as 34%. PPGL may indicate hereditary syndromes requiring stricter, syndrome-specific treatment and surveillance, but clinicians often fail to recognize these associations in routine care. Clinical practice uses GAPP score for PPGL grading, but several limitations remain for PPGL diagnosis: (1) GAPP scoring demands a high workload for clinician because it requires the manual evaluation of six independent components; (2) key components such as cellularity and Ki-67 are often evaluated with subjective criteria; (3) several clinically relevant metastatic risk factors are not captured by GAPP, such as SDHB mutations, which have been associated with reported metastatic rates of 35-75%. Agent-driven diagnostic systems appear promising, but most lack traceable reasoning for decision-making and do not incorporate domain-specific knowledge such as PPGL genotype information. To address these limitations, we present PPGL-Swarm, an agentic PPGL diagnostic system that generates a comprehensive report, including automated GAPP scoring (with quantified cellularity and Ki-67), genotype risk alerts, and multimodal report with integrated evidence. The system provides an auditable reasoning trail by decomposing diagnosis into micro-tasks, each assigned to a specialized agent. The gene and table agents use knowledge enhancement to better interpret genotype and laboratory findings, and during training we use reinforcement learning to refine tool selection and task assignment.

Method: Formulates token reduction as a learnable compression mapping that unifies pruning and merging approaches. Introduces a data-efficient training paradigm to learn optimal compression mappings with limited computational costs, making it prompt-agnostic and suitable for multi-turn settings.

Result: Extensive experiments on MT-VQA benchmarks across multiple LVLM architectures show MetaCompress achieves superior efficiency-accuracy trade-offs while maintaining strong generalization across dialogue turns.

Conclusion: MetaCompress overcomes limitations of heuristic token reduction methods for multi-turn VQA, providing an effective learning-based approach that balances computational efficiency with visual reasoning accuracy in practical dialogue scenarios.

Abstract: Large Vision-Language Models (LVLMs) excel in visual understanding and reasoning, but the excessive visual tokens lead to high inference costs. Although recent token reduction methods mitigate this issue, they mainly target single-turn Visual Question Answering (VQA), leaving the more practical multi-turn VQA (MT-VQA) scenario largely unexplored. MT-VQA introduces additional challenges, as subsequent questions are unknown beforehand and may refer to arbitrary image regions, making existing reduction strategies ineffective. Specifically, current approaches fall into two categories: prompt-dependent methods, which bias toward the initial text prompt and discard information useful for subsequent turns; prompt-agnostic ones, which, though technically applicable to multi-turn settings, rely on heuristic reduction metrics such as attention scores, leading to suboptimal performance. In this paper, we propose a learning-based prompt-agnostic method, termed MetaCompress, overcoming the limitations of heuristic designs. We begin by formulating token reduction as a learnable compression mapping, unifying existing formats such as pruning and merging into a single learning objective. Upon this formulation, we introduce a data-efficient training paradigm capable of learning optimal compression mappings with limited computational costs. Extensive experiments on MT-VQA benchmarks and across multiple LVLM architectures demonstrate that MetaCompress achieves superior efficiency-accuracy trade-offs while maintaining strong generalization across dialogue turns. Our code is available at https://github.com/MArSha1147/MetaCompress.

Method: Fine-tuned state-of-the-art LVLMs with two approaches: Direct Counting (predict only total count) and Point-then-Count (generate object coordinates first, then count). Analyzed spatial biases and conducted mechanistic analyses.

Result: Point-then-Count achieved higher out-of-distribution generalization than Direct Counting. Predicted points were accurately grounded (over 89% F1), but performance varied across image regions revealing spatial biases. Gains attributed to spatial information in coordinates.

Conclusion: Pointing helps LVLMs learn generalizable skills rather than overfitting, coordinates provide valuable spatial information for counting, but spatial biases exist that affect reliability as visual explanations.

Abstract: Pointing increases the accuracy and explainability of Large Vision-Language Models (LVLMs) by modeling grounding and reasoning as explicit sequential steps. The model grounds the objects mentioned in the natural-language query by predicting their coordinates, and then generates an answer conditioned on these points. While pointing has been shown to increase LVLMs’ accuracy, it is unclear which mechanism supports these gains and its relevance in cognitive tasks. In addition, the reliability of the intermediate points remains understudied, limiting their use as visual explanations. In this work, we study the role of pointing in a cognitive task: zero-shot counting from a visual scene. We fine-tune state-of-the-art LVLMs following two approaches: Direct Counting, where models only predict the total number of objects, and Point-then-Count, where LVLMs generate the target objects’ coordinates followed by their count. The results show that Point-then-Count achieves higher out-of-distribution generalization, suggesting that coordinates help LVLMs learn skills rather than overfitting on narrow tasks. Although predicted points are accurately grounded in the image in over 89% of cases (as measured by F1), performance varies across image regions, revealing spatial biases. Finally, mechanistic analyses show that gains in counting arise from the spatial information encoded in the coordinates.

Method: Introduces DaP-ICoT with two key components: (1) Dynamic Visual Thought Integration - adaptively introduces visual inputs based on reasoning needs to reduce redundancy; (2) Precise Visual Thought Guidance - ensures visual representations are semantically coherent and contextually aligned.

Result: DaP-ICoT achieves state-of-the-art performance across multiple benchmarks and models. It significantly reduces the number of inserted images, leading to a 72.6% decrease in token consumption, enabling more efficient ICoT reasoning.

Conclusion: DaP-ICoT addresses key limitations in current ICoT methods by making visual integration dynamic and precise, resulting in both improved performance and computational efficiency for multimodal reasoning.

Abstract: Recently, Interleaved-modal Chain-of-Thought (ICoT) reasoning has achieved remarkable success by leveraging both multimodal inputs and outputs, attracting increasing attention. While achieving promising performance, current ICoT methods still suffer from two major limitations: (1) Static Visual Thought Positioning, which statically inserts visual information at fixed steps, resulting in inefficient and inflexible reasoning; and (2) Broken Visual Thought Representation, which involves discontinuous and semantically incoherent visual tokens. To address these limitations, we introduce Interleaved-modal Chain-of-Thought reasoning with Dynamic and Precise Visual Thoughts (DaP-ICoT), which incorporates two key components: (1) Dynamic Visual Thought Integration adaptively introduces visual inputs based on reasoning needs, reducing redundancy and improving efficiency. (2) Precise Visual Thought Guidance ensures visual semantically coherent and contextually aligned representations. Experiments across multiple benchmarks and models demonstrate that DaP-ICoT achieves state-of-the-art performance. In addition, DaP-ICoT significantly reduces the number of inserted images, leading to a 72.6% decrease in token consumption, enabling more efficient ICoT reasoning.

Method: Two main components: 1) RS-FLUX - fine-tuning FLUX diffusion model on 100k+ curated remote sensing images to build domain prior; 2) SHARP - training-free method with rational fractional time schedule k_rs(t) for dynamic RoPE rescaling that applies strong positional promotion early and relaxes it during detail recovery.

Result: SHARP consistently outperforms all training-free baselines across six square and rectangular resolutions on CLIP Score, Aesthetic Score, and HPSv2 metrics, with widening margins at more aggressive extrapolation factors and negligible computational overhead.

Conclusion: The combination of domain-specialized RS-FLUX prior and dynamic SHARP resolution promotion enables high-quality remote sensing image generation at large resolutions without expensive retraining, addressing key barriers in RS synthesis.

Abstract: Text-to-image generation powered by Diffusion Transformers (DiTs) has made remarkable strides, yet remote sensing (RS) synthesis lags behind due to two barriers: the absence of a domain-specialized DiT prior and the prohibitive cost of training at the large resolutions that RS applications demand. Training-free resolution promotion via Rotary Position Embedding (RoPE) rescaling offers a practical remedy, but every existing method applies a static positional scaling rule throughout the denoising process. This uniform compression is particularly harmful for RS imagery, whose substantially denser medium- and high-frequency energy encodes the fine structures critical for aerial-scene realism, such as vehicles, building contours, and road markings. Addressing both challenges requires a domain-specialized generative prior coupled with a denoising-aware positional adaptation strategy. To this end, we fine-tune FLUX on over 100,000 curated RS images to build a strong domain prior (RS-FLUX), and propose Spectrum-aware Highly-dynamic Adaptation for Resolution Promotion (SHARP), a training-free method that introduces a rational fractional time schedule k_rs(t) into RoPE. SHARP applies strong positional promotion during the early layout-formation stage and progressively relaxes it during detail recovery, aligning extrapolation strength with the frequency-progressive nature of diffusion denoising. Its resolution-agnostic formulation further enables robust multi-scale generation from a single set of hyperparameters. Extensive experiments across six square and rectangular resolutions show that SHARP consistently outperforms all training-free baselines on CLIP Score, Aesthetic Score, and HPSv2, with widening margins at more aggressive extrapolation factors and negligible computational overhead. Code and weights are available at https://github.com/bxuanz/SHARP.

Method: DEBIR consists of two main components: 1) Burst Auto-Exposure Network (BAENet) that estimates optimal exposure times for each burst image based on preview image, motion magnitude, and gain; 2) A burst image restoration network that reconstructs high-quality images from burst images captured with these optimal exposure times. Uses differentiable burst simulator and three-stage training strategy.

Result: The pipeline achieves state-of-the-art restoration quality and has been validated on a real-world camera system, demonstrating practical applicability.

Conclusion: Dynamic exposure prediction significantly improves burst image restoration quality compared to fixed exposure settings, and the approach is practical for real-world camera systems.

Abstract: Burst image restoration aims to reconstruct a high-quality image from burst images, which are typically captured using manually designed exposure settings. Although these exposure settings significantly influence the final restoration performance, the problem of finding optimal exposure settings has been overlooked. In this paper, we present Dynamic Exposure Burst Image Restoration (DEBIR), a novel burst image restoration pipeline that enhances restoration quality by dynamically predicting exposure times tailored to the shooting environment. In our pipeline, Burst Auto-Exposure Network (BAENet) estimates the optimal exposure time for each burst image based on a preview image, as well as motion magnitude and gain. Subsequently, a burst image restoration network reconstructs a high-quality image from burst images captured using these optimal exposure times. For training, we introduce a differentiable burst simulator and a three-stage training strategy. Our experiments demonstrate that our pipeline achieves state-of-the-art restoration quality. Furthermore, we validate the effectiveness of our approach on a real-world camera system, demonstrating its practicality.

Method: Image-conditioned reinforcement learning framework with lightweight texture-aware CNN encoder and privileged critic during training; formulates frontend configuration as sequential decision-making problem

Result: 3x longer feature tracks and 3x lower computational cost on TartanAirV2 and TUM RGB-D datasets, trained entirely in simulation

Conclusion: First RL framework for online VO parameter tuning that proactively adapts to visual input, improving robustness and efficiency

Abstract: Resource-constrained autonomous robots rely on sparse direct and semi-direct visual-(inertial)-odometry (VO) pipelines, as they provide a favorable tradeoff between accuracy, robustness, and computational cost. However, the performance of most systems depends critically on hand-tuned hyperparameters governing feature detection, tracking, and outlier rejection. These parameters are typically fixed during deployment, even though their optimal values vary with scene characteristics such as texture density, illumination, motion blur, and sensor noise, leading to brittle performance in real-world environments. We propose the first image-conditioned reinforcement learning framework for online tuning of VO frontend parameters, effectively embedding the expert into the system. Our key idea is to formulate the frontend configuration as a sequential decision-making problem and learn a policy that directly maps visual input to feature detection and tracking parameters. The policy uses a lightweight texture-aware CNN encoder and a privileged critic during training. Unlike prior RL-based approaches that rely solely on internal VO statistics, our method observes the image content and proactively adapts parameters before tracking degrades. Experiments on TartanAirV2 and TUM RGB-D show 3x longer feature tracks and 3x lower computational cost, despite training entirely in simulation.

Method: Benchmark recurrent ReYOLOv8s on MTEvent dataset for industrial multi-class recognition, using non-recurrent YOLOv8s variant as baseline to analyze temporal memory effects. Compare scratch training vs. event-domain pretraining (GEN1, PEDRo).

Result: Best scratch recurrent model achieves 0.285 mAP50 (9.6% improvement over non-recurrent baseline). GEN1-initialized fine-tuning yields best overall result of 0.329 mAP50. PEDRo initialization drops to 0.251, showing mismatched pretraining can be worse than scratch training. Models improve with clip length when pretrained.

Conclusion: Recurrent event-based detection shows promise for industrial environments, with event-domain pretraining being crucial. Class imbalance and human-object interaction remain challenges. This work provides focused benchmarking for industrial event-based detection.

Abstract: Event cameras are attractive for industrial robotics because they provide high temporal resolution, high dynamic range, and reduced motion blur. However, most event-based object detection studies focus on outdoor driving scenarios or limited class settings. In this work, we benchmark recurrent ReYOLOv8s on MTEvent for industrial multi-class recognition and use a non-recurrent YOLOv8s variant as a baseline to analyze the effect of temporal memory. On the MTEvent validation split, the best scratch recurrent model (C21) reaches 0.285 mAP50, corresponding to a 9.6% relative improvement over the nonrecurrent YOLOv8s baseline (0.260). Event-domain pretraining has a stronger effect: GEN1-initialized fine-tuning yields the best overall result of 0.329 mAP50 at clip length 21, and unlike scratch training, GEN1-pretrained models improve consistently with clip length. PEDRo initialization drops to 0.251, indicating that mismatched source-domain pretraining can be less effective than training from scratch. Persistent failure modes are dominated by class imbalance and human-object interaction. Overall, we position this work as a focused benchmarking and analysis study of recurrent event-based detection in industrial environments.

Method: Combines BERTopic for thematic segmentation, Whisper-AT for laughter detection, YOLOv8-cls for shot classification, and YOLOv8s-pose for skeletal keypoint extraction, with all streams temporally aligned without resampling.

Result: Created dataset with 5,400+ topic segments; found kinetic energy negatively predicts laughter, personal content elicits more laughter than geopolitical themes, and close-up shots correlate positively with laughter.

Conclusion: TIC-TALK enables multimodal analysis of performance dynamics in stand-up comedy, revealing relationships between body language, thematic content, and audience response.

Abstract: Stand-up comedy, and humor in general, are often studied through their verbal content. Yet live performance relies just as much on embodied presence and audience feedback. We introduce TIC-TALK, a multimodal resource with 5,400+ temporally aligned topic segments capturing language, gesture, and audience response across 90 professionally filmed stand-up comedy specials (2015-2024). The pipeline combines BERTopic for 60 s thematic segmentation with dense sentence embeddings, Whisper-AT for 0.8 s laughter detection, a fine-tuned YOLOv8-cls shot classifier, and YOLOv8s-pose for raw keypoint extraction at 1 fps. Raw 17-joint skeletal coordinates are retained without prior clustering, enabling the computation of continuous kinematic signals-arm spread, kinetic energy, and trunk lean-that serve as proxies for performance dynamics. All streams are aligned by hierarchical temporal containment without resampling, and each topic segment stores its sentence-BERT embedding for downstream similarity and clustering tasks. As a concrete use case, we study laughter dynamics across 24 thematic topics: kinetic energy negatively predicts audience laughter rate (r = -0.75, N = 24), consistent with a stillness-before-punchline pattern; personal and bodily content elicits more laughter than geopolitical themes; and shot close-up proportion correlates positively with laughter (r = +0.28), consistent with reactive montage.

Method: Leverages pretrained medical image tokenizer and latent transformer to represent anatomy through quantized visual tokens. Uses token entropy as uncertainty measure to guide active sampling via two strategies: Latent Entropy Selection (LES) projects patch-wise token entropy to k-space to identify informative sampling lines, and Gradient-based Entropy Optimization (GEO) finds regions of maximum uncertainty reduction via k-space gradient of total latent entropy loss.

Result: Evaluated on fastMRI single-coil Knee and Brain datasets at 8× and 16× acceleration. Outperforms state-of-the-art baselines in perceptual metrics and feature-based distances.

Conclusion: The proposed active sampling framework effectively accelerates MRI acquisition by intelligently selecting k-space sampling locations based on latent uncertainty measures from tokenized representations.

Abstract: Full data acquisition in MRI is inherently slow, which limits clinical throughput and increases patient discomfort. Compressed Sensing MRI (CS-MRI) seeks to accelerate acquisition by reconstructing images from under-sampled k-space data, requiring both an optimal sampling trajectory and a high-fidelity reconstruction model. In this work, we propose a novel active sampling framework that leverages the inherent discrete structure of a pretrained medical image tokenizer and a latent transformer. By representing anatomy through a dictionary of quantized visual tokens, the model provides a well-defined probability distribution over the latent space. We utilize this distribution to derive a principled uncertainty measure via token entropy, which guides the active sampling process. We introduce two strategies to exploit this latent uncertainty: (1) Latent Entropy Selection (LES), projecting patch-wise token entropy into the $k$-space domain to identify informative sampling lines, and (2) Gradient-based Entropy Optimization (GEO), which identifies regions of maximum uncertainty reduction via the $k$-space gradient of a total latent entropy loss. We evaluate our framework on the fastMRI singlecoil Knee and Brain datasets at $\times 8$ and $\times 16$ acceleration. Our results demonstrate that our active policies outperform state-of-the-art baselines in perceptual metrics, and feature-based distances. Our code is available at https://github.com/levayz/TRUST-MRI.

Method: Proposes cascade-free architecture based on multitask learning that jointly integrates multiple intermediate representations (phoneme and viseme). Uses semantic-guided local contrastive loss to temporally align features, enabling on-demand activation during inference for efficiency-performance trade-off.

Result: Experiments on publicly available datasets demonstrate superior recognition performance compared to existing methods.

Conclusion: The proposed method effectively addresses tonal language challenges in Chinese Mandarin VSR by eliminating cascade dependencies, reducing error accumulation, and improving inference efficiency while maintaining performance.

Abstract: Chinese mandarin visual speech recognition (VSR) is a task that has advanced in recent years, yet still lags behind the performance on non-tonal languages such as English. One primary challenge arises from the tonal nature of Mandarin, which limits the effectiveness of conventional sequence-to-sequence modeling approaches. To alleviate this issue, existing Chinese VSR systems commonly incorporate intermediate representations, most notably pinyin, within cascade architectures to enhance recognition accuracy. While beneficial, in these cascaded designs, the subsequent stage during inference depends on the output of the preceding stage, leading to error accumulation and increased inference latency. To address these limitations, we propose a cascade-free architecture based on multitask learning that jointly integrates multiple intermediate representations, including phoneme and viseme, to better exploit contextual information. The proposed semantic-guided local contrastive loss temporally aligns the features, enabling on-demand activation during inference, thereby providing a trade-off between inference efficiency and performance while mitigating error accumulation caused by projection and re-embedding. Experiments conducted on publicly available datasets demonstrate that our method achieves superior recognition performance.

Method: Clinical Graph-Mediated Distillation (CGMD) constructs a clinical similarity kNN graph using shared biomarkers to bridge MRI and fundus cohorts, trains an MRI teacher, propagates representations over the graph, imputes brain-informed targets for fundus patients, and trains a fundus student with joint HTN supervision, target distillation, and relational distillation.

Result: CGMD consistently improves fundus-based hypertension prediction over standard distillation and non-graph imputation baselines on newly collected unpaired MRI-fundus-biomarker dataset, with ablations confirming importance of clinically grounded graph connectivity.

Conclusion: CGMD effectively transfers knowledge across unpaired medical imaging modalities using clinical similarity graphs, enabling improved hypertension screening from low-cost fundus images by leveraging brain MRI insights without requiring paired data.

Abstract: Retinal fundus imaging enables low-cost and scalable hypertension (HTN) screening, but HTN-related retinal cues are subtle, yielding high-variance predictions. Brain MRI provides stronger vascular and small-vessel-disease markers of HTN, yet it is expensive and rarely acquired alongside fundus images, resulting in modality-siloed datasets with disjoint MRI and fundus cohorts. We study this unpaired MRI-fundus regime and introduce Clinical Graph-Mediated Distillation (CGMD), a framework that transfers MRI-derived HTN knowledge to a fundus model without paired multimodal data. CGMD leverages shared structured biomarkers as a bridge by constructing a clinical similarity kNN graph spanning both cohorts. We train an MRI teacher, propagate its representations over the graph, and impute brain-informed representation targets for fundus patients. A fundus student is then trained with a joint objective combining HTN supervision, target distillation, and relational distillation. Experiments on our newly collected unpaired MRI-fundus-biomarker dataset show that CGMD consistently improves fundus-based HTN prediction over standard distillation and non-graph imputation baselines, with ablations confirming the importance of clinically grounded graph connectivity. Code is available at https://github.com/DillanImans/CGMD-unpaired-distillation.

