In this paper, we introduce Predicate-State Video Reasoning Segmentation (PS-VRS), where a query defines a dynamic predicate P(o,t) over objects and time rather than a fixed object identity. Existing methods typically treat the target as permanently active once identified, leading to over-emission of masks during vacant intervals and failure to handle target switching. To address this, we construct PS-Bench, a large-scale benchmark with 2,077 videos and 16,250 queries annotated with explicit active/vacant predicate states, temporal intervals, and pixel masks. We further propose Predicate-State Grounding (PSG), a State-Bind-Emit pipeline that: (1) predicts [ACTIVE]/[VACANT] state tokens; (2) aligns them to video frames via dense cross-modal similarity; and (3) emits masks via SAM2 only for active states. Experiments show that PSG outperforms existing baselines by +9.32 𝒥 ℱ on PS-Bench and significantly reduces false-positive mask emissions. Our results demonstrate the effectiveness of explicit predicate-state modeling for video reasoning segmentation with dynamic visual witnesses.

Can financial vision–language models (VLMs) turn chart evidence into reliable action recommendations? Existing hallucination evaluations are mostly claim-centric: they assess whether generated statements are supported, but not whether evidence remains traceable through rationale, confidence, and final action. We introduce E2A-Bench, a 969-query benchmark for financial chart reasoning, constructed from 323 HS300 constituents under three input modalities with deterministic OHLCV-derived evidence anchors. E2A-Bench evaluates grounding, reasoning–action consistency, evidence–confidence calibration, and directional coverage through UCR, RCI, ECI, and NDR, where NDR measures coverage-aware evidence-to-action reliability rather than realized trading performance. Evaluating 20 VLMs reveals three failures hidden by scalar hallucination scores: the lowest-UCR model ranks near the bottom by NDR due to only 6.4% directional coverage; oracle-aided verification reduces unsupported claims but can collapse coverage; and financial fine-tuning amplifies the BUY:SELL ratio by 4.21–4.68× across strict base–fine-tuned pairs. These results show that financial VLM evaluation should trace the full evidence-to-action chain rather than rely on a single hallucination score. Code and data: https://github.com/wanng-ide/E2A-Bench .

Vision-language-action (VLA) models for autonomous driving typically express intermediate reasoning entirely in text, which may discard decision-relevant visual information. We propose Latent Dreaming (LatDM), a latent–text reasoning framework that complements textual chain-of-thought with continuous latent states. LatDM uses latent blocks to preserve visually grounded spatial and temporal information, while text handles semantic abstraction and decision communication. A pretrained visual encoder provides latent supervision during training, and the model autonomously interleaves latent and textual reasoning at inference. Experiments on VLADBench, CODA-LM, and NuInstruct show consistent improvements over the underlying VLA backbone, while ablations confirm that the latent channel contributes beyond standard textual fine-tuning.

Reinforcement learning from verifiable rewards improves the reasoning ability of large language models, but often suffers from entropy collapse, in which increasingly concentrated policies reduce rollout diversity and useful learning signals. Existing remedies either constrain the RL objective (e.g., entropy regularization) or adjust sampling temperature during rollout collection, but these interventions remain external to the model parameters. We propose Temperature-Scaled On-Policy Self-Distillation (TS-OPSD), a lightweight policy reheating method that internalizes the exploratory effect of temperature into model parameters. Starting from an entropy-collapsed RL checkpoint, TS-OPSD constructs a self-teacher by applying high-temperature scaling to the model’s own logits, then distills the resulting smoother distribution back into the student. This policy reheating requires no external teacher, privileged data, or additional inference cost. Experiments on Qwen3-4B-Base and Qwen3-8B-Base show that policy reheating yields a stronger initialization for continued RL than both standard continued RL and rollout-level temperature reheating. Further analyses show that TS-OPSD mainly reduces output sharpness while preserving intermediate representations, top candidate sets, and reasoning capability. These results suggest that entropy restoration can serve as a simple post-collapse intervention for extending reasoning-oriented RL.

Evaluating multimodal large language models (MLLMs) on long-form scientific papers requires measuring not only answer correctness, but also whether answers are supported by traceable cross-modal evidence. Existing answer-only metrics and synthetic “Needle-In-A-Haystack” tests mainly assess answer matching or isolated retrieval, leaving evidence-grounded support under-examined. We introduce “Fish-in-the-Ocean” (FITO), an evidence-chain evaluation lens for native scientific documents, and instantiate it with SIN-Data, a scientific interleaved corpus that preserves text–figure coupling. Built on SIN-Data, SIN-Bench comprises four evidence-auditing tasks: evidence discovery, hypothesis verification, grounded QA, and evidence-anchored synthesis. We further propose “No Evidence, No Score”, which credits predictions only when linked to verifiable anchors and diagnoses evidence quality through matching, relevance, and logic. Experiments on 10 MLLMs show a clear correctness–grounding gap: Gemini-3-pro obtains the best average overall score (0.566), while GPT-5 reaches the highest SIN-QA answer accuracy (0.767) but lower evidence-aligned scores. Code and data are available at https://github.com/IIGROUP/sin-bench .

