Layer-Order Inversion: Rethinking Latent Multi-Hop Reasoning in Large Language Models

Large language models (LLMs) perform well on multi-hop reasoning, yet how they internally compose multiple facts remains unclear. Recent work proposes \emph{hop-aligned circuit hypothesis}, suggesting that bridge entities are computed sequentially across layers before later-hop answers. Through systematic analyses on real-world multi-hop queries, we show that this hop-aligned assumption does not generalize: later-hop answer entities can become decodable earlier than bridge entities, a phenomenon we call \emph{layer-order inversion}, which strengthens with total hops. To explain this behavior, we propose a \emph{probabilistic recall-and-extract} framework that models multi-hop reasoning as broad probabilistic recall in shallow MLP layers followed by selective extraction in deeper attention layers. This framework is empirically validated through systematic probing analyses, reinterpreting prior layer-wise decoding evidence, explaining chain-of-thought gains, and providing a mechanistic diagnosis of multi-hop failures despite correct single-hop knowledge. Code is available at https://github.com/laquabe/Layer-Order-Inversion.

MarineEval: Assessing the Marine Intelligence of Vision-Language Models

We have witnessed promising progress led by large language models (LLMs) and further vision language models (VLMs) in handling various queries as a general-purpose assistant. VLMs, as a bridge to connect the visual world and language corpus, receive both visual content and various text-only user instructions to generate corresponding responses. Though great success has been achieved by VLMs in various fields, in this work, we ask whether the existing VLMs can act as domain experts, accurately answering marine questions, which require significant domain expertise and address special domain challenges/requirements. To comprehensively evaluate the effectiveness and explore the boundary of existing VLMs, we construct the first large-scale marine VLM dataset and benchmark called MarineEval, with 2,000 image-based question-answering pairs. During our dataset construction, we ensure the diversity and coverage of the constructed data: 7 task dimensions and 20 capacity dimensions. The domain requirements are specially integrated into the data construction and further verified by the corresponding marine domain experts. We comprehensively benchmark 17 existing VLMs on our MarineEval and also investigate the limitations of existing models in answering marine research questions. The experimental results reveal that existing VLMs cannot effectively answer the domain-specific questions, and there is still a large room for further performance improvements. We hope our new benchmark and observations will facilitate future research. Project Page: http://marineeval.hkustvgd.com/

LongVideoAgent: Multi-Agent Reasoning with Long Videos

Recent advances in multimodal LLMs and systems that use tools for long-video QA point to the promise of reasoning over hour-long episodes. However, many methods still compress content into lossy summaries or rely on limited toolsets, weakening temporal grounding and missing fine-grained cues. We propose a multi-agent framework in which a master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted textual observations. The master agent plans with a step limit, and is trained with reinforcement learning to encourage concise, correct, and efficient multi-agent cooperation. This design helps the master agent focus on relevant clips via grounding, complements subtitles with visual detail, and yields interpretable trajectories. On our proposed LongTVQA and LongTVQA+ which are episode-level datasets aggregated from TVQA/TVQA+, our multi-agent system significantly outperforms strong non-agent baselines. Experiments also show reinforcement learning further strengthens reasoning and planning for the trained agent. Code and data will be shared at https://longvideoagent.github.io/.