Method: Ctrl-A uses a control-loop architecture with relative operation response curves to dynamically and individually adapt augmentation strength distributions during training. It employs an operation-dependent update procedure that can suppress augmentation styles that negatively impact model performance.

Result: Experiments on CIFAR-10, CIFAR-100, and SVHN-core datasets using WideResNet-28-10 show that Ctrl-A is highly competitive with existing state-of-the-art data augmentation strategies.

Conclusion: Ctrl-A provides an effective automated approach to data augmentation for image-vision tasks by leveraging control theory principles, eliminating the need for manual initialization and policy engineering while achieving competitive performance.

Abstract: We introduce ControlAugment (Ctrl-A), an automated data augmentation algorithm for image-vision tasks, which incorporates principles from control theory for online adjustment of augmentation strength distributions during model training. Ctrl-A eliminates the need for initialization of individual augmentation strengths. Instead, augmentation strength distributions are dynamically, and individually, adapted during training based on a control-loop architecture and what we define as relative operation response curves. Using an operation-dependent update procedure provides Ctrl-A with the potential to suppress augmentation styles that negatively impact model performance, alleviating the need for manually engineering augmentation policies for new image-vision tasks. Experiments on the CIFAR-10, CIFAR-100, and SVHN-core benchmark datasets using the common WideResNet-28-10 architecture demonstrate that Ctrl-A is highly competitive with existing state-of-the-art data augmentation strategies.

Method: Proposes UnPaired and Arbitrarily Paired Training Paradigms (UPTP and APTP) with theoretical objectives. Develops a practical framework with three lightweight baselines (CNN, Transformer, GAN) and innovative loss functions to handle limited, unaligned training data.

Result: APTP and UPTP achieve performance comparable to strictly paired training with 100x larger datasets, demonstrating feasibility with severely limited and content-inconsistent data.

Conclusion: The proposed paradigms fundamentally reduce data collection costs and difficulty while enhancing model robustness, providing a practical solution for infrared-visible image fusion studies.

Abstract: Infrared and visible image fusion(IVIF) combines complementary modalities while preserving natural textures and salient thermal signatures. Existing solutions predominantly rely on extensive sets of rigidly aligned image pairs for training. However, acquiring such data is often impractical due to the costly and labour-intensive alignment process. Besides, maintaining a rigid pairing setting during training restricts the volume of cross-modal relationships, thereby limiting generalisation performance. To this end, this work challenges the necessity of Strictly Paired Training Paradigm (SPTP) by systematically investigating UnPaired and Arbitrarily Paired Training Paradigms (UPTP and APTP) for high-performance IVIF. We establish a theoretical objective of APTP, reflecting the complementary nature between UPTP and SPTP. More importantly, we develop a practical framework capable of significantly enriching cross-modal relationships even with severely limited and unaligned training data. To validate our propositions, three end-to-end lightweight baselines, alongside a set of innovative loss functions, are designed to cover three classic frameworks (CNN, Transformer, GAN). Comprehensive experiments demonstrate that the proposed APTP and UPTP are feasible and capable of training models on a severely limited and content-inconsistent infrared and visible dataset, achieving performance comparable to that of a dataset 100$\times$ larger in SPTP. This finding fundamentally alleviates the cost and difficulty of data collection while enhancing model robustness from the data perspective, delivering a feasible solution for IVIF studies. The code is available at \href{https://github.com/yanglinDeng/IVIF_unpair}{\textcolor{blue}{https://github.com/yanglinDeng/IVIF_unpair}}.

Method: Created SteelDefectX dataset with 7,778 images across 25 defect categories annotated with coarse-to-fine textual descriptions. Coarse level includes defect categories, visual attributes, and industrial causes; fine level includes sample-specific attributes like shape, size, depth, position, contrast. Established benchmark with four tasks: vision-only classification, vision-language classification, few/zero-shot recognition, and zero-shot transfer.

Result: Experiments with baseline models show coarse-to-fine textual annotations significantly improve interpretability, generalization, and transferability. Dataset enables models to learn richer defect representations.

Conclusion: SteelDefectX serves as valuable resource for advancing explainable, generalizable steel surface defect detection. Dataset publicly available on GitHub.

Abstract: Steel surface defect detection is essential for ensuring product quality and reliability in modern manufacturing. Current methods often rely on basic image classification models trained on label-only datasets, which limits their interpretability and generalization. To address these challenges, we introduce SteelDefectX, a vision-language dataset containing 7,778 images across 25 defect categories, annotated with coarse-to-fine textual descriptions. At the coarse-grained level, the dataset provides class-level information, including defect categories, representative visual attributes, and associated industrial causes. At the fine-grained level, it captures sample-specific attributes, such as shape, size, depth, position, and contrast, enabling models to learn richer and more detailed defect representations. We further establish a benchmark comprising four tasks, vision-only classification, vision-language classification, few/zero-shot recognition, and zero-shot transfer, to evaluate model performance and generalization. Experiments with several baseline models demonstrate that coarse-to-fine textual annotations significantly improve interpretability, generalization, and transferability. We hope that SteelDefectX will serve as a valuable resource for advancing research on explainable, generalizable steel surface defect detection. The data will be publicly available on https://github.com/Zhaosxian/SteelDefectX.

Method: Two-stage framework with MDSConv (multidirectional snake convolution) in encoding stage for multi-view feature fusion along anatomical planes, and RVM (residual visual Mamba) in decoding stage for efficient long-range dependency modeling with linear complexity. Uses progressive segmentation: coarse whole-image segmentation followed by fine-grained block-level segmentation.

Result: The paper proposes a novel segmentation framework but does not provide specific quantitative results in the abstract. The method addresses computational efficiency challenges while aiming to capture complex vascular geometry.

Conclusion: MDSVM-UNet offers an efficient solution for coronary artery segmentation by combining MDSConv’s geometric modeling with RVM’s efficient long-range dependency capture, potentially enabling clinical deployment in resource-constrained settings.

Abstract: Accurate segmentation of coronary arteries from computed tomography angiography (CTA) images is of paramount clinical importance for the diagnosis and treatment planning of cardiovascular diseases. However, coronary artery segmentation remains challenging due to the inherent multi-branching and slender tubular morphology of the vasculature, compounded by severe class imbalance between foreground vessels and background tissue. Conventional convolutional neural network (CNN)-based approaches struggle to capture long-range dependencies among spatially distant vascular structures, while Vision Transformer (ViT)-based methods incur prohibitive computational overhead that hinders deployment in resource-constrained clinical settings. Motivated by the recent success of state space models (SSMs) in efficiently modeling long-range sequential dependencies with linear complexity, we propose MDSVM-UNet, a novel two-stage coronary artery segmentation framework that synergistically integrates multidirectional snake convolution (MDSConv) with residual visual Mamba (RVM). In the encoding stage, we introduce MDSConv, a deformable convolution module that learns adaptive offsets along three orthogonal anatomical planes – sagittal, coronal, and axial – thereby enabling comprehensive multi-view feature fusion that faithfully captures the elongated and tortuous geometry of coronary vessels. In the decoding stage, we design an RVM-based upsampling decoder block that leverages selective state space mechanisms to model inter-slice long-range dependencies while preserving linear computational complexity. Furthermore, we propose a progressive two-stage segmentation strategy: the first stage performs coarse whole-image segmentation to guide intelligent block extraction, while the second stage conducts fine-grained block-level segmentation to recover vascular details and suppress false positives..

Method: A video diffusion model trained on atmospheric reanalysis data to generate MJO sequences conditioned on key low-dimensional metrics, with intentional idealized conditionings to create tractable scenarios for analysis.

Result: The generated MJOs capture key features including composites, power spectra, and multiscale structures like convectively coupled waves, despite some biases. The model successfully generates tractable MJOs under idealized conditions.

Conclusion: The approach provides a practical framework for connecting low-dimensional MJO theory with high-resolution atmospheric complexity and will aid tropical atmosphere prediction.

Abstract: Generative Deep Learning is a powerful tool for modeling of the Madden-Julian oscillation (MJO) in the tropics, yet its relationship to traditional theoretical frameworks remains poorly understood. Here we propose a video diffusion model, trained on atmospheric reanalysis, to synthetize long MJO sequences conditioned on key low-dimensional metrics. The generated MJOs capture key features including composites, power spectra and multiscale structures including convectively coupled waves, despite some bias. We then prompt the model to generate more tractable MJOs based on intentionally idealized low-dimensional conditionings, for example a perpetual MJO, an isolated modulation by seasons and/or the El Nino-Southern Oscillation, and so on. This enables deconstructing the underlying processes and identifying physical drivers. The present approach provides a practical framework for bridging the gap between low-dimensional MJO theory and high-resolution atmospheric complexity and will help tropical atmosphere prediction.

Method: Proposes a video-specific distillation framework with three key innovations: (1) adaptive regression loss that dynamically adjusts spatial supervision weights to prevent artifacts from distribution shifts, (2) temporal regularization loss to prevent temporal collapse and promote smooth sampling trajectories, and (3) inference-time frame interpolation to reduce sampling overhead while preserving quality.

Result: Extensive experiments on VBench and VBench2 benchmarks show the method achieves stable few-step video synthesis, significantly enhancing perceptual fidelity and motion realism, consistently outperforming existing distillation baselines across multiple metrics.

Conclusion: The proposed distillation framework effectively addresses video-specific artifacts from image distillation techniques, enabling efficient and high-quality video generation through adaptive spatial supervision, temporal regularization, and optimized inference strategies.

Abstract: Video generation has recently emerged as a central task in the field of generative AI. However, the substantial computational cost inherent in video synthesis makes model distillation a critical technique for efficient deployment. Despite its significance, there is a scarcity of methods specifically designed for video diffusion models. Prevailing approaches often directly adapt image distillation techniques, which frequently lead to artifacts such as oversaturation, temporal inconsistency, and mode collapse. To address these challenges, we propose a novel distillation framework tailored specifically for video diffusion models. Its core innovations include: (1) an adaptive regression loss that dynamically adjusts spatial supervision weights to prevent artifacts arising from excessive distribution shifts; (2) a temporal regularization loss to counteract temporal collapse, promoting smooth and physically plausible sampling trajectories; and (3) an inference-time frame interpolation strategy that reduces sampling overhead while preserving perceptual quality. Extensive experiments and ablation studies on the VBench and VBench2 benchmarks demonstrate that our method achieves stable few-step video synthesis, significantly enhancing perceptual fidelity and motion realism. It consistently outperforms existing distillation baselines across multiple metrics.

Method: Define low-dimensional pattern space (color, shape, angle), optimize patterns to maximize recognition error, project onto semantically valid facial regions, ensure cross-model transferability across architectures.

Result: Significantly degrades performance of all tested state-of-the-art face recognition models in simulations, shows promising real-world results, reveals model robustness differences and attack transferability across architectures.

Conclusion: Adversarial Camouflage provides an effective privacy protection solution against facial recognition systems with demonstrated cross-model effectiveness and real-world applicability.

Abstract: While the rapid development of facial recognition algorithms has enabled numerous beneficial applications, their widespread deployment has raised significant concerns about the risks of mass surveillance and threats to individual privacy. In this paper, we introduce \textit{Adversarial Camouflage} as a novel solution for protecting users’ privacy. This approach is designed to be efficient and simple to reproduce for users in the physical world. The algorithm starts by defining a low-dimensional pattern space parameterized by color, shape, and angle. Optimized patterns, once found, are projected onto semantically valid facial regions for evaluation. Our method maximizes recognition error across multiple architectures, ensuring high cross-model transferability even against black-box systems. It significantly degrades the performance of all tested state-of-the-art face recognition models during simulations and demonstrates promising results in real-world human experiments, while revealing differences in model robustness and evidence of attack transferability across architectures.

Method: SAGE-GRPO applies constraints at micro and macro levels: micro-level includes manifold-aware SDE with logarithmic curvature correction and gradient norm equalizer; macro-level uses dual trust region with periodic moving anchor and stepwise constraints to track checkpoints closer to the manifold and limit drift.

Result: Evaluation on HunyuanVideo1.5 with VideoAlign reward model shows consistent gains over previous methods in VQ, MQ, TA, and visual metrics (CLIPScore, PickScore), demonstrating superior performance in both reward maximization and overall video quality.

Conclusion: Constraining exploration within the pre-trained model’s data manifold through micro and macro-level constraints enables stable alignment for video generation, addressing reliability issues in current GRPO methods.

Abstract: Group Relative Policy Optimization (GRPO) methods for video generation like FlowGRPO remain far less reliable than their counterparts for language models and images. This gap arises because video generation has a complex solution space, and the ODE-to-SDE conversion used for exploration can inject excess noise, lowering rollout quality and making reward estimates less reliable, which destabilizes post-training alignment. To address this problem, we view the pre-trained model as defining a valid video data manifold and formulate the core problem as constraining exploration within the vicinity of this manifold, ensuring that rollout quality is preserved and reward estimates remain reliable. We propose SAGE-GRPO (Stable Alignment via Exploration), which applies constraints at both micro and macro levels. At the micro level, we derive a precise manifold-aware SDE with a logarithmic curvature correction and introduce a gradient norm equalizer to stabilize sampling and updates across timesteps. At the macro level, we use a dual trust region with a periodic moving anchor and stepwise constraints so that the trust region tracks checkpoints that are closer to the manifold and limits long-horizon drift. We evaluate SAGE-GRPO on HunyuanVideo1.5 using the original VideoAlign as the reward model and observe consistent gains over previous methods in VQ, MQ, TA, and visual metrics (CLIPScore, PickScore), demonstrating superior performance in both reward maximization and overall video quality. The code and visual gallery are available at https://dungeonmassster.github.io/SAGE-GRPO-Page/.

Method: UPPA uses geometrically constrained parameterized Bezier blocks to model perturbations, employs Particle Swarm Optimization (PSO) for unified optimization across global data distribution, and materializes digital perturbations into physical cold patches that create continuous low-temperature distributions aligned with infrared thermal radiation characteristics.

Result: Extensive experiments show UPPA achieves outstanding physical attack success rate without online computational overhead, exhibits strong cross-domain generalization, and has reliable black-box transferability.

Conclusion: UPPA is the first universal physical attack method in the infrared domain that addresses limitations of existing methods by providing a computationally efficient, physically robust solution with good generalization capabilities.

Abstract: Although infrared pedestrian detectors have been widely deployed in visual perception tasks, their vulnerability to physical adversarial attacks is becoming increasingly apparent. Existing physical attack methods predominantly rely on instance-specific online optimization and rigid pattern design, leading to high deployment costs and insufficient physical robustness. To address these limitations, this work proposes the Universal Physical Patch Attack (UPPA), the first universal physical attack method in the infrared domain. This method employs geometrically constrained parameterized Bezier blocks to model perturbations and utilizes the Particle Swarm Optimization (PSO) algorithm to perform unified optimization across the global data distribution, thus maintaining topological stability under dynamic deformations. In the physical deployment phase, we materialize the optimized digital perturbations into physical cold patches, achieving a continuous and smooth low-temperature distribution that naturally aligns with the thermal radiation characteristics of infrared imaging. Extensive experiments demonstrate that UPPA achieves an outstanding physical attack success rate without any online computational overhead, while also exhibiting strong cross-domain generalization and reliable black-box transferability.

Method: Proposes LoRA² which learns adaptive ranks for each layer during fine-tuning by imposing importance ordering on rank positions, encouraging higher ranks only when needed. Uses variational methods to let ranks freely adapt based on subject complexity.

Result: Achieves competitive trade-off between DINO, CLIP-I, and CLIP-T metrics across 29 subjects while requiring less memory and lower overall rank than high-rank LoRA versions. Shows both qualitative and quantitative improvements.

Conclusion: LoRA² provides an effective solution to the rank selection problem in diffusion model fine-tuning, enabling adaptive layer-wise rank allocation that balances performance and memory efficiency better than fixed-rank approaches.

Abstract: Low Rank Adaptation (LoRA) is the de facto fine-tuning strategy to generate personalized images from pre-trained diffusion models. Choosing a good rank is extremely critical, since it trades off performance and memory consumption, but today the decision is often left to the community’s consensus, regardless of the personalized subject’s complexity. The reason is evident: the cost of selecting a good rank for each LoRA component is combinatorial, so we opt for practical shortcuts such as fixing the same rank for all components. In this paper, we take a first step to overcome this challenge. Inspired by variational methods that learn an adaptive width of neural networks, we let the ranks of each layer freely adapt during fine-tuning on a subject. We achieve it by imposing an ordering of importance on the rank’s positions, effectively encouraging the creation of higher ranks when strictly needed. Qualitatively and quantitatively, our approach, LoRA$^2$, achieves a competitive trade-off between DINO, CLIP-I, and CLIP-T across 29 subjects while requiring much less memory and lower rank than high rank LoRA versions. Code: https://github.com/donaldssh/NotAllLayersAreCreatedEqual.

Method: Integrated several modern learning-based stereo matchers (StereoAnywhere, MonSter, Foundation Stereo, and satellite fine-tuned MonSter) into the Satellite Stereo Pipeline (S2P), adapting the rectification stage to enforce compatible disparity polarity and range.

Result: Experiments show consistent improvements over classical cost-volume-based approaches in Digital Surface Model accuracy, with substantially improved geometric detail and sharper structures, though metrics like mean absolute error show saturation effects. Performance on challenging surfaces like vegetation remains limited.

Conclusion: Learning-based stereo matchers can improve satellite stereo pipelines but require adaptations for satellite geometry, and evaluation strategies need to better reflect perceptual and structural fidelity rather than just traditional metrics.

Abstract: Digital Surface Model generation from satellite imagery is a core task in Earth observation and is commonly addressed using classical stereoscopic matching algorithms in satellite pipelines as in the Satellite Stereo Pipeline (S2P). While recent learning-based stereo matchers achieve state-of-the-art performance on standard benchmarks, their integration into operational satellite pipelines remains challenging due to differences in viewing geometry and disparity assumptions. In this work, we integrate several modern learning-based stereo matchers, including StereoAnywhere, MonSter, Foundation Stereo, and a satellite fine-tuned variant of MonSter, into the Satellite Stereo Pipeline, adapting the rectification stage to enforce compatible disparity polarity and range. We release the corresponding code to enable reproducible use of these methods in large-scale Earth observation workflows. Experiments on satellite imagery show consistent improvements over classical cost-volume-based approaches in terms of Digital Surface Model accuracy, although commonly used metrics such as mean absolute error exhibit saturation effects. Qualitative results reveal substantially improved geometric detail and sharper structures, highlighting the need for evaluation strategies that better reflect perceptual and structural fidelity. At the same time, performance over challenging surface types such as vegetation remains limited across all evaluated models, indicating open challenges for learning-based stereo in natural environments.

Method: Built on DINOv3 foundation with Hierarchical Feature Modulation (HFM) for high-fidelity class-aware features, Hypergraph Plausibility Estimation (HPE) for global anatomical plausibility assessment, and Structural Anomaly Pruning (SAP) for artifact removal via cross-view stability.

Result: State-of-the-art performance on cardiac and abdominal cross-modality benchmarks: 90.08% (MRI->CT) and 78.51% (CT->MRI) Dice scores on cardiac data; 87.48% (MRI->CT) and 86.89% (CT->MRI) on abdominal data.

Conclusion: SHAPE successfully addresses limitations of existing UDA methods by reframing adaptation towards global anatomical plausibility, achieving superior performance through structure-aware hierarchical adaptation with plausibility evaluation.

Abstract: Unsupervised Domain Adaptation (UDA) is essential for deploying medical segmentation models across diverse clinical environments. Existing methods are fundamentally limited, suffering from semantically unaware feature alignment that results in poor distributional fidelity and from pseudo-label validation that disregards global anatomical constraints, thus failing to prevent the formation of globally implausible structures. To address these issues, we propose SHAPE (Structure-aware Hierarchical Unsupervised Domain Adaptation with Plausibility Evaluation), a framework that reframes adaptation towards global anatomical plausibility. Built on a DINOv3 foundation, its Hierarchical Feature Modulation (HFM) module first generates features with both high fidelity and class-awareness. This shifts the core challenge to robustly validating pseudo-labels. To augment conventional pixel-level validation, we introduce Hypergraph Plausibility Estimation (HPE), which leverages hypergraphs to assess the global anatomical plausibility that standard graphs cannot capture. This is complemented by Structural Anomaly Pruning (SAP) to purge remaining artifacts via cross-view stability. SHAPE significantly outperforms prior methods on cardiac and abdominal cross-modality benchmarks, achieving state-of-the-art average Dice scores of 90.08% (MRI->CT) and 78.51% (CT->MRI) on cardiac data, and 87.48% (MRI->CT) and 86.89% (CT->MRI) on abdominal data. The code is available at https://github.com/BioMedIA-repo/SHAPE.

