As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at https://github.com/zhcz328/CRAFT.
We present Youtu-Parsing-Omni, an omni-modal parser that converts documents, images, audio, and video into structured records grounded in space and time, providing a perception layer for agents that read, locate, and cite evidence across modalities. Its design rests on one encoder for every modality and one schema for every task. Youtu-Omni-Encoder, initialized from a pre-trained language model to strengthen text reading, encodes all modalities with shared weights in a unified space-time coordinate system, and a few fusion layers let video frames attend to concurrent audio, bringing cross-modal fusion and temporal alignment into perception itself. OmniSchema unifies the outputs of all ten parsing tasks, allowing a single self-evolving post-training procedure, OmniSchema-Aware On-Policy Distillation (OSAD), to provide a shared token-role-based objective across tasks. In each round, the model is trained on its own rollouts against token-level targets from a reference-conditioned copy of itself, which is then refreshed with the trained weights, so that the parser and its supervision co-evolve. With a single 5 B-parameter checkpoint, Youtu-Parsing-Omni achieves state-of-the-art performance on OmniDocBench v1.6 (96.96 Overall) and leads all open-weight models on OmniParsingBench (75.08 Avg.), with its advantage concentrated on audio and video. It also achieves the best overall chart-parsing score (95.02) and the highest exact chemical-fingerprint agreement among compared systems, and matches a dedicated optical-music model in character-level transcription of typeset sheet music while staying far ahead of general-purpose systems. Code and model weights are publicly available.
Video understanding keeps moving toward longer recordings, and the longer a recording is, the more questions can be asked about it. But processing it again for every question does not scale, so many systems watch the video once, build a memory, and answer later questions from it. Yet what to remember is decided before anyone knows what will be asked, and whatever is left out cannot be recovered later. Moreover, whether the savings pay off is judged by token counts, which fall when the input gets shorter but not when the model reuses what it has already computed. We present REEL, an evaluation framework that compares three saved representations built from exactly the same observed frames, so that a change in answers can come only from what the system remembered. Text writes a short record of what happened, Visual keeps the encoder outputs, and Prefix State also keeps the language-model prefix. Each is compared with full reprocessing (Frames) by paired correctness changes, cumulative GPU time, and storage size. Across six vision-language models and two benchmarks, Text and Frames differ in correctness on 20\% to 29\% of questions, even when their mean accuracy is almost the same. By contrast, Visual and Prefix State stay much closer to Frames. With unlimited storage, Prefix State is the cheapest in total GPU time among memories that keep almost all of the answers Frames gets right, for every model on both benchmarks. The trade-off is storage: Prefix State keeps up to about 21 GiB per video, whereas Text needs only kilobytes. Reusing computation and discarding information both look like efficiency, but only the second removes evidence that later questions may need, and neither mean accuracy nor token count tells them apart. REEL makes that difference measurable for any system that answers many questions from one saved memory.
Diffusion video models produce striking clips, but they cannot reliably render an exact formula, a correct chart, a rule-consistent game or narration synchronized to on-screen events. A different route is emerging: multimodal agents that write code to render every frame and synthesize every sound. We introduce VibeVideoBench, a benchmark for evaluating agents on code-driven video generation. Its 239 self-contained tasks in English and Chinese span 12 categories, from data stories and music-driven pieces to classroom lessons and paper explainers; each asks an agent, running in its native harness inside a sandbox with no image or video generation model, to deliver a finished MP4 with sound within one hour. Evaluation is hierarchical and evidence-based: program checks of the delivery spec, about 13k task-specific rubric checks traced to the requirements they enforce, and five quality dimensions anchored per task, scored by three VLM judge panels. Across nine frontier agent systems and 2,151 scored rollouts, GPT-6 Astra and Claude Opus 5.5 rank first and second with both the judges and human annotators, and the strongest system satisfies every core requirement on 76.3% of tasks. Delivery is a second axis: counting missing videos as zero changes the ranking, and three quarters of the lost videos are final renders that started too late, ran too slowly or never started, a failure of how agents plan their hour. Agents build their own rendering pipelines and inspect their own frames, yet how much they look does not predict quality. The judges' ranking of the systems closely matches blind human preferences.