Skin-R1: Toward Trustworthy Clinical Reasoning for Dermatological Diagnosis

The emergence of vision-language models (VLMs) has opened new possibilities for clinical reasoning and has shown promising performance in dermatological diagnosis. However, their trustworthiness and clinical utility are often limited by three major factors: (1) Data heterogeneity, where diverse datasets lack consistent diagnostic labels and clinical concept annotations; (2) Absence of grounded diagnostic rationales, leading to a scarcity of reliable reasoning supervision; and (3) Limited scalability and generalization, as models trained on small, densely annotated datasets struggle to transfer nuanced reasoning to large, sparsely-annotated ones. To address these limitations, we propose SkinR1, a novel dermatological VLM that combines deep, textbook-based reasoning with the broad generalization capabilities of reinforcement learning (RL). SkinR1 systematically resolves the key challenges through a unified, end-to-end framework. First, we design a textbook-based reasoning generator that synthesizes high-fidelity, hierarchy-aware, and differential-diagnosis (DDx)-informed trajectories, providing reliable expert-level supervision. Second, we leverage the constructed trajectories for supervised fine-tuning (SFT) empowering the model with grounded reasoning ability. Third, we develop a novel RL paradigm that, by incorporating the hierarchical structure of diseases, effectively transfers these grounded reasoning patterns to large-scale, sparse data. Extensive experiments on multiple dermatology datasets demonstrate that SkinR1 achieves superior diagnostic accuracy. The ablation study demonstrates the importance of the reasoning foundation instilled by SFT.

RIDE: Difficulty Evolving Perturbation with Item Response Theory for Mathematical Reasoning

Large language models (LLMs) achieve high performance on mathematical reasoning, but these results can be inflated by training data leakage or superficial pattern matching rather than genuine reasoning. To this end, an adversarial perturbation-based evaluation is needed to measure true mathematical reasoning ability. Current rule-based perturbation methods often generate ill-posed questions and impede the systematic evaluation of question difficulty and the evolution of benchmarks. To bridge this gap, we propose RIDE, a novel adversarial question-rewriting framework that leverages Item Response Theory (IRT) to rigorously measure question difficulty and to generate intrinsically more challenging, well-posed variations of mathematical problems. We employ 35 LLMs to simulate students and build a difficulty ranker from their responses. This ranker provides a reward signal during reinforcement learning and guides a question-rewriting model to reformulate existing questions across difficulty levels. Applying RIDE to competition-level mathematical benchmarks yields perturbed versions that degrade advanced LLM performance, with experiments showing an average 21.73% drop across 26 models, thereby exposing limited robustness in mathematical reasoning and confirming the validity of our evaluation approach.

Pointing to a Llama and Call it a Camel: On the Sycophancy of Multimodal Large Language Models

Multimodal large language models (MLLMs) have demonstrated extraordinary capabilities in conducting conversations based on image inputs. However, we observe that MLLMs exhibit a pronounced form of visual sycophantic behavior. While similar behavior has also been noted in text-based large language models (LLMs), it becomes significantly more prominent when MLLMs process image inputs. We refer to this phenomenon as the "sycophantic modality gap." To better understand this issue, we further analyze the factors that contribute to the exacerbation of this gap. To mitigate the visual sycophantic behavior, we first experiment with naive supervised fine-tuning to help the MLLM resist misleading instructions from the user. However, we find that this approach also makes the MLLM overly resistant to corrective instructions (i.e., stubborn even if it is wrong). To alleviate this trade-off, we propose Sycophantic Reflective Tuning (SRT), which enables the MLLM to engage in reflective reasoning, allowing it to determine whether a user's instruction is misleading or corrective before drawing a conclusion. After applying SRT, we observe a significant reduction in sycophantic behavior toward misleading instructions, without resulting in excessive stubbornness when receiving corrective instructions.

Remote Sensing Image Intelligent Interpretation with the Language-Centered Perspective: Principles, Methods and Challenges

The mainstream paradigm of remote sensing image interpretation has long been dominated by vision-centered models, which rely on visual features for semantic understanding. However, these models face inherent limitations in handling multi-modal reasoning, semantic abstraction, and interactive decision-making. While recent advances have introduced Large Language Models (LLMs) into remote sensing workflows, existing studies primarily focus on downstream applications, lacking a unified theoretical framework that explains the cognitive role of language. This review advocates a paradigm shift from vision-centered to language-centered remote sensing interpretation. Drawing inspiration from the Global Workspace Theory (GWT) of human cognition, We propose a language-centered framework for remote sensing interpretation that treats LLMs as the cognitive central hub integrating perceptual, task, knowledge and action spaces to enable unified understanding, reasoning, and decision-making. We first explore the potential of LLMs as the central cognitive component in remote sensing interpretation, and then summarize core technical challenges, including unified multimodal representation, knowledge association, and reasoning and decision-making. Furthermore, we construct a global workspace-driven interpretation mechanism and review how language-centered solutions address each challenge. Finally, we outline future research directions from four perspectives: adaptive alignment of multimodal data, task understanding under dynamic knowledge constraints, trustworthy reasoning, and autonomous interaction. This work aims to provide a conceptual foundation for the next generation of remote sensing interpretation systems and establish a roadmap toward cognition-driven intelligent geospatial analysis.