Method: Two-stage design: Stage I learns disentangled subtitle representations via self-supervised orthogonality constraints on dual encoders. Stage II employs LoRA-based adaptation with generation feedback for dynamic context adjustment. Only requires 0.77% of base diffusion model parameters for training.

Result: Outperforms mask-dependent baselines by +6.77dB PSNR and -74.7% VFID on Chinese subtitle benchmarks. Demonstrates superior zero-shot generalization across six languages (English, Korean, French, Japanese, Russian, German).

Conclusion: CLEAR achieves truly end-to-end inference without mask requirements, enabling practical deployment with strong performance and cross-language generalization through generation-driven feedback.

Abstract: Video subtitle removal aims to distinguish text overlays from background content while preserving temporal coherence. Existing diffusion-based methods necessitate explicit mask sequences during both training and inference phases, which restricts their practical deployment. In this paper, we present CLEAR (Context-aware Learning for End-to-end Adaptive Video Subtitle Removal), a mask-free framework that achieves truly end-to-end inference through context-aware adaptive learning. Our two-stage design decouples prior extraction from generative refinement: Stage I learns disentangled subtitle representations via self-supervised orthogonality constraints on dual encoders, while Stage II employs LoRA-based adaptation with generation feedback for dynamic context adjustment. Notably, our method only requires 0.77% of the parameters of the base diffusion model for training. On Chinese subtitle benchmarks, CLEAR outperforms mask-dependent baselines by + 6.77dB PSNR and -74.7% VFID, while demonstrating superior zero-shot generalization across six languages (English, Korean, French, Japanese, Russian, German), a performance enabled by our generation-driven feedback mechanism that ensures robust subtitle removal without ground-truth masks during inference.

Method: Proposes a camera-agnostic, one-shot, post-training pruning method using hybrid descriptor framework capturing structural and appearance consistency directly from splat representation. Formulates pruning as statistical evidence estimation problem with Beta evidence model that quantifies per-splat reliability through probabilistic confidence scores.

Result: Experiments on ISO/IEC MPEG CTC test sequences show substantial pruning while preserving reconstruction quality, establishing a practical and generalizable alternative to camera-dependent pruning strategies.

Conclusion: The method provides an effective camera-agnostic pruning solution for 3D Gaussian splats that works in emerging exchange settings where splats are shared directly as point-based representations.

Abstract: The pruning of 3D Gaussian splats is essential for reducing their complexity to enable efficient storage, transmission, and downstream processing. However, most of the existing pruning strategies depend on camera parameters, rendered images, or view-dependent measures. This dependency becomes a hindrance in emerging camera-agnostic exchange settings, where splats are shared directly as point-based representations (e.g., .ply). In this paper, we propose a camera-agnostic, one-shot, post-training pruning method for 3D Gaussian splats that relies solely on attribute-derived neighbourhood descriptors. As our primary contribution, we introduce a hybrid descriptor framework that captures structural and appearance consistency directly from the splat representation. Building on these descriptors, we formulate pruning as a statistical evidence estimation problem and introduce a Beta evidence model that quantifies per-splat reliability through a probabilistic confidence score. Experiments conducted on standardized test sequences defined by the ISO/IEC MPEG Common Test Conditions (CTC) demonstrate that our approach achieves substantial pruning while preserving reconstruction quality, establishing a practical and generalizable alternative to existing camera-dependent pruning strategies.

Method: The paper proves the existence of a “golden subspace” for adaptation that coincides with the row space of the pretrained classifier. They introduce sample-wise Average Gradient Outer Product (AGOP) to efficiently estimate classifier weights without retraining. GOLD uses a lightweight adapter to project features onto this subspace and learns a compact scaling vector while dynamically updating the subspace via AGOP.

Result: Extensive experiments on classification and segmentation benchmarks, including autonomous-driving scenarios, demonstrate that GOLD attains superior efficiency, stability, and overall performance compared to existing CTTA methods.

Conclusion: GOLD successfully addresses the efficiency-generalization trade-off in CTTA by identifying and maintaining the golden subspace for adaptation, enabling efficient online adaptation with minimal feature updates while maintaining strong performance.

Abstract: Continual Test-Time Adaptation (CTTA) aims to enable models to adapt online to unlabeled data streams under distribution shift without accessing source data. Existing CTTA methods face an efficiency-generalization trade-off: updating more parameters improves adaptation but severely reduces online inference efficiency. An ideal solution is to achieve comparable adaptation with minimal feature updates; we call this minimal subspace the golden subspace. We prove its existence in a single-step adaptation setting and show that it coincides with the row space of the pretrained classifier. To enable online maintenance of this subspace, we introduce the sample-wise Average Gradient Outer Product (AGOP) as an efficient proxy for estimating the classifier weights without retraining. Building on these insights, we propose Guided Online Low-rank Directional adaptation (GOLD), which uses a lightweight adapter to project features onto the golden subspace and learns a compact scaling vector while the subspace is dynamically updated via AGOP. Extensive experiments on classification and segmentation benchmarks, including autonomous-driving scenarios, demonstrate that GOLD attains superior efficiency, stability, and overall performance. Our code is available at https://github.com/AIGNLAI/GOLD.

Method: ChronoCon uses contrastive learning with rankings derived from visitation order of longitudinal scans, replacing label-based ranking losses with temporal ordering. It assumes monotonic progression in irreversible diseases to learn disease-relevant representations without expert labels, generalizing Rank-N-Contrast from label distances to temporal ordering.

Result: On rheumatoid arthritis radiographs, ChronoCon significantly improves label efficiency, outperforming fully supervised baselines in low-label settings. Fine-tuning on expert scores from only five patients yields 86% intraclass correlation coefficient for severity prediction.

Conclusion: Chronological contrastive learning can exploit routinely available imaging metadata to reduce annotation requirements in irreversible disease domains, demonstrating practical value for medical imaging applications.

Abstract: Quantitative disease severity scoring in medical imaging is costly, time-consuming, and subject to inter-reader variability. At the same time, clinical archives contain far more longitudinal imaging data than expert-annotated severity scores. Existing self-supervised methods typically ignore this chronological structure. We introduce ChronoCon, a contrastive learning approach that replaces label-based ranking losses with rankings derived solely from the visitation order of a patient’s longitudinal scans. Under the clinically plausible assumption of monotonic progression in irreversible diseases, the method learns disease-relevant representations without using any expert labels. This generalizes the idea of Rank-N-Contrast from label distances to temporal ordering. Evaluated on rheumatoid arthritis radiographs for severity assessment, the learned representations substantially improve label efficiency. In low-label settings, ChronoCon significantly outperforms a fully supervised baseline initialized from ImageNet weights. In a few-shot learning experiment, fine-tuning ChronoCon on expert scores from only five patients yields an intraclass correlation coefficient of 86% for severity score prediction. These results demonstrate the potential of chronological contrastive learning to exploit routinely available imaging metadata to reduce annotation requirements in the irreversible disease domain. Code is available at https://github.com/cirmuw/ChronoCon.

Method: Three model-agnostic regularizers: 1) Gravity-Aligned Planarity Regularization aligns surface normals with gravity axis, 2) Granularity Regularization enforces coarse-to-fine geometry learning, 3) Depth-Supervised Regularization stabilizes early training.

Result: On DFC2019 benchmark, SatGeo-NeRF improves Mean Altitude Error by 13.9% and 11.7% relative to state-of-the-art baselines EO-NeRF and EO-GS.

Conclusion: The proposed geometric regularizers effectively mitigate overfitting artifacts in satellite NeRF reconstruction, significantly improving geometric accuracy.

Abstract: We present SatGeo-NeRF, a geometrically regularized NeRF for satellite imagery that mitigates overfitting-induced geometric artifacts observed in current state-of-the-art models using three model-agnostic regularizers. Gravity-Aligned Planarity Regularization aligns depth-inferred, approximated surface normals with the gravity axis to promote local planarity, coupling adjacent rays via a corresponding surface approximation to facilitate cross-ray gradient flow. Granularity Regularization enforces a coarse-to-fine geometry-learning scheme, and Depth-Supervised Regularization stabilizes early training for improved geometric accuracy. On the DFC2019 satellite reconstruction benchmark, SatGeo-NeRF improves the Mean Altitude Error by 13.9% and 11.7% relative to state-of-the-art baselines such as EO-NeRF and EO-GS.

Method: Two-step optimization framework: 1) Iterative feature-guided absolute orientation solver for coarse registration (robust to poor initialization), 2) Fine registration enforcing multi-view feature consistency inspired by inverse radiance-field formulations. Uses viewpoint-guided spherical map features for robust correspondences.

Result: GSA outperforms prior works in same-object cases, often by large margins, even when competitors are given true scale. For different objects in same category, GSA vastly surpasses existing methods, providing first effective solution for category-level 3DGS registration.

Conclusion: GSA enables effective category-level 3D Gaussian Splatting registration, unlocking new applications by aligning different objects within the same category without requiring scale priors or same-object constraints.

Abstract: We present Gaussian Splatting Alignment (GSA), a novel method for aligning two independent 3D Gaussian Splatting (3DGS) models via a similarity transformation (rotation, translation, and scale), even when they are of different objects in the same category (e.g., different cars). In contrast, existing methods can only align 3DGS models of the same object (e.g., the same car) and often must be given true scale as input, while we estimate it successfully. GSA leverages viewpoint-guided spherical map features to obtain robust correspondences and introduces a two-step optimization framework that aligns 3DGS models while keeping them fixed. First, we apply an iterative feature-guided absolute orientation solver as our coarse registration, which is robust to poor initialization (e.g., 180 degrees misalignment or a 10x scale gap). Next, we use a fine registration step that enforces multi-view feature consistency, inspired by inverse radiance-field formulations. The first step already achieves state-of-the-art performance, and the second further improves results. In the same-object case, GSA outperforms prior works, often by a large margin, even when the other methods are given the true scale. In the harder case of different objects in the same category, GSA vastly surpasses them, providing the first effective solution for category-level 3DGS registration and unlocking new applications. Project webpage: https://bgu-cs-vil.github.io/GSA-project/

Method: Proposes LRC-WeatherNet with early fusion using unified Bird’s Eye View representation and mid-level gated fusion of modality-specific feature maps. The framework adapts to varying sensor reliability under changing weather conditions.

Result: Evaluated on MSU-4S dataset covering nine weather types, LRC-WeatherNet achieves superior classification performance and computational efficiency, significantly outperforming unimodal baselines in adverse conditions.

Conclusion: This is the first work to combine all three modalities (LiDAR, RADAR, camera) for robust, real-time weather classification in autonomous driving, with models and code released publicly.

Abstract: Autonomous vehicles face major perception and navigation challenges in adverse weather such as rain, fog, and snow, which degrade the performance of LiDAR, RADAR, and RGB camera sensors. While each sensor type offers unique strengths, such as RADAR robustness in poor visibility and LiDAR precision in clear conditions, they also suffer distinct limitations when exposed to environmental obstructions. This study proposes LRC-WeatherNet, a novel multi-sensor fusion framework that integrates LiDAR, RADAR, and camera data for real-time classification of weather conditions. By employing both early fusion using a unified Bird’s Eye View representation and mid-level gated fusion of modality-specific feature maps, our approach adapts to the varying reliability of each sensor under changing weather. Evaluated on the extensive MSU-4S dataset covering nine weather types, LRC-WeatherNet achieves superior classification performance and computational efficiency, significantly outperforming unimodal baselines in adverse conditions. This work is the first to combine all three modalities for robust, real-time weather classification in autonomous driving. We release our trained models and source code in https://github.com/nouralhudaalbashir/LRC-WeatherNet.

Method: Created MultiBind benchmark from real multi-person photographs with slot-ordered subject crops, masks, bounding boxes, canonicalized references, inpainted backgrounds, and entity-indexed prompts. Proposed dimension-wise confusion evaluation protocol that matches generated subjects to ground-truth slots and measures slot-to-slot similarity using specialists for face identity, appearance, pose, and expression.

Result: MultiBind reveals binding failures that conventional reconstruction metrics miss, exposing interpretable failure patterns like drift, swap, dominance, and blending in modern multi-reference generators.

Conclusion: MultiBind provides a specialized benchmark and evaluation protocol for diagnosing cross-subject attribute misbinding in multi-reference image generation, enabling better understanding and improvement of binding capabilities.

Abstract: Subject-driven image generation is increasingly expected to support fine-grained control over multiple entities within a single image. In multi-reference workflows, users may provide several subject images, a background reference, and long, entity-indexed prompts to control multiple people within one scene. In this setting, a key failure mode is cross-subject attribute misbinding: attributes are preserved, edited, or transferred to the wrong subject. Existing benchmarks and metrics largely emphasize holistic fidelity or per-subject self-similarity, making such failures hard to diagnose. We introduce MultiBind, a benchmark built from real multi-person photographs. Each instance provides slot-ordered subject crops with masks and bounding boxes, canonicalized subject references, an inpainted background reference, and a dense entity-indexed prompt derived from structured annotations. We also propose a dimension-wise confusion evaluation protocol that matches generated subjects to ground-truth slots and measures slot-to-slot similarity using specialists for face identity, appearance, pose, and expression. By subtracting the corresponding ground-truth similarity matrices, our method separates self-degradation from true cross-subject interference and exposes interpretable failure patterns such as drift, swap, dominance, and blending. Experiments on modern multi-reference generators show that MultiBind reveals binding failures that conventional reconstruction metrics miss.

Method: Hybrid architecture synergizing Mamba and Transformer modules in hierarchical volumetric encoder. Uses Mamba-based layers in early high-resolution stages to reduce computation, and self-attention in later low-resolution stages to refine features. Augmented with generalized rotary position embeddings.

Result: Achieves competitive performance on Synapse, BraTS, and ACDC benchmarks, matching Dice coefficient of larger models. Reduces parameters by up to 75x and substantially decreases FLOPs compared to state-of-the-art models.

Conclusion: SegMaFormer establishes an efficient and high-performing solution for 3D medical image segmentation by balancing computational efficiency with modeling capability through strategic Mamba-Transformer integration.

Abstract: The advent of Transformer and Mamba-based architectures has significantly advanced 3D medical image segmentation by enabling global contextual modeling, a capability traditionally limited in Convolutional Neural Networks (CNNs). However, state-of-the-art Transformer models often entail substantial computational complexity and parameter counts, which is particularly prohibitive for volumetric data and further exacerbated by the limited availability of annotated medical imaging datasets. To address these limitations, this work introduces SegMaFormer, a lightweight hybrid architecture that synergizes Mamba and Transformer modules within a hierarchical volumetric encoder for efficient long-range dependency modeling. The model strategically employs Mamba-based layers in early, high-resolution stages to reduce computational overhead while capturing essential spatial context, and reserves self-attention mechanisms for later, lower-resolution stages to refine feature representation. This design is augmented with generalized rotary position embeddings to enhance spatial awareness. Despite its compact structure, SegMaFormer achieves competitive performance on three public benchmarks (Synapse, BraTS, and ACDC), matching the Dice coefficient of significantly larger models. Empirically, our approach reduces parameters by up to 75x and substantially decreases FLOPs compared to current state-of-the-art models, establishing an efficient and high-performing solution for 3D medical image segmentation.

Method: GeoFlow learns a direct probabilistic mapping to predict displacement (distance and direction) to correct location hypotheses, combined with Iterative Refinement Sampling (IRS) that refines a population of hypotheses iteratively to converge to a robust consensus.

Result: Experiments on KITTI and VIGOR datasets show GeoFlow achieves state-of-the-art efficiency with real-time speeds of 29 FPS while maintaining competitive localization accuracy.

Conclusion: GeoFlow offers a new approach for practical real-time geolocalization systems with flexible inference-time scaling that allows direct trade-off between performance and computation without retraining.

Abstract: Accurate and fast localization is vital for safe autonomous navigation in GPS-denied areas. Fine-Grained Cross-View Geolocalization (FG-CVG) aims to estimate the precise 2-Degree-of-Freedom (2-DoF) location of a ground image relative to a satellite image. However, current methods force a difficult trade-off, with high-accuracy models being slow for real-time use. In this paper, we introduce GeoFlow, a new approach that offers a lightweight and highly efficient framework that breaks this accuracy-speed trade-off. Our technique learns a direct probabilistic mapping, predicting the displacement (in distance and direction) required to correct any given location hypothesis. This is complemented by our novel inference algorithm, Iterative Refinement Sampling (IRS). Instead of trusting a single prediction, IRS refines a population of hypotheses, allowing them to iteratively ‘flow’ from random starting points to a robust, converged consensus. Even its iterative nature, this approach offers flexible inference-time scaling, allowing a direct trade-off between performance and computation without any re-training. Experiments on the KITTI and VIGOR datasets show that GeoFlow achieves state-of-the-art efficiency, running at real-time speeds of 29 FPS while maintaining competitive localization accuracy. This work opens a new path for the development of practical real-time geolocalization systems.

Method: Proposes UNcertainty-guided Compositional Hyperbolic Alignment (UNCHA) that: 1) models part-to-whole semantic representativeness with hyperbolic uncertainty, 2) incorporates this representativeness into contrastive objectives with uncertainty-guided weights, and 3) calibrates uncertainty with an entailment loss regularized by entropy-based term.

Result: Achieves state-of-the-art performance on zero-shot classification, retrieval, and multi-label classification benchmarks, learning hyperbolic embeddings with more accurate part-whole ordering and better compositional structure understanding.

Conclusion: UNCHA successfully enhances hyperbolic VLMs by modeling semantic representativeness through uncertainty, leading to improved understanding of complex multi-object scenes and better hierarchical structure preservation.

Abstract: While Vision-Language Models (VLMs) have achieved remarkable performance, their Euclidean embeddings remain limited in capturing hierarchical relationships such as part-to-whole or parent-child structures, and often face challenges in multi-object compositional scenarios. Hyperbolic VLMs mitigate this issue by better preserving hierarchical structures and modeling part-whole relations (i.e., whole scene and its part images) through entailment. However, existing approaches do not model that each part has a different level of semantic representativeness to the whole. We propose UNcertainty-guided Compositional Hyperbolic Alignment (UNCHA) for enhancing hyperbolic VLMs. UNCHA models part-to-whole semantic representativeness with hyperbolic uncertainty, by assigning lower uncertainty to more representative parts and higher uncertainty to less representative ones for the whole scene. This representativeness is then incorporated into the contrastive objective with uncertainty-guided weights. Finally, the uncertainty is further calibrated with an entailment loss regularized by entropy-based term. With the proposed losses, UNCHA learns hyperbolic embeddings with more accurate part-whole ordering, capturing the underlying compositional structure in an image and improving its understanding of complex multi-object scenes. UNCHA achieves state-of-the-art performance on zero-shot classification, retrieval, and multi-label classification benchmarks. Our code and models are available at: https://github.com/jeeit17/UNCHA.git.

Method: Group3D uses a multimodal large language model to create scene-adaptive vocabulary organized into semantic compatibility groups. These groups act as merge-time constraints, allowing 3D fragment association only when both semantic compatibility and geometric consistency are satisfied.

Result: Group3D achieves state-of-the-art performance on ScanNet and ARKitScenes for multi-view open-vocabulary 3D detection, with strong generalization in zero-shot scenarios.

Conclusion: Integrating semantic constraints directly into instance construction via MLLM-derived compatibility groups effectively mitigates geometry-driven errors and improves open-vocabulary 3D detection performance.

Abstract: Open-vocabulary 3D object detection aims to localize and recognize objects beyond a fixed training taxonomy. In multi-view RGB settings, recent approaches often decouple geometry-based instance construction from semantic labeling, generating class-agnostic fragments and assigning open-vocabulary categories post hoc. While flexible, such decoupling leaves instance construction governed primarily by geometric consistency, without semantic constraints during merging. When geometric evidence is view-dependent and incomplete, this geometry-only merging can lead to irreversible association errors, including over-merging of distinct objects or fragmentation of a single instance. We propose Group3D, a multi-view open-vocabulary 3D detection framework that integrates semantic constraints directly into the instance construction process. Group3D maintains a scene-adaptive vocabulary derived from a multimodal large language model (MLLM) and organizes it into semantic compatibility groups that encode plausible cross-view category equivalence. These groups act as merge-time constraints: 3D fragments are associated only when they satisfy both semantic compatibility and geometric consistency. This semantically gated merging mitigates geometry-driven over-merging while absorbing multi-view category variability. Group3D supports both pose-known and pose-free settings, relying only on RGB observations. Experiments on ScanNet and ARKitScenes demonstrate that Group3D achieves state-of-the-art performance in multi-view open-vocabulary 3D detection, while exhibiting strong generalization in zero-shot scenarios. The project page is available at https://ubin108.github.io/Group3D/.