Supervised fine-tuning (SFT) is a fundamental post-training strategy to align Large Language Models (LLMs) with human intent. However, traditional SFT often ignores the one-to-many nature of language by forcing alignment with a single reference answer, leading to the model overfitting to non-core expressions. Although our empirical analysis suggests that introducing multiple reference answers can mitigate this issue, the prohibitive data and computational costs necessitate a strategic shift: prioritizing the mitigation of single-reference overfitting over the costly pursuit of answer diversity. To achieve this, we reveal the intrinsic connection between token probability and semantic importance: high-probability tokens carry the core logical framework, while low-probability tokens are mostly replaceable expressions. Based on this insight, we propose ProFit, which selectively masks low-probability tokens to prevent surface-level overfitting. Extensive experiments confirm that ProFit consistently outperforms traditional SFT baselines on general reasoning and mathematical benchmarks

The increasing adoption of large language models (LLMs) in software engineering necessitates rigorous security evaluation of their generated code. However, existing benchmarks often lack relevance to real-world AI-assisted programming scenarios, making them inadequate for assessing the practical security risks associated with AI-generated code in production environments. To address this gap, we introduce A.S.E (AI Code Generation Security Evaluation), a repository-level evaluation benchmark designed to closely mirror real-world AI programming tasks, offering a comprehensive and reliable framework for assessing the security of AI-generated code. Our evaluation of leading LLMs on A.S.E reveals several key findings. In particular, current LLMs still struggle with secure coding. The complexity in repository-level scenarios presents challenges for LLMs that typically perform well on snippet-level tasks. Moreover, a larger reasoning budget does not necessarily lead to better code generation. These observations offer valuable insights into the current state of AI code generation and help developers identify the most suitable models for practical tasks. They also lay the groundwork for refining LLMs to generate secure and efficient code in real-world applications.

Existing benchmarks for MLLM-generated web artifacts assess interaction through local evidence and miss the requirement-induced states and transitions that determine whether a page works. We introduce WebRISE, which compiles task requirements into Interaction Contract Graphs (ICGs) of observable states, user-intent transitions, and DOM/visual assertions for implementation-agnostic browser execution. WebRISE spans 442 tasks across five input modalities (Text, Markdown, Sketch, Image, Video), with 5,495 transitions and 5,271 requirement checks that separate user-stated functions from implicit product-level constraints. Across 14 MLLMs, even the strongest model reaches only 65.6% transition validity and 66.3% requirement coverage, and visual quality is no proxy for behavior (Qwen3.6-35B-A3B on Markdown: V=80.8 yet T=15.5). Video gives the strongest interaction signal (+10.6pp implicit coverage over Text), while implicit constraints persist; defect injection shows ICG-based scoring detects state errors at 2–16× the rate of checkpoint-style evaluation.

Universal multimodal retrieval (UMR) increasingly adopts multimodal large language models (MLLMs) as unified embedding backbones, but their strong retrieval performance comes at substantial inference cost. Existing methods typically rely on uniformly dense inference, where all input tokens are processed through the entire model and matched using the final-layer [EOS] representation. However, this paradigm overlooks two key forms of heterogeneity in multimodal retrieval: token contributions to the final retrieval embedding are highly uneven, and different queries require markedly different amounts of inference depth. To address this, we propose **Skim and Skip (SAS)**, a hierarchical adaptive inference framework for efficient multimodal retrieval. SAS first performs token-level evidence selection to preserve only the input information most relevant to the final retrieval embedding, and then performs depth-adaptive inference to determine whether the current representation is already sufficient for reliable matching. Experiments on 12 MMEB retrieval tasks show that SAS retains about 99% of the dense baseline’s average retrieval performance while reducing LLM-side FLOPs by up to **66.3%** and achieving up to **1.64×** end-to-end speedup. Our code is available at [https://github.com/GaoMengGladys/SAS](https://github.com/GaoMengGladys/SAS).

We present AnyCap, a plug-and-play framework that brings instruction alignment to omni-modal captioning. Captions offer a unified language interface for multimodal learning, and users increasingly expect instruction-driven control over their content and style. Current caption models lack explicit instruction supervision and are weak at instruction following, while directly tuning them can degrade general language ability. Achieving instruction alignment in an omni-modal setting is harder still, as each modality calls for separate models and custom designs. To address these challenges, AnyCap leverages a multimodal revision-based refinement paradigm that refines uncontrolled captions from existing models to instruction-aligned ones, without re-training base models. By processing multi-modality features in a unified framework, it enables one model to serve images, videos, and audio. To address the lack of instruction-based data, we construct AnyCapData, a large-scale, high-quality corpus spanning three modalities with 28 well-designed instruction types, each with multiple instruction samples. For evaluation, we address the limitations of current metrics for instruction-oriented captioning by designing AnyCapEval. Its key insight is to decouple evaluation into content and style for fine-grained assessment. Extensive experiments show that on AnyCapEval and diverse public benchmarks, AnyCap consistently improves both caption quality and instruction adherence for both open-source and API-based models. Notably, AnyCap-8B boosts GPT-4o’s content scores by 46% and style scores by 12%. Our code and models will be made publicly available.