LegalReasoner: Step-wised Verification-Correction for Legal Judgment Reasoning

Legal judgment prediction (LJP) aims to function as a judge by making final rulings based on case claims and facts, which plays a vital role in the judicial domain for supporting court decision-making and improving judicial efficiency. However, existing methods often struggle with logical errors when conducting complex legal reasoning. We propose LegalReasoner, which enhances LJP reliability through step-wise verification and correction of the reasoning process. Specifically, it first identifies dispute points to decompose complex cases, and then conducts step-wise reasoning while employing a process verifier to validate each step's logic from correctness, progressiveness, and potential perspectives. When errors are detected, expert-designed attribution and resolution strategies are applied for correction. To fine-tune LegalReasoner, we release the LegalHK dataset, containing 58,130 Hong Kong court cases with detailed annotations of dispute points, step-by-step reasoning chains, and process verification labels. Experiments demonstrate that LegalReasoner significantly improves concordance with court decisions from 72.37 to 80.27 on LLAMA-3.1-70B. The data is available at https://huggingface.co/datasets/weijiezz/LegalHK.

ExeSQL: Self-Taught Text-to-SQL Models with Execution-Driven Bootstrapping for SQL Dialects

Recent text-to-SQL models have achieved strong performance, but their effectiveness remains largely confined to SQLite due to dataset limitations. However, real-world applications require SQL generation across multiple dialects with varying syntax and specialized features, which remains a challenge for current models. The main obstacle in building a dialect-aware model lies in acquiring high-quality dialect-specific data. Data generated purely through static prompting - without validating SQLs via execution - tends to be noisy and unreliable. Moreover, the lack of real execution environments in the training loop prevents models from grounding their predictions in executable semantics, limiting generalization despite surface-level improvements from data filtering. This work introduces ExeSQL, a text-to-SQL framework with execution-driven, agentic bootstrapping. The method consists of iterative query generation, execution-based filtering (e.g., rejection sampling), and preference-based training, enabling the model to adapt to new SQL dialects through verifiable, feedback-guided learning. Experiments show that ExeSQL bridges the dialect gap in text-to-SQL, achieving average improvements of 15.2%, 10.38%, and 4.49% over GPT-4o on PostgreSQL, MySQL, and Oracle, respectively, across multiple datasets of varying difficulty.

MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly

The rapid extension of context windows in large vision-language models has given rise to long-context vision-language models (LCVLMs), which are capable of handling hundreds of images with interleaved text tokens in a single forward pass. In this work, we introduce MMLongBench, the first benchmark covering a diverse set of long-context vision-language tasks, to evaluate LCVLMs effectively and thoroughly. MMLongBench is composed of 13,331 examples spanning five different categories of downstream tasks, such as Visual RAG and Many-Shot ICL. It also provides broad coverage of image types, including various natural and synthetic images. To assess the robustness of the models to different input lengths, all examples are delivered at five standardized input lengths (8K-128K tokens) via a cross-modal tokenization scheme that combines vision patches and text tokens. Through a thorough benchmarking of 46 closed-source and open-source LCVLMs, we provide a comprehensive analysis of the current models' vision-language long-context ability. Our results show that: i) performance on a single task is a weak proxy for overall long-context capability; ii) both closed-source and open-source models face challenges in long-context vision-language tasks, indicating substantial room for future improvement; iii) models with stronger reasoning ability tend to exhibit better long-context performance. By offering wide task coverage, various image types, and rigorous length control, MMLongBench provides the missing foundation for diagnosing and advancing the next generation of LCVLMs.