Method: Two-stage framework: 1) LLM-guided subtitle matching identifies relevant textual cues from video subtitles, fused with queries to generate auxiliary short videos via text-to-video models, creating temporal priors. 2) Augmented queries processed through a multimodal controlled Mamba network with video-guided gating for efficient fusion of generated priors and long sequences while filtering noise.

Result: Experimental evaluations on TVR benchmark show significant improvements over state-of-the-art methods, including reduced computational overhead and higher recall in long-sequence grounding.

Conclusion: The proposed framework effectively addresses limitations in traditional VMR by integrating subtitle contexts with generated temporal priors through an efficient multimodal architecture, achieving better performance in long-sequence video grounding with lower computational costs.

Abstract: Text-driven video moment retrieval (VMR) remains challenging due to limited capture of hidden temporal dynamics in untrimmed videos, leading to imprecise grounding in long sequences. Traditional methods rely on natural language queries (NLQs) or static image augmentations, overlooking motion sequences and suffering from high computational costs in Transformer-based architectures. Existing approaches fail to integrate subtitle contexts and generated temporal priors effectively, we therefore propose a novel two-stage framework for enhanced temporal grounding. In the first stage, LLM-guided subtitle matching identifies relevant textual cues from video subtitles, fused with the query to generate auxiliary short videos via text-to-video models, capturing implicit motion information as temporal priors. In the second stage, augmented queries are processed through a multi-modal controlled Mamba network, extending text-controlled selection with video-guided gating for efficient fusion of generated priors and long sequences while filtering noise. Our framework is agnostic to base retrieval models and widely applicable for multimodal VMR. Experimental evaluations on the TVR benchmark demonstrate significant improvements over state-of-the-art methods, including reduced computational overhead and higher recall in long-sequence grounding.

Method: Reformulates token compression as a spatiotemporal allocation task within a global token retention pool. Uses unified selection mechanism integrating attention weights and semantic similarity to globally select tokens with high contribution and low redundancy. Unselected tokens are merged via clustering and refilled. Inside LLM, introduces text-aware merging for secondary compression based on query relevance. Plug-and-play module requiring no retraining.

Result: Retaining only about 2% of visual tokens preserves 90.1% of baseline performance across multiple benchmarks, reducing FLOPs to roughly 2.6%. Benefits generalize across diverse backbones, decreasing end-to-end inference latency and memory consumption.

Conclusion: The unified spatiotemporal token compression strategy establishes state-of-the-art in video understanding under ultra-low token retention, offering efficient plug-and-play solution for existing Video-LLMs.

Abstract: Video large language models (Video-LLMs) face high computational costs due to large volumes of visual tokens. Existing token compression methods typically adopt a two-stage spatiotemporal compression strategy, relying on stage-specific metrics and an implicit assumption of spatiotemporal separability. Under extremely low retention ratios, however, such approaches often result in unbalanced allocation and loss of visual evidence essential for question answering. We reformulate token compression as a spatiotemporal allocation task within a global token retention pool. We propose a unified selection mechanism that integrates attention weights and semantic similarity to globally select tokens with high contribution and low redundancy. Unselected tokens are merged via clustering and refilled, preserving information integrity. Inside the LLM, we further introduce text-aware merging to perform secondary compression based on query relevance. Without requiring retraining, our method serves as a plug-and-play module compatible with existing Video-LLMs. Experiments show that retaining only about 2% of visual tokens preserves 90.1% of baseline performance across multiple benchmarks, while reducing FLOPs to roughly 2.6%. These benefits generalize across diverse backbones, decreasing end-to-end inference latency and memory consumption. Our unified spatiotemporal token compression strategy establishes the state-of-the-art in video understanding under ultra-low token retention.

Method: Proposes GeoFusion-CAD: encodes CAD programs as hierarchical trees capturing geometry and topology, uses state-space diffusion with lightweight C-Mamba blocks for long-range dependency modeling through selective state transitions.

Result: Achieves superior performance on both short and long command ranges, maintains high geometric fidelity and topological consistency where Transformer models degrade, sets new SOTA for long-sequence parametric CAD generation.

Conclusion: GeoFusion-CAD establishes a scalable foundation for next-generation CAD modeling systems, with code and datasets available.

Abstract: Parametric Computer-Aided Design (CAD) is fundamental to modern 3D modeling, yet existing methods struggle to generate long command sequences, especially under complex geometric and topological dependencies. Transformer-based architectures dominate CAD sequence generation due to their strong dependency modeling, but their quadratic attention cost and limited context windowing hinder scalability to long programs. We propose GeoFusion-CAD, an end-to-end diffusion framework for scalable and structure-aware generation. Our proposal encodes CAD programs as hierarchical trees, jointly capturing geometry and topology within a state-space diffusion process. Specifically, a lightweight C-Mamba block models long-range structural dependencies through selective state transitions, enabling coherent generation across extended command sequences. To support long-sequence evaluation, we introduce DeepCAD-240, an extended benchmark that increases the sequence length ranging from 40 to 240 while preserving sketch-extrusion semantics from the ABC dataset. Extensive experiments demonstrate that GeoFusion-CAD achieves superior performance on both short and long command ranges, maintaining high geometric fidelity and topological consistency where Transformer-based models degrade. Our approach sets new state-of-the-art scores for long-sequence parametric CAD generation, establishing a scalable foundation for next-generation CAD modeling systems. Code and datasets are available at GitHub.

Method: Proposes Bearing-UAV, a purely vision-driven method that leverages global and local structural features and explicitly encodes relative spatial relationships. It jointly predicts UAV absolute location and heading from neighboring features rather than using matching/retrieval paradigm.

Result: Extensive experiments show Bearing-UAV yields lower localization error than previous matching/retrieval methods across diverse terrains. The method also introduces Bearing-UAV-90k, a multi-city benchmark for cross-view localization and navigation evaluation.

Conclusion: Bearing-UAV enables accurate, lightweight, and robust vision-based UAV navigation in GNSS-denied environments by jointly predicting location and heading, addressing limitations of existing cross-view geo-localization approaches.

Abstract: Recent advances in cross-view geo-localization (CVGL) methods have shown strong potential for supporting unmanned aerial vehicle (UAV) navigation in GNSS-denied environments. However, existing work predominantly focuses on matching UAV views to onboard map tiles, which introduces an inherent trade-off between accuracy and storage overhead, and overlooks the importance of the UAV’s heading during navigation. Moreover, the substantial discrepancies and varying overlaps in cross-view scenarios have been insufficiently considered, limiting their generalization to real-world scenarios. In this paper, we present Bearing-UAV, a purely vision-driven cross-view navigation method that jointly predicts UAV absolute location and heading from neighboring features, enabling accurate, lightweight, and robust navigation in the wild. Our method leverages global and local structural features and explicitly encodes relative spatial relationships, making it robust to cross-view variations, misalignment, and feature-sparse conditions. We also present Bearing-UAV-90k, a multi-city benchmark for evaluating cross-view localization and navigation. Extensive experiments show encouraging results that Bearing-UAV yields lower localization error than previous matching/retrieval paradigm across diverse terrains. Our code and dataset will be made publicly available.

Method: Proposes WiFi-GEN network that treats WiFi indoor imaging as a multi-modal image generation task. The generative AI network absorbs challenges into massive parameters and is designed to fit measured WiFi signals to desired imaging output. A large-scale dataset of 80,000 WiFi signal-image pairs is created.

Result: Achieves shape reconstruction accuracy 275% of physical model-based methods, reduces Frechet Inception Distance by 82%, and releases the first large-scale WiFi imaging dataset.

Conclusion: WiFi-GEN demonstrates superior performance for WiFi-based indoor imaging by framing it as a multi-modal image generation task, enabling high-resolution reconstruction from WiFi signals.

Abstract: Indoor imaging is a critical task for robotics and internet-ofthings. WiFi as an omnipresent signal is a promising candidate for carrying out passive imaging and synchronizing the up-to-date information to all connected devices. This is the first research work to consider WiFi indoor imaging as a multi-modal image generation task that converts the measured WiFi power into a high-resolution indoor image. Our proposedWiFi-GEN network achieves a shape reconstruction accuracy that is 275% of that achieved by physical model-based inversion methods. Additionally, the Frechet Inception Distance score has been significantly reduced by 82%. To examine the effectiveness of models for this task, the first large-scale dataset is released containing 80,000 pairs of WiFi signal and imaging target. Our model absorbs challenges for the model-based methods including the nonlinearity, ill-posedness and non-certainty into massive parameters of our generative AI network. The network is also designed to best fit measured WiFi signals and the desired imaging output. Code: https://github.com/CNFightingSjy/WiFiGEN

Method: Uses a single-stream Transformer that processes text, video, and audio within a unified token sequence via self-attention only. Combines this with model distillation, latent-space super-resolution, and a Turbo VAE decoder for efficient inference.

Result: Achieves highest visual quality and text alignment among leading open models, with lowest word error rate (14.60%) for speech intelligibility. Wins 80.0% against Ovi 1.1 and 60.9% against LTX 2.3 in human evaluations. Generates 5-second 256p video in 2 seconds on single H100 GPU.

Conclusion: daVinci-MagiHuman demonstrates that a simple single-stream Transformer architecture can effectively generate synchronized audio-video content with strong human-centric performance, multilingual capabilities, and efficient inference.

Abstract: We present daVinci-MagiHuman, an open-source audio-video generative foundation model for human-centric generation. daVinci-MagiHuman jointly generates synchronized video and audio using a single-stream Transformer that processes text, video, and audio within a unified token sequence via self-attention only. This single-stream design avoids the complexity of multi-stream or cross-attention architectures while remaining easy to optimize with standard training and inference infrastructure. The model is particularly strong in human-centric scenarios, producing expressive facial performance, natural speech-expression coordination, realistic body motion, and precise audio-video synchronization. It supports multilingual spoken generation across Chinese (Mandarin and Cantonese), English, Japanese, Korean, German, and French. For efficient inference, we combine the single-stream backbone with model distillation, latent-space super-resolution, and a Turbo VAE decoder, enabling generation of a 5-second 256p video in 2 seconds on a single H100 GPU. In automatic evaluation, daVinci-MagiHuman achieves the highest visual quality and text alignment among leading open models, along with the lowest word error rate (14.60%) for speech intelligibility. In pairwise human evaluation, it achieves win rates of 80.0% against Ovi 1.1 and 60.9% against LTX 2.3 over 2000 comparisons. We open-source the complete model stack, including the base model, the distilled model, the super-resolution model, and the inference codebase.

Method: Proposes Visual Feedback Layout Model (VFLM), a self-improving framework that leverages visual feedback for iterative refinement. Uses reinforcement learning with a visually grounded reward model that incorporates OCR accuracy. The model performs adaptive reflective generation by leveraging visual information to reflect on previous issues and iteratively generates outputs until satisfactory quality is achieved.

Result: Experiments across multiple benchmarks show that VFLM consistently outperforms advanced MLLMs, existing layout models, and code-only baselines. The framework establishes visual feedback as critical for design-oriented MLLMs.

Conclusion: Visual feedback is a critical factor in layout generation, and VFLM’s self-improving framework with iterative refinement based on visual feedback significantly improves layout quality, readability, and aesthetics compared to existing methods.

Abstract: Recent advances in Multimodal Large Language Models (MLLMs) have enabled automated generation of structured layouts from natural language descriptions. Existing methods typically follow a code-only paradigm that generates code to represent layouts, which are then rendered by graphic engines to produce final images. However, they are blind to the rendered visual outcome, making it difficult to guarantee readability and aesthetics. In this paper, we identify visual feedback as a critical factor in layout generation and propose Visual Feedback Layout Model (VFLM), a self-improving framework that leverages visual feedback iterative refinement. VFLM is capable of performing adaptive reflective generation, which leverages visual information to reflect on previous issues and iteratively generates outputs until satisfactory quality is achieved. It is achieved through reinforcement learning with a visually grounded reward model that incorporates OCR accuracy. By rewarding only the final generated outcome, we can effectively stimulate the model’s iterative and reflective generative capabilities. Experiments across multiple benchmarks show that VFLM consistently outperforms advanced MLLMs, existing layout models, and code-only baselines, establishing visual feedback as critical for design-oriented MLLMs. Our code and data are available at https://github.com/FolSpark/VFLM.

Method: Proposes Superpixel Token Enhancing Network (STENet) with two superpixel-driven cross-modal interaction modules: 1) Superpixel Attention Global Enhancing Module for region-level information with reduced complexity, and 2) Superpixel Attention Local Refining Module for local detail enhancement. Uses updated superpixel generation with expanded neighborhood range.

Result: Experiments on seven RGB-D SOD datasets show STENet achieves competitive performance compared to state-of-the-art methods.

Conclusion: STENet effectively addresses transformer limitations in RGB-D SOD by introducing superpixel-based cross-modal interaction, achieving good performance with reduced computational complexity.

Abstract: Transformer-based methods for RGB-D Salient Object Detection (SOD) have gained significant interest, owing to the transformer’s exceptional capacity to capture long-range pixel dependencies. Nevertheless, current RGB-D SOD methods face challenges, such as the quadratic complexity of the attention mechanism and the limited local detail extraction. To overcome these limitations, we propose a novel Superpixel Token Enhancing Network (STENet), which introduces superpixels into cross-modal interaction. STENet follows the two-stream encoder-decoder structure. Its cores are two tailored superpixel-driven cross-modal interaction modules, responsible for global and local feature enhancement. Specifically, we update the superpixel generation method by expanding the neighborhood range of each superpixel, allowing for flexible transformation between pixels and superpixels. With the updated superpixel generation method, we first propose the Superpixel Attention Global Enhancing Module to model the global pixel-to-superpixel relationship rather than the traditional global pixel-to-pixel relationship, which can capture region-level information and reduce computational complexity. We also propose the Superpixel Attention Local Refining Module, which leverages pixel similarity within superpixels to filter out a subset of pixels (i.e., local pixels) and then performs feature enhancement on these local pixels, thereby capturing concerned local details. Furthermore, we fuse the globally and locally enhanced features along with the cross-scale features to achieve comprehensive feature representation. Experiments on seven RGB-D SOD datasets reveal that our STENet achieves competitive performance compared to state-of-the-art methods. The code and results of our method are available at https://github.com/Mark9010/STENet.

Method: Developed a novel LoRA layer-based model merging method with high cosine similarity threshold and ‘winner-takes-all’ layer selection strategy to integrate HPE capability into CogVLM while preserving original object detection knowledge.

Result: HPE-CogVLM achieves 31.5% reduction in Mean Absolute Error over state-of-the-art CNN model 6DRepNet in cross-dataset evaluation, and outperforms both directly LoRA fine-tuned and task arithmetic-based merged VLMs across all HPE metrics.

Conclusion: The proposed model merging approach successfully integrates head pose estimation into Vision Language Models while maintaining object detection capabilities, demonstrating superior performance over traditional CNN methods and other VLM adaptation techniques.

Abstract: Head pose estimation (HPE) requires a sophisticated understanding of 3D spatial relationships to generate precise yaw, pitch, and roll angles. Previous HPE models, primarily CNN-based, rely on cropped close-up human head images as inputs and often lack robustness in real-world scenario. Vision Language Models (VLMs) can analyze entire images while focusing on specific objects through their attention mechanisms. In this paper, we propose a novel framework to improve the HPE accuracy by leveraging the object detection grounding capability of a VLM, referred to as CogVLM. We empirically find that directly LoRA fine-tuning of this VLM for the HPE task fails to achieve desirable HPE accuracy, while some model merging methods can improve accuracy but frequently produce blended invalid response formats, struggling to handle both object detection and HPE tasks simultaneously. To integrate HPE capability into CogVLM effectively, we develop a novel LoRA layer-based model merging method. This merging approach applies a high cosine similarity threshold and a ‘winner-takes-all’ layer selection strategy, aligning attention to the HPE task while preserving original object detection knowledge. It successfully resolves issues with blended invalid response formats and improves accuracy. Results show that our HPE-CogVLM achieves a 31.5% reduction in Mean Absolute Error over the current state-of-the-art CNN model, 6DRepNet, in cross-dataset evaluation. Furthermore, HPE-CogVLM outperforms both directly LoRA fine-tuned and task arithmetic-based merged VLMs across all HPE metrics.

Method: Proposes a marker-based approach for full six-dimensional hand-eye calibration of robot-mounted OCT probes. The method enables accurate transformation estimation between the robot and OCT probe coordinate systems, allowing for consistent scanning of curved surfaces without relying on image registration.

Result: The calibration method produces highly repeatable transformation estimates. Robotic scanning experiments on two phantom surfaces demonstrate that the proposed calibration enables consistent scanning of large, curved tissue surfaces. The approach avoids error accumulation along scan paths and shows improvement over conventional 3D-translational robotic scanning.

Conclusion: The marker-based 6D hand-eye calibration enables effective robotic OCT scanning of curved tissue surfaces, overcoming limitations of existing methods that rely on image registration or translational-only motion. This approach provides more accurate and consistent scanning for larger tissue structures.

Abstract: Optical coherence tomography (OCT) is a non-invasive volumetric imaging modality with high spatial and temporal resolution. For imaging larger tissue structures, OCT probes need to be moved to scan the respective area. For handheld scanning, stitching of the acquired OCT volumes requires overlap to register the images. For robotic scanning and stitching, a typical approach is to restrict the motion to translations, as this avoids a full hand-eye calibration, which is complicated by the small field of view of most OCT probes. However, stitching by registration or by translational scanning are limited when curved tissue surfaces need to be scanned. We propose a marker for full six-dimensional hand-eye calibration of a robot mounted OCT probe. We show that the calibration results in highly repeatable estimates of the transformation. Moreover, we evaluate robotic scanning of two phantom surfaces to demonstrate that the proposed calibration allows for consistent scanning of large, curved tissue surfaces. As the proposed approach is not relying on image registration, it does not suffer from a potential accumulation of errors along a scan path. We also illustrate the improvement compared to conventional 3D-translational robotic scanning.

Method: Uses FLUX.1-dev flow matching model with unified multi-modal fusion (encoding multi-modal conditions into unified sequence) and training-free test-time scaling paradigm that dynamically steers denoising direction using reward model feedback during inference.

Result: Achieves state-of-the-art performance across multiple standard benchmarks, validating flow matching models for low-level vision tasks and proposing efficient inference-time scaling paradigm for large pre-trained models.

Conclusion: ResFlow-Tuner demonstrates powerful capabilities of flow matching models in image restoration and introduces a novel, efficient inference-time scaling approach suitable for large pre-trained models.

Abstract: Although diffusion-based real-world image restoration (Real-IR) has achieved remarkable progress, efficiently leveraging ultra-large-scale pre-trained text-to-image (T2I) models and fully exploiting their potential remain significant challenges. To address this issue, we propose ResFlow-Tuner, an image restoration framework based on the state-of-the-art flow matching model, FLUX.1-dev, which integrates unified multi-modal fusion (UMMF) with test-time scaling (TTS) to achieve unprecedented restoration performance. Our approach fully leverages the advantages of the Multi-Modal Diffusion Transformer (MM-DiT) architecture by encoding multi-modal conditions into a unified sequence that guides the synthesis of high-quality images. Furthermore, we introduce a training-free test-time scaling paradigm tailored for image restoration. During inference, this technique dynamically steers the denoising direction through feedback from a reward model (RM), thereby achieving significant performance gains with controllable computational overhead. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple standard benchmarks. This work not only validates the powerful capabilities of the flow matching model in low-level vision tasks but, more importantly, proposes a novel and efficient inference-time scaling paradigm suitable for large pre-trained models.

Method: Multi-stage pipeline: 1) Prompt Decomposer extracts entities, attributes, and spatial metadata; 2) Expert detectors provide visual grounding of object positions; 3) Vision-language model uses chain-of-thought reasoning to assess complex spatial relations. Also introduces SpatRelBench benchmark for comprehensive evaluation.

Result: Incorporating SpatialReward into RL training for Stable Diffusion and FLUX consistently improves spatial consistency and overall generation quality, with results better aligned with human judgments.

Conclusion: Verifiable reward models like SpatialReward have significant potential for enabling more accurate and controllable optimization in text-to-image generation by addressing spatial relationship evaluation.