Large Language Models (LLMs) have made notable progress in mathematical reasoning, yet they often rely on single-paradigm reasoning that limits their effectiveness across diverse tasks. In this paper, we introduce Chain-of-Reasoning (CoR), a novel unified framework that integrates multiple reasoning paradigms — Natural Language Reasoning (NLR), Algorithmic Reasoning (AR), and Symbolic Reasoning (SR) — to enable synergistic collaboration. CoR generates multiple potential answers using different reasoning paradigms and synthesizes them into a coherent final solution. We propose a Progressive Paradigm Training (PPT) strategy that allows models to progressively master these paradigms, culminating in the development of at CoR-Math-7B. Experimental results demonstrate that CoR-Math-7B significantly outperforms current SOTA models, achieving up to a 41.0% absolute improvement over GPT-4o in theorem proving tasks and a 15% improvement over RL-based methods on the MATH benchmark in arithmetic tasks. These results show the enhanced mathematical comprehensive ability of our model, enabling zero-shot generalization across tasks.The code is available at https://github.com/microsoft/CoR.

Generative AI has demonstrated unprecedented creativity in the field of computer vision, yet such phenomena have not been observed in natural language processing. In particular, large language models (LLMs) can hardly produce written works at the level of human experts due to the extremely high complexity of literature writing. In this paper, we present HoLLMwood, an automated framework for unleashing the creativity of LLMs and exploring their potential in screenwriting, which is a highly demanding task. Mimicking the human creative process, we assign LLMs to different roles involved in the real-world scenario. In addition to the common practice of treating LLMs as Writer, we also apply LLMs as Editor, who is responsible for providing feedback and revision advice to Writer. Besides, to enrich the characters and deepen the plots, we introduce a role-playing mechanism and adopt LLMs as Actors that can communicate and interact with each other. Evaluations on automatically generated screenplays show that HoLLMwood substantially outperforms strong baselines in terms of coherence, relevance, interestingness and overall quality.

Tool-augmented large language models (LLMs) are rapidly being integrated into real-world applications. Due to the lack of benchmarks, the community has yet to fully understand the hallucination issues within these models. To address this challenge, we introduce a comprehensive diagnostic benchmark, ToolBH. Specifically, we assess the LLM’s hallucinations through two perspectives: depth and breadth. In terms of depth, we propose a multi-level diagnostic process, including (1) solvability detection, (2) solution planning, and (3) missing-tool analysis. For breadth, we consider three scenarios based on the characteristics of the toolset: missing necessary tools, potential tools, and limited functionality tools. Furthermore, we developed seven tasks and collected 700 evaluation samples through multiple rounds of manual annotation. The results show the significant challenges presented by the ToolBH benchmark. The current advanced models Gemini-1.5-Pro and GPT-4o only achieve total scores of 45.3 and 37.0, respectively, on a scale of 100. In this benchmark, larger model parameters do not guarantee better performance; the training data and response strategies also play crucial roles in tool-enhanced LLM scenarios. Our diagnostic analysis indicates that the primary reason for model errors lies in assessing task solvability. Additionally, open-weight models suffer from performance drops with verbose replies, whereas proprietary models excel with longer reasoning.

We propose a new paradigm for universal information extraction (IE) that is compatible with any schema format and applicable to a list of IE tasks, such as named entity recognition, relation extraction, event extraction and sentiment analysis. Our approach converts the text-based IE tasks as the token-pair problem, which uniformly disassembles all extraction targets into joint span detection, classification and association problems with a unified extractive framework, namely UniEX. UniEX can synchronously encode schema-based prompt and textual information, and collaboratively learn the generalized knowledge from pre-defined information using the auto-encoder language models. We develop a traffine attention mechanism to integrate heterogeneous factors including tasks, labels and inside tokens, and obtain the extraction target via a scoring matrix. Experiment results show that UniEX can outperform generative universal IE models in terms of performance and inference-speed on 14 benchmarks IE datasets with the supervised setting. The state-of-the-art performance in low-resource scenarios also verifies the transferability and effectiveness of UniEX.

Large-scale pre-trained language models (PLMs) bring new opportunities to challenging problems, especially those that need high-level intelligence, such as the math word problem (MWPs). However, directly applying existing PLMs to MWPs can fail as the generation process lacks sufficient supervision and thus lacks fast adaptivity as humans. We notice that human reasoning has a dual reasoning framework that consists of an immediate reaction system (system 1) and a delicate reasoning system (system 2), where the entire reasoning is determined by their interaction. This inspires us to develop a cooperative reasoning-induced PLM for solving MWPs, called Cooperative Reasoning (CoRe), resulting in a human-like reasoning architecture with system 1 as the generator and system 2 as the verifier. In our approach, the generator is responsible for generating reasoning paths, and the verifiers are used to supervise the evaluation in order to obtain reliable feedback for the generator. We evaluate our CoRe framework on several mathematical reasoning datasets and achieve decent improvement over state-of-the-art methods, up to 9.6% increase over best baselines.