Abstract: Recent advances in text-to-image (T2I) generation via reinforcement learning (RL) have benefited from reward models that assess semantic alignment and visual quality. However, most existing reward models pay limited attention to fine-grained spatial relationships, often producing images that appear plausible overall yet contain inaccuracies in object positioning. In this work, we present \textbf{SpatialReward}, a verifiable reward model explicitly designed to evaluate spatial layouts in generated images. SpatialReward adopts a multi-stage pipeline: a \emph{Prompt Decomposer} extracts entities, attributes, and spatial metadata from free-form prompts; expert detectors provide accurate visual grounding of object positions and attributes; and a vision-language model applies chain-of-thought reasoning over grounded observations to assess complex spatial relations that are challenging for rule-based methods. To more comprehensively evaluate spatial relationships in generated images, we introduce \textbf{SpatRelBench}, a benchmark covering object attributes, orientation, inter-object relations, and rendered text placement. Experiments on Stable Diffusion and FLUX show that incorporating SpatialReward into RL training consistently improves spatial consistency and overall generation quality, with results aligned more closely to human judgments. These findings indicate that verifiable reward models hold considerable potential for enabling more accurate and controllable optimization in text-to-image generation models.

Method: Proposes GTSR using two sets of Gaussians: surface Gaussians for geometry and interior Gaussians for light scattering. Uses Fresnel term to blend both sets for rendering. Incorporates Disney BSDF model with deferred rendering to enhance normal and depth constraints for better detail reconstruction.

Result: Outperforms baseline methods on NeuralTO Syn dataset with real-time rendering performance. Successfully adapts to different translucent materials on extended dataset with varying material properties.

Conclusion: GTSR provides an efficient and effective solution for translucent object reconstruction that properly models optical properties while maintaining real-time rendering capabilities.

Abstract: Reconstructing translucent objects from multi-view images is a difficult problem. Previously, researchers have used differentiable path tracing and the neural implicit field, which require relatively large computational costs. Recently, many works have achieved good reconstruction results for opaque objects based on a 3DGS pipeline with much higher efficiency. However, such methods have difficulty dealing with translucent objects, because they do not consider the optical properties of translucent objects. In this paper, we propose a novel 3DGS-based pipeline (GTSR) to reconstruct the surface geometry of translucent objects. GTSR combines two sets of Gaussians, surface and interior Gaussians, which are used to model the surface and scattering color when lights pass translucent objects. To render the appearance of translucent objects, we introduce a method that uses the Fresnel term to blend two sets of Gaussians. Furthermore, to improve the reconstructed details of non-contour areas, we introduce the Disney BSDF model with deferred rendering to enhance constraints of the normal and depth. Experimental results demonstrate that our method outperforms baseline reconstruction methods on the NeuralTO Syn dataset while showing great real-time rendering performance. We also extend the dataset with new translucent objects of varying material properties and demonstrate our method can adapt to different translucent materials.

Method: Proposes DTVI with two stages: 1) Category-aware sequence-level intervention on full prompt embeddings to capture distributed malicious semantics, and 2) Attenuation of remaining unsafe influences during visual generation stage.

Result: Achieves average Defense Success Rate (DSR) of 94.43% across sexual-category benchmarks and 88.56% across seven unsafe categories, while maintaining generation quality on benign prompts.

Conclusion: DTVI provides effective and robust defense against unsafe content generation in T2I models while preserving reasonable generation quality for benign prompts.

Abstract: Text-to-Image (T2I) diffusion models have demonstrated strong generation ability, but their potential to generate unsafe content raises significant safety concerns. Existing inference-time defense methods typically perform category-agnostic token-level intervention in the text embedding space, which fails to capture malicious semantics distributed across the full token sequence and remains vulnerable to adversarial prompts. In this paper, we propose DTVI, a dual-stage inference-time defense framework for safe T2I generation. Unlike existing methods that intervene on specific token embeddings, our method introduces category-aware sequence-level intervention on the full prompt embedding to better capture distributed malicious semantics, and further attenuates the remaining unsafe influences during the visual generation stage. Experimental results on real-world unsafe prompts, adversarial prompts, and multiple harmful categories show that our method achieves effective and robust defense while preserving reasonable generation quality on benign prompts, obtaining an average Defense Success Rate (DSR) of 94.43% across sexual-category benchmarks and 88.56 across seven unsafe categories, while maintaining generation quality on benign prompts.

Method: Introduces a Structured Reasoning framework that performs text-conditioned spatial layout editing via scene-graph reasoning. Given an input scene graph and natural-language instruction, the model reasons over the graph to generate an updated scene graph that satisfies the text condition while maintaining spatial coherence. Uses explicit structured relational representations to guide the reasoning process.

Result: Achieves 15% improvement in IoU and 25% reduction in center-distance error compared to CoT-SFT and vanilla GRPO baselines. Compared to SOTA zero-shot LLMs, achieves up to 20% higher mIoU, demonstrating markedly improved spatial precision. Evaluated on a new text-guided layout editing benchmark with sorting, spatial alignment, and room-editing tasks.

Conclusion: The structured reasoning approach with scene-graph representations significantly improves spatial understanding and layout consistency in visual editing tasks, offering better interpretability and control over spatial relationships compared to existing LLM/VLM approaches.

Abstract: Large Language Models (LLMs) and Vision Language Models (VLMs) have shown impressive reasoning abilities, yet they struggle with spatial understanding and layout consistency when performing fine-grained visual editing. We introduce a Structured Reasoning framework that performs text-conditioned spatial layout editing via scene-graph reasoning. Given an input scene graph and a natural-language instruction, the model reasons over the graph to generate an updated scene graph that satisfies the text condition while maintaining spatial coherence. By explicitly guiding the reasoning process through structured relational representations, our approach improves both interpretability and control over spatial relationships. We evaluate our method on a new text-guided layout editing benchmark encompassing sorting, spatial alignment, and room-editing tasks. Our training paradigm yields an average 15% improvement in IoU and 25% reduction in center-distance error compared to Chain of Thought Fine-tuning (CoT-SFT) and vanilla GRPO baselines. Compared to SOTA zero-shot LLMs, our best models achieve up to 20% higher mIoU, demonstrating markedly improved spatial precision.

Method: Introduces FontCrafter framework with: 1) in-context generation strategy treating element images as visual context using inpainting model to transfer styles at pixel level, 2) Context-aware Mask Adapter (CMA) to inject shape information for glyph control, 3) training-free attention redirection for region-aware style control and stroke hallucination suppression, and 4) edge repainting for natural boundaries.

Result: FontCrafter achieves strong zero-shot generation performance, particularly in preserving structural and textural fidelity of both object elements (structured) and amorphous elements (unstructured). It also supports flexible controls like style mixture.

Conclusion: The proposed element-driven approach with FontCrafter framework successfully addresses limitations in artistic font generation, enabling diverse style creation with fine-grained control while maintaining high fidelity to reference elements.

Abstract: Artistic font generation aims to synthesize stylized glyphs based on a reference style. However, existing approaches suffer from limited style diversity and coarse control. In this work, we explore the potential of element-driven artistic font generation. Elements are the fundamental visual units of a font, serving as reference images for the desired style. Conceptually, we categorize elements into object elements (e.g., flowers or stones) with distinct structures and amorphous elements (e.g., flames or clouds) with unstructured textures. We introduce FontCrafter, an element-driven framework for font creation, and construct a large-scale dataset, ElementFont, which contains diverse element types and high-quality glyph images. However, achieving high-fidelity reconstruction of both texture and structure of reference elements remains challenging. To address this, we propose an in-context generation strategy that treats element images as visual context and uses an inpainting model to transfer element styles into glyph regions at the pixel level. To further control glyph shapes, we design a lightweight Context-aware Mask Adapter (CMA) that injects shape information. Moreover, a training-free attention redirection mechanism enables region-aware style control and suppresses stroke hallucination. In addition, edge repainting is applied to make boundaries more natural. Extensive experiments demonstrate that FontCrafter achieves strong zero-shot generation performance, particularly in preserving structural and textural fidelity, while also supporting flexible controls such as style mixture.

Method: UniMotion uses a Cross-Modal Aligned Motion VAE (CMA-VAE) and symmetric dual-path embedders to create parallel continuous pathways for Motion and RGB within a shared LLM backbone. It introduces Dual-Posterior KL Alignment (DPA) to inject visual-semantic priors into motion representations without requiring images at inference, and Latent Reconstruction Alignment (LRA) to address the cold-start problem through self-supervised pre-training.

Result: UniMotion achieves state-of-the-art performance across seven tasks spanning any-to-any understanding, generation, and editing among the three modalities (motion, language, RGB images), with especially strong advantages on cross-modal compositional tasks.

Conclusion: UniMotion demonstrates that treating motion as a first-class continuous modality and using novel alignment techniques enables effective unified modeling of motion, language, and vision, opening new possibilities for multimodal understanding and generation.

Abstract: We present UniMotion, to our knowledge the first unified framework for simultaneous understanding and generation of human motion, natural language, and RGB images within a single architecture. Existing unified models handle only restricted modality subsets (e.g., Motion-Text or static Pose-Image) and predominantly rely on discrete tokenization, which introduces quantization errors and disrupts temporal continuity. UniMotion overcomes both limitations through a core principle: treating motion as a first-class continuous modality on equal footing with RGB. A novel Cross-Modal Aligned Motion VAE (CMA-VAE) and symmetric dual-path embedders construct parallel continuous pathways for Motion and RGB within a shared LLM backbone. To inject visual-semantic priors into motion representations without requiring images at inference, we propose Dual-Posterior KL Alignment (DPA), which distills a vision-fused encoder’s richer posterior into the motion-only encoder. To address the cold-start problem – where text supervision alone is too sparse to calibrate the newly introduced motion pathway – we further propose Latent Reconstruction Alignment (LRA), a self-supervised pre-training strategy that uses dense motion latents as unambiguous conditions to co-calibrate the embedder, backbone, and flow head, establishing a stable motion-aware foundation for all downstream tasks. UniMotion achieves state-of-the-art performance across seven tasks spanning any-to-any understanding, generation, and editing among the three modalities, with especially strong advantages on cross-modal compositional tasks.

Method: Converts dense 3D spatial information from 2D images into linguistic expressions, then uses LLMs with multi-turn Chain-of-Thought reasoning to progressively inject spatial knowledge into vision encoders, building hierarchical spatial understanding.

Result: SpatialBoost improves DINOv3 performance from 55.9 to 59.7 mIoU on ADE20K, achieving state-of-the-art with 3.8% gain, showing effectiveness across benchmarks requiring 3D perception and general vision abilities.

Conclusion: SpatialBoost successfully enhances spatial awareness of vision encoders by leveraging linguistic descriptions of 3D spatial information through LLM reasoning, bridging the gap between 2D training and 3D understanding.

Abstract: Despite the remarkable success of large-scale pre-trained image representation models (i.e., vision encoders) across various vision tasks, they are predominantly trained on 2D image data and therefore often fail to capture 3D spatial relationships between objects and backgrounds in the real world, constraining their effectiveness in many downstream applications. To address this, we propose SpatialBoost, a scalable framework that enhances the spatial awareness of existing pre-trained vision encoders by injecting 3D spatial knowledge expressed in linguistic descriptions. The core idea involves converting dense 3D spatial information from 2D images into linguistic expressions, which is then used to inject such spatial knowledge into vision encoders through a Large Language Model (LLM). To this end, we adopt a multi-turn Chain-of-Thought (CoT) reasoning process that progressively incorporates dense spatial knowledge and builds hierarchical spatial understanding. To validate effectiveness, we adapt SpatialBoost to state-of-the-art vision encoders such as DINOv3, and evaluate its performance gains on a wide range of benchmarks requiring both 3D perception and general vision abilities. For instance, SpatialBoost improves DINOv3 performance from 55.9 to 59.7 mIoU on ADE20K, achieving state-of-the-art performance with 3.8% gain over the pre-trained DINOv3.

Method: UNITE uses a Generative Encoder that serves as both image tokenizer and latent generator via weight sharing. Tokenization and generation are treated as the same latent inference problem under different conditioning regimes. Single-stage training jointly optimizes both tasks via two forward passes through the same network.

Result: Achieves FID 2.12 and 1.73 for Base and Large models on ImageNet 256×256, near state-of-the-art performance without adversarial losses or pretrained encoders. Works across image and molecule modalities.

Conclusion: Single-stage joint training of tokenization and generation from scratch is feasible and effective, creating a “common latent language” that benefits both tasks through shared parameter optimization.

Abstract: Latent diffusion models (LDMs) enable high-fidelity synthesis by operating in learned latent spaces. However, training state-of-the-art LDMs requires complex staging: a tokenizer must be trained first, before the diffusion model can be trained in the frozen latent space. We propose UNITE - an autoencoder architecture for unified tokenization and latent diffusion. UNITE consists of a Generative Encoder that serves as both image tokenizer and latent generator via weight sharing. Our key insight is that tokenization and generation can be viewed as the same latent inference problem under different conditioning regimes: tokenization infers latents from fully observed images, whereas generation infers them from noise together with text or class conditioning. Motivated by this, we introduce a single-stage training procedure that jointly optimizes both tasks via two forward passes through the same Generative Encoder. The shared parameters enable gradients to jointly shape the latent space, encouraging a “common latent language”. Across image and molecule modalities, UNITE achieves near state of the art performance without adversarial losses or pretrained encoders (e.g., DINO), reaching FID 2.12 and 1.73 for Base and Large models on ImageNet 256 x 256. We further analyze the Generative Encoder through the lenses of representation alignment and compression. These results show that single stage joint training of tokenization & generation from scratch is feasible.

Method: BayesMM models textual priors (from semantic prompts) and streaming visual features as Gaussian distributions, fuses them via Bayesian model averaging that automatically adjusts modality contributions based on posterior evidence, enabling continual adaptation to test-time data without training.

Result: Extensive experiments on multiple point cloud benchmarks show BayesMM maintains robustness under distributional shifts, achieving over 4% average improvement compared to existing methods.

Conclusion: BayesMM provides an effective Bayesian framework for test-time adaptation in multimodal 3D point cloud analysis that addresses limitations of cache-based methods through principled distribution learning and fusion.

Abstract: Multimodal 3D vision-language models show strong generalization across diverse 3D tasks, but their performance still degrades notably under domain shifts. This has motivated recent studies on test-time adaptation (TTA), which enables models to adapt online using test-time data. Among existing TTA methods, cache-based mechanisms are widely adopted for leveraging previously observed samples in online prediction refinement. However, they store only limited historical information, leading to progressive information loss as the test stream evolves. In addition, their prediction logits are fused heuristically, making adaptation unstable. To address these limitations, we propose BayesMM, a Multimodal Bayesian Distribution Learning framework for test-time point cloud analysis. BayesMM models textual priors and streaming visual features of each class as Gaussian distributions: textual parameters are derived from semantic prompts, while visual parameters are updated online with arriving samples. The two modalities are fused via Bayesian model averaging, which automatically adjusts their contributions based on posterior evidence, yielding a unified prediction that adapts continually to evolving test-time data without training. Extensive experiments on multiple point cloud benchmarks demonstrate that BayesMM maintains robustness under distributional shifts, yielding over 4% average improvement.

Method: P-Flow uses vision-language models to perform test-time prompt optimization, refining prompts based on discrepancy between reference video effects and generated output, enabling iterative improvement without modifying the underlying video generation model.

Result: P-Flow achieves high-fidelity and diverse visual effect customization, outperforming other models on both text-to-video and image-to-video generation tasks.

Conclusion: The proposed training-free framework effectively customizes dynamic visual effects through iterative prompt optimization, addressing the challenge of complex temporal reasoning in video generation.

Abstract: Recent advancements in video generation models have significantly improved their ability to follow text prompts. However, the customization of dynamic visual effects, defined as temporally evolving and appearance-driven visual phenomena like object crushing or explosion, remains underexplored. Prior works on motion customization or control mainly focus on low-level motions of the subject or camera, which can be guided using explicit control signals such as motion trajectories. In contrast, dynamic visual effects involve higher-level semantics that are more naturally suited for control via text prompts. However, it is hard and time-consuming for humans to craft a single prompt that accurately specifies these effects, as they require complex temporal reasoning and iterative refinement over time. To address this challenge, we propose P-Flow, a novel training-free framework for customizing dynamic visual effects in video generation without modifying the underlying model. By leveraging the semantic and temporal reasoning capabilities of vision-language models, P-Flow performs test-time prompt optimization, refining prompts based on the discrepancy between the visual effects of the reference video and the generated output. Through iterative refinement, the prompts evolve to better induce the desired dynamic effect in novel scenes. Experiments demonstrate that P-Flow achieves high-fidelity and diverse visual effect customization and outperforms other models on both text-to-video and image-to-video generation tasks. Code is available at https://github.com/showlab/P-Flow.

Method: Proposes NullSteer, a null-space projected activation defense framework that constructs refusal directions within model activations through linear transformation. Maintains zero perturbation in benign subspace while dynamically inducing refusal along potentially harmful directions.

Result: Significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15% on MiniGPT-4) while maintaining comparable performance to original model on general benchmarks.

Conclusion: NullSteer effectively balances safety and utility by theoretically achieving safety enhancement without impairing model’s general capabilities, addressing limitations of existing activation steering methods.

Abstract: As vision-language models (VLMs) are increasingly deployed in open-world scenarios, they can be easily induced by visual jailbreak attacks to generate harmful content, posing serious risks to model safety and trustworthy usage. Recent activation steering methods inject directional vectors into model activations during inference to induce refusal behaviors and have demonstrated effectiveness. However, a steering vector may both enhance refusal ability and cause over-refusal, thereby degrading model performance on benign inputs. Moreover, due to the lack of theoretical interpretability, these methods still suffer from limited robustness and effectiveness. To better balance safety and utility, we propose NullSteer, a null-space projected activation defense framework. Our method constructs refusal directions within model activations through a linear transformation: it maintains zero perturbation within the benign subspace while dynamically inducing refusal along potentially harmful directions, thereby theoretically achieving safety enhancement without impairing the model’s general capabilities. Extensive experiments show that NullSteer significantly reduces harmful outputs under various jailbreak attacks (average ASR reduction over 15 percent on MiniGPT-4) while maintaining comparable performance to the original model on general benchmarks.

Method: Proposes IAG (Input-Aware, text-Guided attack) using a text-conditioned UNet to embed imperceptible target semantic cues into visual inputs. Employs joint training objective balancing language capability with perceptual reconstruction to ensure imperceptibility, effectiveness, and stealth.

Result: Achieves best attack success rates compared to baselines on multiple VLMs (LLaVA, InternVL, Ferret) and benchmarks (RefCOCO, RefCOCO+, RefCOCOg, Flickr30k Entities, ShowUI) without compromising clean accuracy. Maintains robustness against existing defenses and shows transferability across datasets and models.

Conclusion: Reveals critical security risks in grounding-capable VLMs, demonstrating the feasibility of realistic multi-target backdoor attacks. Highlights the need for further research on trustworthy multimodal understanding and security in vision-language systems.

Abstract: Recent advances in vision-language models (VLMs) have significantly enhanced the visual grounding task, which involves locating objects in an image based on natural language queries. Despite these advancements, the security of VLM-based grounding systems has not been thoroughly investigated. This paper reveals a novel and realistic vulnerability: the first multi-target backdoor attack on VLM-based visual grounding. Unlike prior attacks that rely on static triggers or fixed targets, we propose IAG, a method that dynamically generates input-aware, text-guided triggers conditioned on any specified target object description to execute the attack. This is achieved through a text-conditioned UNet that embeds imperceptible target semantic cues into visual inputs while preserving normal grounding performance on benign samples. We further develop a joint training objective that balances language capability with perceptual reconstruction to ensure imperceptibility, effectiveness, and stealth. Extensive experiments on multiple VLMs (e.g., LLaVA, InternVL, Ferret) and benchmarks (RefCOCO, RefCOCO+, RefCOCOg, Flickr30k Entities, and ShowUI) demonstrate that IAG achieves the best ASRs compared with other baselines on almost all settings without compromising clean accuracy, maintaining robustness against existing defenses, and exhibiting transferability across datasets and models. These findings underscore critical security risks in grounding-capable VLMs and highlight the need for further research on trustworthy multimodal understanding.

Method: Combines free-moving part segmentation (using point-tracking and feature model priors) with joint estimation (calibrating object-to-camera poses, recovering joint type/axis), and 3DGS-based end-to-end optimization to jointly reconstruct textures, geometry, and joint angles.

Result: Outperforms existing methods on two benchmarks and real-world free-moving articulated objects, excelling in free-moving scenarios while remaining competitive in previous reconstruction settings.

Conclusion: FreeArtGS provides a practical and effective solution for realistic asset generation in articulated object reconstruction, particularly valuable for AR and robotics applications.

Abstract: The increasing demand for augmented reality and robotics is driving the need for articulated object reconstruction with high scalability. However, existing settings for reconstructing from discrete articulation states or casual monocular videos require non-trivial axis alignment or suffer from insufficient coverage, limiting their applicability. In this paper, we introduce FreeArtGS, a novel method for reconstructing articulated objects under free-moving scenario, a new setting with a simple setup and high scalability. FreeArtGS combines free-moving part segmentation with joint estimation and end-to-end optimization, taking only a monocular RGB-D video as input. By optimizing with the priors from off-the-shelf point-tracking and feature models, the free-moving part segmentation module identifies rigid parts from relative motion under unconstrained capture. The joint estimation module calibrates the unified object-to-camera poses and recovers joint type and axis robustly from part segmentation. Finally, 3DGS-based end-to-end optimization is implemented to jointly reconstruct visual textures, geometry, and joint angles of the articulated object. We conduct experiments on two benchmarks and real-world free-moving articulated objects. Experimental results demonstrate that FreeArtGS consistently excels in reconstructing free-moving articulated objects and remains highly competitive in previous reconstruction settings, proving itself a practical and effective solution for realistic asset generation. The project page is available at: https://freeartgs.github.io/

Method: StreamingClaw integrates five core capabilities: 1) real-time streaming reasoning, 2) reasoning about future events and proactive interaction with evolving objectives, 3) multimodal long-term storage with hierarchical evolution and efficient retrieval across agents, 4) perception-decision-action closed loop with streaming tools and action-centric skills, and 5) compatibility with OpenClaw framework for community resources.

Result: The framework integrates online real-time reasoning, multimodal long-term memory, and proactive interaction within a unified system, enabling direct control of the physical world through executable actions for practical embodied interaction deployment.

Conclusion: StreamingClaw addresses key bottlenecks in streaming video understanding for embodied intelligence by providing a comprehensive framework that supports real-time multimodal interaction, memory, and action execution in physical environments.

Abstract: Applications such as embodied intelligence rely on a real-time perception-decision-action closed loop, posing stringent challenges for streaming video understanding. However, current agents suffer from fragmented capabilities, such as supporting only offline video understanding, lacking long-term multimodal memory mechanisms, or struggling to achieve real-time reasoning and proactive interaction under streaming inputs. These shortcomings have become a key bottleneck for preventing them from sustaining perception, making real-time decisions, and executing actions in real-world environments. To alleviate these issues, we propose StreamingClaw, a unified agent framework for streaming video understanding and embodied intelligence. It is also an OpenClaw-compatible framework that supports real-time, multimodal streaming interaction. StreamingClaw integrates five core capabilities: (1) It supports real-time streaming reasoning. (2) It supports reasoning about future events and proactive interaction under the online evolution of interaction objectives. (3) It supports multimodal long-term storage, hierarchical evolution, and efficient retrieval of shared memory across multiple agents. (4) It supports a closed-loop of perception-decision-action. In addition to conventional tools and skills, it also provides streaming tools and action-centric skills tailored for real-world physical environments. (5) It is compatible with the OpenClaw framework, allowing it to fully leverage the resources and support of the open-source community. With these designs, StreamingClaw integrates online real-time reasoning, multimodal long-term memory, and proactive interaction within a unified framework. Moreover, by translating decisions into executable actions, it enables direct control of the physical world, supporting practical deployment of embodied interaction.

Method: Proposes physics-regularized implicit neural representations (INR) for surrogate-based motion modeling. Uses trajectory-aware spatio-temporally continuous diffeomorphic motion representation with biophysical constraints for physiological plausibility.

Result: Trajectory-aware approach performs on par in interpolation and improves extrapolation compared to initial INR approach. Both INR approaches perform equally well in interpolation but underperform in extrapolation compared to sequential registration methods.

Conclusion: INRs show strong potential for respiratory motion modeling with methodological advantages, though current extrapolation performance needs improvement. Physics regularization enhances physiological plausibility.

Abstract: A precise spatial delivery of the radiation dose is crucial for the treatment success in radiotherapy. In the lung and upper abdominal region, respiratory motion introduces significant treatment uncertainties, requiring special motion management techniques. To address this, respiratory motion models are commonly used to infer the patient-specific respiratory motion and target the dose more efficiently. In this work, we investigate the possibility of using implicit neural representations (INR) for surrogate-based motion modeling. Therefore, we propose physics-regularized implicit surrogate-based modeling for respiratory motion (PRISM-RM). Our new integrated respiratory motion model is free of a fixed reference breathing state. Unlike conventional pairwise registration techniques, our approach provides a trajectory-aware spatio-temporally continuous and diffeomorphic motion representation, improving generalization to extrapolation scenarios. We introduce biophysical constraints, ensuring physiologically plausible motion estimation across time beyond the training data. Our results show that our trajectory-aware approach performs on par in interpolation and improves the extrapolation ability compared to our initially proposed INR-based approach. Compared to sequential registration-based approaches both our approaches perform equally well in interpolation, but underperform in extrapolation scenarios. However, the methodical features of INRs make them particularly effective for respiratory motion modeling, and with their performance steadily improving, they demonstrate strong potential for advancing this field.

Method: DA-VAE uses explicit latent layout: first C channels from pretrained VAE at base resolution, plus D channels for higher-resolution details. Detail-alignment mechanism preserves original structure. Warm-start fine-tuning adapts pretrained diffusion backbone lightly.

Result: Enables 1024×1024 generation with Stable Diffusion 3.5 using only 32×32 tokens (4× fewer), within 5 H100-days. Unlocks 2048×2048 generation with SD3.5, achieving 6× speedup while preserving quality. Validated quantitatively on ImageNet.

Conclusion: DA-VAE provides efficient high-resolution generation for latent diffusion models by expanding VAE compression while preserving pretrained structure, requiring minimal backbone adaptation.

Abstract: Reducing token count is crucial for efficient training and inference of latent diffusion models, especially at high resolution. A common strategy is to build high-compression image tokenizers with more channels per token. However, when trained only for reconstruction, high-dimensional latent spaces often lose meaningful structure, making diffusion training harder. Existing methods address this with extra objectives such as semantic alignment or selective dropout, but usually require costly diffusion retraining. Pretrained diffusion models, however, already exhibit a structured, lower-dimensional latent space; thus, a simpler idea is to expand the latent dimensionality while preserving this structure. We therefore propose \textbf{D}etail-\textbf{A}ligned VAE, which increases the compression ratio of a pretrained VAE with only lightweight adaptation of the pretrained diffusion backbone. DA-VAE uses an explicit latent layout: the first $C$ channels come directly from the pretrained VAE at a base resolution, while an additional $D$ channels encode higher-resolution details. A simple detail-alignment mechanism encourages the expanded latent space to retain the structure of the original one. With a warm-start fine-tuning strategy, our method enables $1024 \times 1024$ image generation with Stable Diffusion 3.5 using only $32 \times 32$ tokens, $4\times$ fewer than the original model, within 5 H100-days. It further unlocks $2048 \times 2048$ generation with SD3.5, achieving a $6\times$ speedup while preserving image quality. We also validate the method and its design choices quantitatively on ImageNet.

Method: Introduces OpenEarth-Agent framework with adaptive workflow planning and tool creation capabilities instead of calling predefined tools. Uses open-ended integration of multi-stage tools and cross-domain knowledge bases for robust execution across the entire EO pipeline.

Result: OpenEarth-Agent successfully masters full-pipeline EO across multiple domains in open environments. On cross-benchmark Earth-Bench, with only 6 essential pre-trained models, it achieves performance comparable to tool-calling agents using 104 specialized tools, and outperforms them with complete toolset. Created tools show superior robustness to data anomalies.

Conclusion: OpenEarth-Agent represents a significant advancement for autonomous Earth Observation in open environments, demonstrating that tool-creation agents can effectively generalize to diverse data and tasks beyond the limitations of traditional tool-calling approaches.

Abstract: Earth Observation (EO) is essential for perceiving dynamic land surface changes, yet deploying autonomous EO in open environments is hindered by the immense diversity of multi-source data and heterogeneous tasks. While remote sensing agents have emerged to streamline EO workflows, existing tool-calling agents are confined to closed environments. They rely on pre-defined tools and are restricted to narrow scope, limiting their generalization to the diverse data and tasks. To overcome these limitations, we introduce OpenEarth-Agent, the first tool-creation agent framework tailored for open-environment EO. Rather than calling predefined tools, OpenEarth-Agent employs adaptive workflow planning and tool creation to generalize to unseen data and tasks. This adaptability is bolstered by an open-ended integration of multi-stage tools and cross-domain knowledge bases, enabling robust execution in the entire EO pipeline across multiple application domains. To comprehensively evaluate EO agents in open environments, we propose OpenEarth-Bench, a novel benchmark comprising 596 real-world, full-pipeline cases across seven application domains, explicitly designed to assess agents’ adaptive planning and tool creation capabilities. Only essential pre-trained model tools are provided in this benchmark, devoid of any other predefined task-specific tools. Extensive experiments demonstrate that OpenEarth-Agent successfully masters full-pipeline EO across multiple domains in the open environment. Notably, on the cross-benchmark Earth-Bench, our tool-creating agent equipped with 6 essential pre-trained models achieves performance comparable to tool-calling agents relying on 104 specialized tools, and significantly outperforms them when provided with the complete toolset. In several cases, the created tools exhibit superior robustness to data anomalies compared to human-engineered counterparts.

Method: Redesigned fine-tuning process with: 1) Controlled generation of synthetic data and annotations free from bias, distribution imbalance, and errors by comprehensively sampling object attributes (color, shape, size, position); 2) Fine-tuning state-of-the-art VLMs on this balanced synthetic dataset and evaluating transferability to real-world data on absolute position tasks.

Result: Two key findings: 1) Fine-tuning on balanced synthetic data yields uniform performance across visual scenes and mitigates common biases; 2) Fine-tuning on synthetic stimuli improves performance by 13% on real-world data (COCO), outperforming models fine-tuned on the full COCO train set.

Conclusion: Controlled synthetic data generation for VLM fine-tuning is an effective approach to address biases and distribution issues in real-world data, leading to better performance and more balanced models.

Abstract: Fine-tuning Vision-Language Models (VLMs) is a common strategy to improve performance following an ad-hoc data collection and annotation of real-world scenes. However, this process is often prone to biases, errors, and distribution imbalance, resulting in overfitting and imbalanced performance. Although a few studies have tried to address this problem by generating synthetic data, they lacked control over distribution bias and annotation quality. To address these challenges, we redesign the fine-tuning process in two ways. First, we control the generation of data and its annotations, ensuring it is free from bias, distribution imbalance, and annotation errors. We automatically construct the dataset by comprehensively sampling objects’ attributes, including color, shape, size, and position within the scene. Secondly, using this annotated dataset, we fine-tune state-of-the-art VLMs and assess performance transferability to real-world data on the absolute position task. We conduct exhaustive evaluations on both synthetic and real-world benchmarks. Our experiments reveal two key findings: 1) fine-tuning on balanced synthetic data yields uniform performance across the visual scene and mitigates common biases; and 2) fine-tuning on synthetic stimuli improves performance by 13% on real-world data (COCO), outperforming models fine-tuned on the full COCO train set.

Method: Proposes Asymmetric Constrained Preference Optimization (ACPO) - a modality-agnostic alignment mechanism that applies dynamic, target-oriented scaling to preference optimization. It uses complexity-aware scaling coefficients applied exclusively to rejected rewards, asymmetrically suppressing gradient flow on rejected terms while preserving chosen distribution as gradient-stable reference.

Result: ACPO effectively reverses chosen-reward degradation of standard DPO, halts Visual Anchor Collapse, outperforms baselines on hallucination benchmarks (HallusionBench, MM-IFEval) and general leaderboards (MMBench, MMStar, OCRBenchV2) while improving general capabilities.

Conclusion: ACPO addresses fundamental optimization flaws in DPO for multimodal alignment, preventing visual evidence abandonment and reducing hallucinations through asymmetric gradient suppression, making it crucial for vision-language model alignment.

Abstract: While Direct Preference Optimization (DPO) has become the de facto approach for aligning Large Vision-Language Models (LVLMs), it suffers from Likelihood Displacement, where the probability of both chosen and rejected responses collapses. This optimization flaw is especially detrimental in multimodal settings: the erosion of chosen likelihoods – a failure we term Visual Anchor Collapse – causes models to abandon visual evidence for strong language priors, precipitating significant hallucinations. To address this, we propose Asymmetric Constrained Preference Optimization (ACPO), a modality-agnostic alignment mechanism that applies dynamic, target-oriented scaling to preference optimization. ACPO derives a complexity-aware scaling coefficient applied exclusively to the rejected reward, asymmetrically suppressing the gradient flow on the rejected term while preserving the chosen distribution as a gradient-stable reference. While fundamentally a general-purpose objective, breaking this gradient symmetry is crucial for multimodal tasks, as it mitigates the suppression of visual tokens by language priors. Experiments on InternVL models demonstrate that ACPO effectively reverses the chosen-reward degradation of standard DPO. By halting Visual Anchor Collapse, ACPO generally outperforms baselines on hallucination benchmarks (HallusionBench, MM-IFEval) and general leaderboards (MMBench, MMStar, OCRBenchV2) while driving concurrent improvements in general capabilities.

Method: Proposes AdaptVision with coarse-to-fine visual token acquisition: starts with compressed low-resolution tokens, selectively invokes bounding box tool to crop key regions when needed. Uses reinforcement learning with Decoupled Turn Policy Optimization (DTPO) that separates tool learning (correct tool usage) and accuracy improvement (response refinement) objectives.

Result: AdaptVision achieves superior performance on multiple VQA benchmarks while consuming substantially fewer visual tokens than state-of-the-art efficient VLM methods.

Conclusion: The proposed active vision approach enables adaptive visual token acquisition, balancing accuracy and efficiency through decoupled reinforcement learning optimization.

Abstract: Vision-Language Models (VLMs) have achieved remarkable success in visual question answering tasks, but their reliance on large numbers of visual tokens introduces significant computational overhead. While existing efficient VLM approaches reduce visual tokens through fixed-ratio compression, they operate passively and lack the ability to adapt to varying task requirements. This motivates a fundamental question: Can VLMs autonomously determine the minimum number of visual tokens required for each sample? Inspired by human active vision mechanisms, we introduce AdaptVision, an efficient VLM paradigm that enables adaptive visual token acquisition through a coarse-to-fine approach. Our model initially processes compressed visual tokens from low-resolution images and selectively acquires additional visual information by invoking a bounding box tool to crop key regions when necessary. We train AdaptVision using a reinforcement learning framework that carefully balances accuracy and efficiency. Central to our approach is Decoupled Turn Policy Optimization (DTPO), which decouples the learning objective into two components: (1) tool learning, which optimizes correct tool utilization, and (2) accuracy improvement, which refines the generated responses to improve answer correctness. Based on this formulation, we further decouple advantage estimation by computing separate advantages for tokens associated with each objective. This formulation enables more effective optimization for AdaptVision compared to vanilla GRPO. Comprehensive experiments across multiple VQA benchmarks demonstrate that AdaptVision achieves superior performance while consuming substantially fewer visual tokens than state-of-the-art efficient VLM methods.

Method: Proposed Local Pattern SSAT (L-SSAT) framework combining primary face analysis tasks with SSL auxiliary task of Masked Auto-Encoder (MAE) for texture feature reconstruction. Conducted comprehensive benchmarking across multiple backbone architectures (shallow to deep) and evaluated on three face analysis tasks: face attribute prediction, emotion classification, and deepfake detection.

Result: Achieved average accuracies of 0.94 on FaceForensics++ (deepfake detection), 0.87 on CelebA (face attribute prediction), and 0.88 on AffectNet (emotion classification). Found that backbone effectiveness is highly task-dependent with no single unified backbone performing best across all face analysis paradigms.

Conclusion: Backbone selection for SSL-based face analysis is highly dependent on the specific downstream task, with no single generalized backbone performing optimally across different face analysis paradigms. The study provides guidance for backbone selection based on target application.

Abstract: In this work, we benchmark with different backbones and study their impact for self-supervised learning (SSL) as an auxiliary task to blend texture-based local descriptors into feature modelling for efficient face analysis. It is established in previous work that combining a primary task and a self-supervised auxiliary task enables more robust and discriminative representation learning. We employed different shallow to deep backbones for the SSL task of Masked Auto-Encoder (MAE) as an auxiliary objective to reconstruct texture features such as local patterns alongside the primary task in local pattern SSAT (L-SSAT), ensuring robust and unbiased face analysis. To expand the benchmark, we conducted a comprehensive comparative analysis across multiple model configurations within the proposed framework. To this end, we address the three research questions: “What is the role of the backbone in performance L-SSAT?”, “What type of backbone is effective for different face analysis tasks?”, and “Is there any generalized backbone for effective face analysis with L-SSAT?”. Towards answering these questions, we provide a detailed study and experiments. The performance evaluation demonstrates that the backbone for the proposed method is highly dependent on the downstream task, achieving average accuracies of 0.94 on FaceForensics++, 0.87 on CelebA, and 0.88 on AffectNet. For consistency of feature representation quality and generalisation capability across various face analysis paradigms, including face attribute prediction, emotion classification, and deepfake detection, there is no unified backbone.

Method: PAM (Pose-Appearance-Motion Engine) is a unified framework for controllable HOI video generation. It combines depth, segmentation, and keypoints as input conditions to generate high-resolution videos. The method validates performance through comprehensive experiments and ablation studies.

Result: On DexYCB: FVD of 29.13 (vs. 38.83 for InterDyn), MPJPE of 19.37 mm (vs. 30.05 mm for CosHand), generating 480x720 videos vs. 256x256/256x384 baselines. On OAKINK2: full multi-condition model improves FVD from 68.76 to 46.31. Ablation shows combining depth, segmentation, and keypoints yields best results. Synthetic data augmentation with 3,400 videos enables model trained on 50% real data to match 100% real baseline.

Conclusion: PAM successfully unifies pose, appearance, and motion for HOI video generation, achieving state-of-the-art performance on benchmarks and demonstrating practical value through effective synthetic data augmentation for downstream tasks like hand pose estimation.

Abstract: Hand-object interaction (HOI) reconstruction and synthesis are becoming central to embodied AI and AR/VR. Yet, despite rapid progress, existing HOI generation research remains fragmented across three disjoint tracks: (1) pose-only synthesis that predicts MANO trajectories without producing pixels; (2) single-image HOI generation that hallucinates appearance from masks or 2D cues but lacks dynamics; and (3) video generation methods that require both the entire pose sequence and the ground-truth first frame as inputs, preventing true sim-to-real deployment. Inspired by the philosophy of Joo et al. (2018), we think that HOI generation requires a unified engine that brings together pose, appearance, and motion within one coherent framework. Thus we introduce PAM: a Pose-Appearance-Motion Engine for controllable HOI video generation. The performance of our engine is validated by: (1) On DexYCB, we obtain an FVD of 29.13 (vs. 38.83 for InterDyn), and MPJPE of 19.37 mm (vs. 30.05 mm for CosHand), while generating higher-resolution 480x720 videos compared to 256x256 and 256x384 baselines. (2) On OAKINK2, our full multi-condition model improves FVD from 68.76 to 46.31. (3) An ablation over input conditions on DexYCB shows that combining depth, segmentation, and keypoints consistently yields the best results. (4) For a downstream hand pose estimation task using SimpleHand, augmenting training with 3,400 synthetic videos (207k frames) allows a model trained on only 50% of the real data plus our synthetic data to match the 100% real baseline.

Method: MAMMOTH uses a parameter-efficient, multi-head mixture of experts module that applies low-rank transformations tailored to each patch’s phenotype, compatible with any existing MIL aggregation method.

Result: Across 8 MIL methods and 19 classification tasks, MAMMOTH improved performance in 130 of 152 configurations, with average +3.8% performance gain, showing task-specific transformation has larger impact than aggregation method choice.

Conclusion: The linear transformation layer in MIL frameworks is a significant performance bottleneck, and MAMMOTH’s task-specific transformations substantially improve whole-slide image classification across diverse methods and tasks.

Abstract: Multiple Instance Learning (MIL) is the predominant framework for classifying gigapixel whole-slide images in computational pathology. MIL follows a sequence of 1) extracting patch features, 2) applying a linear layer to obtain task-specific patch features, and 3) aggregating the patches into a slide feature for classification. While substantial efforts have been devoted to optimizing patch feature extraction and aggregation, none have yet addressed the second point, the critical layer which transforms general-purpose features into task-specific features. We hypothesize that this layer constitutes an overlooked performance bottleneck and that stronger representations can be achieved with a low-rank transformation tailored to each patch’s phenotype, yielding synergistic effects with any of the existing MIL approaches. To this end, we introduce MAMMOTH, a parameter-efficient, multi-head mixture of experts module designed to improve the performance of any MIL model with minimal alterations to the total number of parameters. Across eight MIL methods and 19 different classification tasks, we find that such task-specific transformation has a larger effect on performance than the choice of aggregation method. For instance, when equipped with MAMMOTH, even simple methods such as max or mean pooling attain higher average performance than any method with the standard linear layer. Overall, MAMMOTH improves performance in 130 of the 152 examined configurations, with an average $+3.8%$ change in performance. Code is available at https://github.com/mahmoodlab/mammoth.

Method: Proposes Omni-WorldBench with two components: 1) Omni-WorldSuite - systematic prompt suite spanning diverse interaction levels and scene types; 2) Omni-Metrics - agent-based evaluation framework quantifying causal impact of interaction actions on both final outcomes and intermediate state evolution trajectories.

Result: Extensive evaluation of 18 representative world models reveals critical limitations in interactive response capabilities. The benchmark provides actionable insights for future research in interactive 4D world modeling.

Conclusion: Omni-WorldBench addresses a critical gap in evaluating interactive 4D world models and will be publicly released to foster progress in this important research direction.

Abstract: Video–based world models have emerged along two dominant paradigms: video generation and 3D reconstruction. However, existing evaluation benchmarks either focus narrowly on visual fidelity and text–video alignment for generative models, or rely on static 3D reconstruction metrics that fundamentally neglect temporal dynamics. We argue that the future of world modeling lies in 4D generation, which jointly models spatial structure and temporal evolution. In this paradigm, the core capability is interactive response: the ability to faithfully reflect how interaction actions drive state transitions across space and time. Yet no existing benchmark systematically evaluates this critical dimension. To address this gap, we propose Omni–WorldBench, a comprehensive benchmark specifically designed to evaluate the interactive response capabilities of world models in 4D settings. Omni–WorldBench comprises two key components: Omni–WorldSuite, a systematic prompt suite spanning diverse interaction levels and scene types; and Omni–Metrics, an agent-based evaluation framework that quantifies world modeling capabilities by measuring the causal impact of interaction actions on both final outcomes and intermediate state evolution trajectories. We conduct extensive evaluations of 18 representative world models across multiple paradigms. Our analysis reveals critical limitations of current world models in interactive response, providing actionable insights for future research. Omni-WorldBench will be publicly released to foster progress in interactive 4D world modeling.

Method: Experimental benchmark comparing four representative deep learning models on a large-scale aerial LiDAR dataset from Navarre, Spain, covering urban, rural, and industrial landscapes. Evaluated across five semantic classes (ground, vegetation, buildings, vehicles, etc.) with focus on class imbalance and geometric variability challenges.

Result: All models achieved high overall accuracy (>93%). KPConv attained highest mean IoU (78.51%) with consistent performance across classes. Point Transformer V3 showed superior performance on underrepresented vehicle class (75.11% IoU). Superpoint Transformer and RandLA-Net traded segmentation robustness for computational efficiency.

Conclusion: The study provides valuable insights into 3D semantic segmentation performance on aerial LiDAR data, highlighting KPConv’s overall superiority and Point Transformer V3’s strength on challenging underrepresented classes, while revealing trade-offs between segmentation quality and computational efficiency.

Abstract: Recent advances in deep learning have significantly improved 3D semantic segmentation, but most models focus on indoor or terrestrial datasets. Their behavior under real aerial acquisition conditions remains insufficiently explored, and although a few studies have addressed similar scenarios, they differ in dataset design, acquisition conditions, and model selection. To address this gap, we conduct an experimental benchmark evaluating several state-of-the-art architectures on a large-scale aerial LiDAR dataset acquired under operational flight conditions in Navarre, Spain, covering heterogeneous urban, rural, and industrial landscapes. This study compares four representative deep learning models, including KPConv, RandLA-Net, Superpoint Transformer, and Point Transformer V3, across five semantic classes commonly found in airborne surveys, such as ground, vegetation, buildings, and vehicles, highlighting the inherent challenges of class imbalance and geometric variability in aerial data. Results show that all tested models achieve high overall accuracy exceeding 93%, with KPConv attaining the highest mean IoU (78.51%) through consistent performance across classes, particularly on challenging and underrepresented categories. Point Transformer V3 demonstrates superior performance on the underrepresented vehicle class (75.11% IoU), while Superpoint Transformer and RandLA-Net trade off segmentation robustness for computational efficiency.

Method: Used Point Transformer v2 (PTv2) architecture for semantic segmentation of 3-channel LiDAR point clouds (geometry + spectral features). Trained on Oulanka river data with geometry and spectral features, conducted ablation studies on feature importance, and investigated multi-dataset training with additional sparsely annotated river data to improve generalization.

Result: Full-feature configuration achieved mIoU of 0.950, significantly outperforming geometry-only baseline. Intensity and reflectance features were identified as key for accurate land cover mapping. Multi-dataset training showed improved generalization performance despite limited high-quality annotations.

Conclusion: Transformer-based architectures show strong potential for multispectral point cloud analysis in riverine environments, offering new capabilities for sediment transport monitoring and river management applications. Multi-dataset training can improve model robustness with limited annotated data.

Abstract: Accurate land cover mapping in riverine environments is essential for effective river management, ecological understanding, and geomorphic change monitoring. This study explores the use of Point Transformer v2 (PTv2), an advanced deep neural network architecture designed for point cloud data, for land cover mapping through semantic segmentation of multispectral LiDAR data in real-world riverine environments. We utilize the geometric and spectral information from the 3-channel LiDAR point cloud to map land cover classes, including sand, gravel, low vegetation, high vegetation, forest floor, and water. The PTv2 model was trained and evaluated on point cloud data from the Oulanka river in northern Finland using both geometry and spectral features. To improve the model’s generalization in new riverine environments, we additionally investigate multi-dataset training that adds sparsely annotated data from an additional river dataset. Results demonstrated that using the full-feature configuration resulted in performance with a mean Intersection over Union (mIoU) of 0.950, significantly outperforming the geometry baseline. Other ablation studies revealed that intensity and reflectance features were the key for accurate land cover mapping. The multi-dataset training experiment showed improved generalization performance, suggesting potential for developing more robust models despite limited high-quality annotated data. Our work demonstrates the potential of applying transformer-based architectures to multispectral point clouds in riverine environments. The approach offers new capabilities for monitoring sediment transport and other river management applications.

Method: Introduced EgoGroups dataset with first-person views from 65 countries covering various crowd densities and weather/time conditions, with dense human annotations for persons and social groups plus geographic/scene metadata. Evaluated state-of-the-art VLM/LLMs and supervised models on group detection capabilities.

Result: VLMs and LLMs can outperform supervised baselines in zero-shot settings, while crowd density and cultural regions significantly influence model performance.

Conclusion: EgoGroups addresses limitations of existing benchmarks and reveals important insights about VLM/LLM capabilities for social group detection in diverse real-world contexts.

Abstract: Social group detection, or the identification of humans involved in reciprocal interpersonal interactions (e.g., family members, friends, and customers and merchants), is a crucial component of social intelligence needed for agents transacting in the world. The few existing benchmarks for social group detection are limited by low scene diversity and reliance on third-person camera sources (e.g., surveillance footage). Consequently, these benchmarks generally lack real-world evaluation on how groups form and evolve in diverse cultural contexts and unconstrained settings. To address this gap, we introduce EgoGroups, a first-person view dataset that captures social dynamics in cities around the world. EgoGroups spans 65 countries covering low, medium, and high-crowd settings under four weather/time-of-day conditions. We include dense human annotations for person and social groups, along with rich geographic and scene metadata. Using this dataset, we performed an extensive evaluation of state-of-the-art VLM/LLMs and supervised models on their group detection capabilities. We found several interesting findings, including VLMs and LLMs can outperform supervised baselines in a zero-shot setting, while crowd density and cultural regions clearly influence model performance.

Method: Leverages pre-trained depth estimation network to generate pseudo optical flows, which serve as conditioning inputs for a next-frame generation model trained to produce high-fidelity, pixel-aligned subsequent frames. Also proposes inconsistent pixel filtering strategy to identify and remove unreliable pixels in generated frames.

Result: Extensive experiments on KITTI2012, KITTI2015, and Sintel demonstrate competitive or superior results compared to existing unsupervised and semi-supervised approaches.

Conclusion: FlowGen presents a scalable and annotation-free solution for optical flow learning that achieves state-of-the-art performance without human supervision.

Abstract: Optical flow estimation is a fundamental problem in computer vision, yet the reliance on expensive ground-truth annotations limits the scalability of supervised approaches. Although unsupervised and semi-supervised methods alleviate this issue, they often suffer from unreliable supervision signals based on brightness constancy and smoothness assumptions, leading to inaccurate motion estimation in complex real-world scenarios. To overcome these limitations, we introduce \textbf{\modelname}, a novel framework that synthesizes large-scale, perfectly aligned frame–flow data pairs for supervised optical flow training without human annotations. Specifically, our method leverages a pre-trained depth estimation network to generate pseudo optical flows, which serve as conditioning inputs for a next-frame generation model trained to produce high-fidelity, pixel-aligned subsequent frames. This process enables the creation of abundant, high-quality synthetic data with precise motion correspondence. Furthermore, we propose an \textit{inconsistent pixel filtering} strategy that identifies and removes unreliable pixels in generated frames, effectively enhancing fine-tuning performance on real-world datasets. Extensive experiments on KITTI2012, KITTI2015, and Sintel demonstrate that \textbf{\modelname} achieves competitive or superior results compared to existing unsupervised and semi-supervised approaches, highlighting its potential as a scalable and annotation-free solution for optical flow learning. We will release our code upon acceptance.

Method: Three-stage framework: 1) Progressive Guided Distillation Initialization for stabilization, 2) Dual-Stream Distillation combining DMD and RFS-GAN streams for complementary supervision, 3) Preference-Guided Refinement for perceptual quality alignment.

Result: DUO-VSR achieves superior visual quality and efficiency over previous one-step VSR approaches, demonstrating effective acceleration of diffusion models for video tasks.

Conclusion: The proposed dual-stream distillation framework successfully addresses training instability and supervision insufficiency in accelerating diffusion-based VSR, enabling high-quality one-step generation.

Abstract: Diffusion-based video super-resolution (VSR) has recently achieved remarkable fidelity but still suffers from prohibitive sampling costs. While distribution matching distillation (DMD) can accelerate diffusion models toward one-step generation, directly applying it to VSR often results in training instability alongside degraded and insufficient supervision. To address these issues, we propose DUO-VSR, a three-stage framework built upon a Dual-Stream Distillation strategy that unifies distribution matching and adversarial supervision for one-step VSR. Firstly, a Progressive Guided Distillation Initialization is employed to stabilize subsequent training through trajectory-preserving distillation. Next, the Dual-Stream Distillation jointly optimizes the DMD and Real-Fake Score Feature GAN (RFS-GAN) streams, with the latter providing complementary adversarial supervision leveraging discriminative features from both real and fake score models. Finally, a Preference-Guided Refinement stage further aligns the student with perceptual quality preferences. Extensive experiments demonstrate that DUO-VSR achieves superior visual quality and efficiency over previous one-step VSR approaches.

Method: Cannot determine method as paper content is unavailable.

Result: Cannot determine results as paper content is unavailable.

Conclusion: Cannot draw conclusions about paper content due to access limitations.

Abstract: Failed to fetch summary for 2601.10744: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.10744&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes Geometric Latent Diffusion (GLD) that repurposes geometrically consistent feature spaces from geometric foundation models as the latent space for multi-view diffusion models.

Result: GLD outperforms VAE and RAE on 2D image quality and 3D consistency metrics, accelerates training by 4.4x compared to VAE, and remains competitive with state-of-the-art methods despite training from scratch without text-to-image pretraining.

Conclusion: Geometric foundation model features provide an effective latent space for novel view synthesis, enabling geometrically consistent multi-view generation with improved quality and efficiency.

Abstract: While recent advances in generative latent spaces have driven substantial progress in single-image generation, the optimal latent space for novel view synthesis (NVS) remains largely unexplored. In particular, NVS requires geometrically consistent generation across viewpoints, but existing approaches typically operate in a view-independent VAE latent space. In this paper, we propose Geometric Latent Diffusion (GLD), a framework that repurposes the geometrically consistent feature space of geometric foundation models as the latent space for multi-view diffusion. We show that these features not only support high-fidelity RGB reconstruction but also encode strong cross-view geometric correspondences, providing a well-suited latent space for NVS. Our experiments demonstrate that GLD outperforms both VAE and RAE on 2D image quality and 3D consistency metrics, while accelerating training by more than 4.4x compared to the VAE latent space. Notably, GLD remains competitive with state-of-the-art methods that leverage large-scale text-to-image pretraining, despite training its diffusion model from scratch without such generative pretraining.

Method: Analyzed where and how spatial associations are computed within VLMs by examining two concurrent mechanisms: (1) language model backbone layers representing content-independent spatial relations on visual tokens, and (2) vision encoder representations encoding object layouts globally across visual tokens.

Result: Found that while language model layers play a secondary role, the dominant spatial information originates from vision encoders whose representations encode object layouts and extend beyond object regions into surrounding background areas. Enhancing these vision-derived spatial representations globally improves spatial reasoning performance.

Conclusion: Spatial association in VLMs is computed through two mechanisms, with vision encoders playing the central role in enabling spatial reasoning by providing globally distributed spatial signals across all image tokens.

Abstract: Many multimodal tasks, such as image captioning and visual question answering, require vision-language models (VLMs) to associate objects with their properties and spatial relations. Yet it remains unclear where and how such associations are computed within VLMs. In this work, we show that VLMs rely on two concurrent mechanisms to represent such associations. In the language model backbone, intermediate layers represent content-independent spatial relations on top of visual tokens corresponding to objects. However, this mechanism plays only a secondary role in shaping model predictions. Instead, the dominant source of spatial information originates in the vision encoder, whose representations encode the layout of objects and are directly exploited by the language model backbone. Notably, this spatial signal is distributed globally across visual tokens, extending beyond object regions into surrounding background areas. We show that enhancing these vision-derived spatial representations globally across all image tokens improves spatial reasoning performance on naturalistic images. Together, our results clarify how spatial association is computed within VLMs and highlight the central role of vision encoders in enabling spatial reasoning.

Method: Proposes DualCoT-VLA with parallel reasoning mechanism: integrates visual CoT for low-level spatial understanding and linguistic CoT for high-level task planning. Introduces parallel CoT mechanism with two sets of learnable query tokens, shifting from autoregressive step-by-step reasoning to single-step forward reasoning to reduce latency.

Result: Achieves state-of-the-art performance on LIBERO and RoboCasa GR1 benchmarks, and demonstrates effectiveness on real-world platforms.

Conclusion: DualCoT-VLA successfully addresses limitations of current CoT-based VLA models by enabling comprehensive multi-modal reasoning while significantly reducing inference latency through parallel processing.

Abstract: Vision-Language-Action (VLA) models map visual observations and language instructions directly to robotic actions. While effective for simple tasks, standard VLA models often struggle with complex, multi-step tasks requiring logical planning, as well as precise manipulations demanding fine-grained spatial perception. Recent efforts have incorporated Chain-of-Thought (CoT) reasoning to endow VLA models with a ``thinking before acting’’ capability. However, current CoT-based VLA models face two critical limitations: 1) an inability to simultaneously capture low-level visual details and high-level logical planning due to their reliance on isolated, single-modal CoT; 2) high inference latency with compounding errors caused by step-by-step autoregressive decoding. To address these limitations, we propose DualCoT-VLA, a visual-linguistic CoT method for VLA models with a parallel reasoning mechanism. To achieve comprehensive multi-modal reasoning, our method integrates a visual CoT for low-level spatial understanding and a linguistic CoT for high-level task planning. Furthermore, to overcome the latency bottleneck, we introduce a parallel CoT mechanism that incorporates two sets of learnable query tokens, shifting autoregressive reasoning to single-step forward reasoning. Extensive experiments demonstrate that our DualCoT-VLA achieves state-of-the-art performance on the LIBERO and RoboCasa GR1 benchmarks, as well as in real-world platforms.

Method: Divides videos into segments and creates visual-temporal affinity graphs based on visual similarity and temporal proximity. Uses a Hypothesis-Verification-Refinement loop to estimate relevance scores of observed segments and propagate them to unseen segments, producing a global relevance distribution to guide localization of critical segments.

Result: Achieves substantial gains across mainstream MLLMs on representative benchmarks, with accuracy improvements up to 7.5% on VideoMME-long. Consistently outperforms existing methods.

Conclusion: VideoDetective effectively addresses long-video understanding challenges by integrating query-to-segment relevance and inter-segment affinity, enabling efficient clue hunting and improved performance on long-video QA tasks.

Abstract: Long video understanding remains challenging for multimodal large language models (MLLMs) due to limited context windows, which necessitate identifying sparse query-relevant video segments. However, existing methods predominantly localize clues based solely on the query, overlooking the video’s intrinsic structure and varying relevance across segments. To address this, we propose VideoDetective, a framework that integrates query-to-segment relevance and inter-segment affinity for effective clue hunting in long-video question answering. Specifically, we divide a video into various segments and represent them as a visual-temporal affinity graph built from visual similarity and temporal proximity. We then perform a Hypothesis-Verification-Refinement loop to estimate relevance scores of observed segments to the query and propagate them to unseen segments, yielding a global relevance distribution that guides the localization of the most critical segments for final answering with sparse observation. Experiments show our method consistently achieves substantial gains across a wide range of mainstream MLLMs on representative benchmarks, with accuracy improvements of up to 7.5% on VideoMME-long. Our code is available at https://videodetective.github.io/

Method: Cannot analyze method as paper content is unavailable

Result: No results available due to technical error accessing paper

Conclusion: Cannot provide analysis due to API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2312.17251: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2312.17251&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2402.13876: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2402.13876&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval failure

Result: Unable to determine results due to retrieval failure

Conclusion: Unable to draw conclusions due to retrieval failure

Abstract: Failed to fetch summary for 2404.12352: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2404.12352&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2405.13859: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.13859&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2406.03017: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.03017&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2411.10495: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.10495&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot determine conclusion as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2411.16196: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.16196&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2411.17292: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.17292&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests)

Result: No results available - paper content could not be retrieved

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper abstract

Abstract: Failed to fetch summary for 2411.18064: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.18064&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2411.19093: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.19093&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2412.01583: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.01583&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2502.00618: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.00618&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2502.09125: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.09125&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2502.18220: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.18220&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2501.00725: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.00725&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2503.13074: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.13074&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access limitations preventing paper retrieval

Result: Unable to determine results due to API access limitations preventing paper retrieval

Conclusion: Unable to draw conclusions about the paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2503.13617: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.13617&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2503.19740: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.19740&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2504.01396: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.01396&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2504.02519: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.02519&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2504.21247: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.21247&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable.

Result: Cannot determine results as paper content is unavailable.

Conclusion: Cannot draw conclusions about the paper without access to its content.

Abstract: Failed to fetch summary for 2505.12200: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.12200&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2505.12656: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.12656&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable

Result: Cannot determine results as abstract is unavailable

Conclusion: Cannot draw conclusions as abstract is unavailable

Abstract: Failed to fetch summary for 2506.13925: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.13925&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval failure

Result: Unable to determine results due to retrieval failure

Conclusion: Unable to determine conclusion due to retrieval failure

Abstract: Failed to fetch summary for 2505.16474: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.16474&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2505.17782: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17782&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to analyze paper due to technical error fetching content

Abstract: Failed to fetch summary for 2505.22701: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.22701&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting.

Result: Cannot determine results as paper content is unavailable due to API rate limiting.

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting.

Abstract: Failed to fetch summary for 2506.02459: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.02459&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2506.02845: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.02845&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2506.04559: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.04559&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2506.09935: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.09935&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.15130: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.15130&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to access limitations

Abstract: Failed to fetch summary for 2506.20294: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.20294&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2506.21076: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.21076&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2509.18801: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.18801&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2507.00462: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.00462&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2507.11474: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.11474&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content

Result: Cannot determine results without access to the paper content

Conclusion: Cannot determine conclusion without access to the paper content

Abstract: Failed to fetch summary for 2507.15085: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.15085&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2510.02284: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02284&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract content is unavailable

Result: Cannot determine results as abstract content is unavailable

Conclusion: Cannot draw conclusions about paper content due to data retrieval failure

Abstract: Failed to fetch summary for 2508.20066: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.20066&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2509.04597: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.04597&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2510.06638: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.06638&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2509.12544: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.12544&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract content is unavailable

Result: Cannot determine results as abstract content is unavailable

Conclusion: Cannot draw conclusions about paper content due to data retrieval failure

Abstract: Failed to fetch summary for 2510.12060: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.12060&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2509.23774: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23774&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2510.13232: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.13232&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved due to API rate limiting.

Result: Unable to determine results as the paper content could not be retrieved due to API rate limiting.

Conclusion: Unable to determine conclusion as the paper content could not be retrieved due to API rate limiting.

Abstract: Failed to fetch summary for 2509.24817: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24817&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2510.01049: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01049&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2510.01641: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01641&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2510.08771: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.08771&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2510.11026: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.11026&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2510.21271: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.21271&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2510.17568: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.17568&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without paper content. The arXiv ID format suggests this is a recent paper from October 2024.

Result: No results available due to access limitations. The paper may be relevant to multimodal LLMs but cannot be assessed without content.

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper content. Need to wait for rate limits to reset or use alternative access methods.

Abstract: Failed to fetch summary for 2510.17699: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.17699&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.01571: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.01571&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.17805: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17805&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.05421: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.05421&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2511.20011: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.20011&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to technical access issues

Conclusion: Paper analysis impossible due to API rate limiting preventing content retrieval

Abstract: Failed to fetch summary for 2511.11236: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.11236&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2511.22169: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22169&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.13945: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.13945&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.00065: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00065&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2511.15700: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.15700&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.16993: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.16993&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2512.11438: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.11438&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusion due to missing paper content

Abstract: Failed to fetch summary for 2511.17487: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17487&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to analyze paper due to technical fetching error

Abstract: Failed to fetch summary for 2512.13157: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13157&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2511.19172: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.19172&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error fetching paper content

Result: Unable to determine results due to technical error fetching paper content

Conclusion: Unable to determine conclusion due to technical error fetching paper content

Abstract: Failed to fetch summary for 2512.16523: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.16523&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.19235: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.19235&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2512.16975: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.16975&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.20279: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.20279&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content.

Result: Cannot determine results due to inability to access paper content.

Conclusion: Cannot draw conclusions due to inability to access paper content.

Abstract: Failed to fetch summary for 2511.21565: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21565&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2512.00385: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00385&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.01547: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.01547&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2512.00422: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00422&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2512.01495: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.01495&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.05848: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.05848&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.01755: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.01755&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2512.04619: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.04619&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2601.16296: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.16296&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available

Result: No results available

Conclusion: Cannot draw conclusions without paper content

Abstract: Failed to fetch summary for 2512.04821: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.04821&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API rate limiting preventing access to paper content

Result: Unable to determine results due to API rate limiting preventing access to paper content

Conclusion: Unable to determine conclusion due to API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2601.22725: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.22725&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.05905: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.05905&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2512.07698: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.07698&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2512.08441: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08441&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2512.09278: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.09278&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2512.10571: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.10571&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Attempted to use arXiv API query with specific paper ID (2602.14408) but received HTTP 429 status indicating too many requests

Result: Failed to fetch paper summary; no content available for analysis due to API rate limiting

Conclusion: Cannot analyze the paper as the content could not be retrieved; need to wait for rate limits to reset or use alternative methods to access the paper

Abstract: Failed to fetch summary for 2602.14408: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.14408&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2512.13285: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13285&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to the paper’s abstract

Abstract: Failed to fetch summary for 2602.19248: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19248&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2512.17495: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.17495&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper abstract

Result: Cannot determine results without access to paper abstract

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2512.21284: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.21284&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.21778: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.21778&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval error

Result: Unable to determine results due to retrieval error

Conclusion: Unable to determine conclusion due to retrieval error

Abstract: Failed to fetch summary for 2603.00431: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00431&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2601.02763: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.02763&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract

Result: Unable to determine results due to missing abstract

Conclusion: Unable to determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2603.03710: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03710&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2601.04090: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04090&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.04453: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04453&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.10935: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10935&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2601.05175: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.05175&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2601.05237: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.05237&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in accessing paper content

Result: Unable to determine results due to technical error in accessing paper content

Conclusion: Unable to draw conclusions due to technical error in accessing paper content

Abstract: Failed to fetch summary for 2601.11930: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.11930&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available since the paper content could not be retrieved.

Result: No results available as the paper content could not be accessed.

Conclusion: Unable to draw any conclusions about the paper due to lack of accessible content.

Abstract: Failed to fetch summary for 2603.13406: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.13406&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.12193: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.12193&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to access limitations

Result: Cannot determine results due to access limitations

Conclusion: Cannot draw conclusions due to access limitations

Abstract: Failed to fetch summary for 2603.13432: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.13432&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.17468: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17468&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2601.17657: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17657&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without paper content

Result: Cannot determine results without paper content

Conclusion: Cannot determine conclusion without paper content

Abstract: Failed to fetch summary for 2601.21998: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.21998&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request.

Result: No results available due to failed data retrieval.

Conclusion: Unable to analyze paper content due to technical limitations in accessing the abstract.

Abstract: Failed to fetch summary for 2602.02223: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02223&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.07446: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.07446&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2603.16666: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16666&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.13091: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13091&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to abstract fetch failure

Result: Cannot determine results due to abstract fetch failure

Conclusion: Cannot determine conclusion due to abstract fetch failure

Abstract: Failed to fetch summary for 2602.20880: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20880&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to HTTP 429 error preventing access to paper details.

Result: No results available as the paper summary could not be retrieved from arXiv API.

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper content.

Abstract: Failed to fetch summary for 2603.17655: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17655&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2602.21499: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21499&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper fetch

Result: Cannot determine results due to failed paper fetch

Conclusion: Cannot determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2602.23306: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23306&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot determine conclusion as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2602.23783: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23783&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.23956: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23956&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without paper content

Result: Cannot determine results without paper content

Conclusion: Cannot determine conclusion without paper content

Abstract: Failed to fetch summary for 2603.00919: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00919&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to analyze paper due to technical access issues

Abstract: Failed to fetch summary for 2603.01164: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01164&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2603.01498: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01498&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2603.03447: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03447&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.03744: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03744&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2603.04839: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.04839&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2603.04846: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.04846&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2603.06917: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.06917&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.06989: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.06989&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.08536: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08536&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2603.09418: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.09418&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2603.10446: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10446&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content inaccessible

Result: No results available - failed to fetch paper summary

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2603.10568: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10568&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.10748: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10748&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.12055: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12055&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2603.12469: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12469&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2603.12760: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12760&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot determine conclusion without access to the paper abstract

Abstract: Failed to fetch summary for 2603.12918: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12918&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable due to rate limiting error

Result: Cannot determine results as abstract is unavailable due to rate limiting error

Conclusion: Cannot draw conclusions about the paper due to inability to access the abstract content

Abstract: Failed to fetch summary for 2603.14188: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.14188&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2603.15919: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15919&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.16271: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16271&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown due to API access restrictions. The paper ID format suggests it’s a computer science/machine learning paper from March 2023.

Result: No results available for analysis. The HTTP 429 error indicates too many requests to arXiv API.

Conclusion: Cannot provide conclusion without paper content. Need to wait for API rate limits to reset or use alternative methods to access paper details.

Abstract: Failed to fetch summary for 2603.16451: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16451&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.17240: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17240&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.17508: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17508&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved

Result: Unable to determine results as the paper content could not be retrieved

Conclusion: Unable to determine conclusion as the paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.17779: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17779&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.17889: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17889&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.19610: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.19610&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.19964: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.19964&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to draw conclusions due to access error

Abstract: Failed to fetch summary for 2404.12339: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2404.12339&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2511.08935: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08935&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2512.00021: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00021&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request

Result: No results available as paper content could not be accessed

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2512.12984: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.12984&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.19402: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.19402&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.05162: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.05162&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request

Result: No results available as the paper content could not be accessed

Conclusion: Cannot provide conclusion due to inability to fetch paper content from arXiv

Abstract: Failed to fetch summary for 2601.07242: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.07242&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes AgenticGEO framework with MAP-Elites archive to evolve diverse compositional strategies, and a Co-Evolving Critic as a lightweight surrogate to approximate engine feedback for content-specific strategy selection and refinement, guiding both evolutionary search and inference-time planning.

Result: Achieves state-of-the-art performance and demonstrates robust transferability, outperforming 14 baselines across 3 datasets in both in-domain and cross-domain experiments on two representative generative engines.

Conclusion: AgenticGEO provides an effective self-evolving framework for Generative Engine Optimization that can robustly adapt to unpredictable black-box engine behaviors while reducing interaction costs.

Abstract: Generative search engines represent a transition from traditional ranking-based retrieval to Large Language Model (LLM)-based synthesis, transforming optimization goals from ranking prominence towards content inclusion. Generative Engine Optimization (GEO), specifically, aims to maximize visibility and attribution in black-box summarized outputs by strategically manipulating source content. However, existing methods rely on static heuristics, single-prompt optimization, or engine preference rule distillation that is prone to overfitting. They cannot flexibly adapt to diverse content or the changing behaviors of generative engines. Moreover, effectively optimizing these strategies requires an impractical amount of interaction feedback from the engines. To address these challenges, we propose AgenticGEO, a self-evolving agentic framework formulating optimization as a content-conditioned control problem, which enhances intrinsic content quality to robustly adapt to the unpredictable behaviors of black-box engines. Unlike fixed-strategy methods, AgenticGEO employs a MAP-Elites archive to evolve diverse, compositional strategies. To mitigate interaction costs, we introduce a Co-Evolving Critic, a lightweight surrogate that approximates engine feedback for content-specific strategy selection and refinement, efficiently guiding both evolutionary search and inference-time planning. Through extensive in-domain and cross-domain experiments on two representative engines, AgenticGEO achieves state-of-the-art performance and demonstrates robust transferability, outperforming 14 baselines across 3 datasets. Our code and model are available at: https://github.com/AIcling/agentic_geo.

Method: PROMAS uses Markov transitions for predictive error analysis. It extracts Causal Delta Features to capture semantic displacement, maps them to a quantized Vector Markov Space to model reasoning as probabilistic transitions, and integrates a Proactive Prediction Head with Jump Detection to localize errors via risk acceleration rather than static thresholds.

Result: On the Who&When benchmark, PROMAS achieves 22.97% step-level accuracy while processing only 27% of reasoning logs. This performance rivals reactive monitors like MASC while reducing data overhead by 73%. The method trades some accuracy for significantly improved intervention latency.

Conclusion: PROMAS provides a proactive framework for error detection in multi-agent LLM systems that balances diagnostic precision with real-time intervention needs, offering a practical solution for preventing error propagation in collaborative reasoning systems.

Abstract: The integration of Large Language Models into Multi-Agent Systems (MAS) has enabled the so-lution of complex, long-horizon tasks through collaborative reasoning. However, this collec-tive intelligence is inherently fragile, as a single logical fallacy can rapidly propagate and lead to system-wide failure. Most current research re-lies on post-hoc failure analysis, thereby hinder-ing real-time intervention. To address this, we propose PROMAS, a proactive framework utiliz-ing Markov transitions for predictive error anal-ysis. PROMAS extracts Causal Delta Features to capture semantic displacement, mapping them to a quantized Vector Markov Space to model reasoning as probabilistic transitions. By inte-grating a Proactive Prediction Head with Jump Detection, the method localizes errors via risk acceleration rather than static thresholds. On the Who&When benchmark, PROMAS achieves 22.97% step-level accuracy while processing only 27% of reasoning logs. This performance rivals reactive monitors like MASC while reducing data overhead by 73%. Although this strategy entails an accuracy trade-off compared to post-hoc meth-ods, it significantly improves intervention latency, balancing diagnostic precision with the real-time demands of autonomous reasoning.

Method: DST is an adaptable, plug-and-play predictor that serves as a lightweight supervised heuristic to guide the ToT search process, enabling dynamic context-aware pruning that expands search beams only when encountering uncertainty or task complexity.

Result: The method achieves accuracy competitive with or superior to strong baselines including standard ToT while reducing computational overhead by 26-75% across mathematical reasoning, general reasoning, and complex logical reasoning benchmarks.

Conclusion: DST effectively resolves the accuracy-efficiency trade-off in tree-based reasoning, transforming ToT from a resource-intensive technique into a scalable and practical paradigm for complex problem-solving in LLMs.

Abstract: While Large Language Models (LLMs) have advanced complex reasoning, prominent methods like the Tree of Thoughts (ToT) framework face a critical trade-off between exploration depth and computational efficiency. Existing ToT implementations often rely on heavyweight LLM-based self-evaluation or rigid heuristics for branch pruning, making them prohibitively expensive and inflexible for broad application. To address this, we introduce DST, an adaptable, plug-and-play predictor that serves as a lightweight, supervised heuristic to guide the ToT search process. Our predictor enables dynamic, context-aware pruning, allowing the search to proceed with near-greedy efficiency on simpler reasoning steps while adaptively expanding the search beam only when encountering uncertainty or task complexity. We evaluate our approach on a diverse suite of benchmarks spanning mathematical reasoning, general reasoning, and complex logical reasoning. Experimental results demonstrate that our method achieves accuracy competitive with or superior to strong baselines, including standard ToT, while reducing computational overhead by 26-75%. Our work effectively resolves the accuracy-efficiency trade-off in tree-based reasoning, transforming ToT from a resource-intensive technique into a scalable and practical paradigm for complex problem-solving in LLMs.

Method: Combines factored POMDP decomposition (from FactorSim) to reduce context by modularizing simulation steps, with a hierarchical planner-designer-critic agentic workflow (inspired by SceneSmith) for iterative refinement at each generation step.

Result: Experiments on PyGame Learning Environment show improved prompt alignment, fewer runtime errors, and higher code quality compared to non-agentic factored baselines.

Conclusion: The combination of factored decomposition and agentic workflows effectively addresses LLM reasoning limitations in simulation generation, producing higher quality executable code from natural language specifications.

Abstract: Generating executable simulations from natural language specifications remains a challenging problem due to the limited reasoning capacity of large language models (LLMs) when confronted with large, interconnected codebases. This paper presents FactorSmith, a framework that synthesizes playable game simulations in code from textual descriptions by combining two complementary ideas: factored POMDP decomposition for principled context reduction and a hierarchical planner-designer-critic agentic workflow for iterative quality refinement at every generation step. Drawing on the factored partially observable Markov decision process (POMDP) representation introduced by FactorSim [Sun et al., 2024], the proposed method decomposes a simulation specification into modular steps where each step operates only on a minimal subset of relevant state variables, limiting the context window that any single LLM call must process. Inspired by the agentic trio architecture of SceneSmith [Pfaff et al., 2025], FactorSmith embeds within every factored step a three-agent interaction: a planner that orchestrates workflow, a designer that proposes code artifacts, and a critic that evaluates quality through structured scoring, enabling iterative refinement with checkpoint rollback. This paper formalizes the combined approach, presents the mathematical framework underpinning context selection and agentic refinement, and describes the open-source implementation. Experiments on the PyGame Learning Environment benchmark demonstrate that FactorSmith generates simulations with improved prompt alignment, fewer runtime errors, and higher code quality compared to non-agentic factored baselines.

Method: Proposed a formal taxonomy of introspection as latent computation over model policies/parameters. Created Introspect-Bench, a multifaceted evaluation suite for rigorous capability testing. Analyzed attention diffusion mechanisms.

Result: Frontier models show privileged access to their own policies, outperforming peer models in predicting their own behavior. Provided causal, mechanistic evidence showing how LLMs learn to introspect without explicit training via attention diffusion.

Conclusion: LLMs exhibit genuine introspective capabilities that go beyond mere application of general knowledge, with attention mechanisms playing a key role in enabling meta-cognitive processes.

Abstract: A hallmark of human intelligence is Introspection-the ability to assess and reason about one’s own cognitive processes. Introspection has emerged as a promising but contested capability in large language models (LLMs). However, current evaluations often fail to distinguish genuine meta-cognition from the mere application of general world knowledge or text-based self-simulation. In this work, we propose a principled taxonomy that formalizes introspection as the latent computation of specific operators over a model’s policy and parameters. To isolate the components of generalized introspection, we present Introspect-Bench, a multifaceted evaluation suite designed for rigorous capability testing. Our results show that frontier models exhibit privileged access to their own policies, outperforming peer models in predicting their own behavior. Furthermore, we provide causal, mechanistic evidence explaining both how LLMs learn to introspect without explicit training, and how the mechanism of introspection emerges via attention diffusion.

Method: Introduces AgentComm-Bench benchmark suite with evaluation protocol that systematically stress-tests cooperative embodied AI under six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence. Covers three task families: cooperative perception, multi-agent waypoint navigation, and cooperative zone search. Evaluates five communication strategies including a proposed lightweight method based on redundant message coding with staleness-aware fusion.

Result: Communication-dependent tasks degrade catastrophically: stale memory and bandwidth collapse cause over 96% performance drops in navigation, while content corruption reduces perception F1 by over 85%. Vulnerability depends on impairment type-task design interaction. Redundant message coding more than doubles navigation performance under 80% packet loss. Perception fusion is robust to packet loss but amplifies corrupted data.

Conclusion: AgentComm-Bench provides practical evaluation protocol for cooperative embodied AI under realistic communication conditions. Recommends that future work report performance under multiple impairment conditions to ensure robustness in real-world deployments.

Abstract: Cooperative multi-agent methods for embodied AI are almost universally evaluated under idealized communication: zero latency, no packet loss, and unlimited bandwidth. Real-world deployment on robots with wireless links, autonomous vehicles on congested networks, or drone swarms in contested spectrum offers no such guarantees. We introduce AgentComm-Bench, a benchmark suite and evaluation protocol that systematically stress-tests cooperative embodied AI under six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence. AgentComm-Bench spans three task families: cooperative perception, multi-agent waypoint navigation, and cooperative zone search, and evaluates five communication strategies, including a lightweight method we propose based on redundant message coding with staleness-aware fusion. Our experiments reveal that communication-dependent tasks degrade catastrophically: stale memory and bandwidth collapse cause over 96% performance drops in navigation, while content corruption (stale or conflicting data) reduces perception F1 by over 85%. Vulnerability depends on the interaction between impairment type and task design; perception fusion is robust to packet loss but amplifies corrupted data. Redundant message coding more than doubles navigation performance under 80% packet loss. We release AgentComm-Bench as a practical evaluation protocol and recommend that cooperative embodied AI work report performance under multiple impairment conditions.

Method: Integrates LLMs and energy-based contrastive learning. Uses LLMs to generate dependency-aware pseudo-OOD nodes by leveraging semantic understanding, then applies energy contrastive learning to distinguish IND/OOD nodes.

Result: Extensive experiments on six benchmark datasets show consistent outperformance over state-of-the-art baselines, achieving high classification accuracy and robust OOD detection capabilities.

Conclusion: LECT effectively addresses OOD detection in text-attributed graphs by combining LLMs for sample generation and energy contrastive learning for discrimination, demonstrating strong performance across multiple datasets.

Abstract: Text-attributed graphs, where nodes are enriched with textual attributes, have become a powerful tool for modeling real-world networks such as citation, social, and transaction networks. However, existing methods for learning from these graphs often assume that the distributions of training and testing data are consistent. This assumption leads to significant performance degradation when faced with out-of-distribution (OOD) data. In this paper, we address the challenge of node-level OOD detection in text-attributed graphs, with the goal of maintaining accurate node classification while simultaneously identifying OOD nodes. We propose a novel approach, LLM-Enhanced Energy Contrastive Learning for Out-of-Distribution Detection in Text-Attributed Graphs (LECT), which integrates large language models (LLMs) and energy-based contrastive learning. The proposed method involves generating high-quality OOD samples by leveraging the semantic understanding and contextual knowledge of LLMs to create dependency-aware pseudo-OOD nodes, and applying contrastive learning based on energy functions to distinguish between in-distribution (IND) and OOD nodes. The effectiveness of our method is demonstrated through extensive experiments on six benchmark datasets, where our method consistently outperforms state-of-the-art baselines, achieving both high classification accuracy and robust OOD detection capabilities.