Agents end many steps in a typed decision over a set of candidates: which screen element to click, which product to pick. System One models answer in one forward pass, and Jeva showed that four properties the setting demands - a "none of the above" probability, invariance to candidate order, scalability in the number of candidates, and reuse of candidate computation - follow from encoding the request and each candidate separately and letting a request state read a candidate memory. Jeva is text-only, and existing multimodal System One models (OneJev, Jev-Omni) put images in the request while candidates stay text in letter slots. This paper asks what happens to the four properties, and to accuracy, when the candidates themselves are images. Iris admits image, text, or mixed candidates and leaves Jeva's decision module untouched. On a visual-candidate benchmark built from Multimodal-Mind2Web (every candidate is the element's screenshot crop, K=16, three splits), independent encoding reaches 78.4/72.1/74.8% with a flip rate of exactly zero, while joint encoding (OneJev-4B fine-tuned on the same data) reaches 90.6/89.5/91.6% but changes its answer on 8-10% of records when the crops are reordered. Most of the gap is the candidate not seeing the request: a third design that encodes each candidate after the state - but never after another candidate - keeps order invariance exactly, shares the state pass through a forked cache, and reaches 92.1/87.0/90.2% at a third of OneJev's latency with a calibrated null probability (AUROC 0.93-0.97). Letting pooled candidates see each other (set attention, token-level reads, multi-vector memory) buys nothing, because the encoder maps visually distinct elements to vectors within cosine 0.999 of each other. Reusing image candidates across requests cuts latency 37x at K=256 and 191x at K=2048. Code: https://github.com/lllfx/Iris
General-purpose embodied intelligence requires robots to coordinate diverse skills over long horizons, track task progress, and recover from failures. Specialist policies alone do not ensure task-level reliability, while powerful multimodal agents rely on costly reasoning and external experience that compact models struggle to reproduce. We introduce OmniBrain-0, a lightweight unified embodied agent that coordinates manipulation and navigation through a shared high-level decision model, integrating task understanding, planning, progress tracking, and feedback-driven recovery. Our progressive training framework combines embodied pretraining, domain-specialized agent distillation, and multi-teacher on-policy distillation (MOPD). It distills powerful teachers' interaction experience into manipulation and navigation experts, then consolidates their decision capabilities into OmniBrain-0. Unification therefore occurs within the learned decision model rather than only at the tool interface. We further introduce Rubric-Guided Recursive Self-Improvement (RSI), a data flywheel in which a Meta-Agent turns execution evidence into refined rubrics, step-level reminders, and verified trajectories for domain-expert updates before their reconsolidation into OmniBrain-0 through MOPD. We evaluate OmniBrain-0 on LIBERO-Pro, LIBERO-Plus, RoboCasa365, RoboTwin 2.0, and R2R-CE Val-Unseen, as well as multiple real-world manipulation tasks. OmniBrain-0 combines strong cross-domain task performance with substantially lower inference and execution overhead than its closed-source teacher, demonstrating efficient transfer and consolidation of embodied capabilities into a compact unified agent.
Zhengxin Li, Chongzhen Tian, Hui Yuan, Hao Liu, Sam Kwong
Deploying large vision-language models (LVLMs) on resource-constrained embodied agents for 3D visual question answering necessitates cloud-edge collaboration. However, offloading massive visual data to the cloud over bandwidth-limited and noisy channels introduces severe latency and transmission distortion. To address this, we propose TP-JSCC, an efficient token pruning and joint source-channel coding method for LVLMs-based 3D-VQA in Cloud-Edge Collaboration. First, we develop a Local and Global Semantics-based Hybrid Token Pruning (LGS-HTP) module to significantly reduce the visual token count. By decoupling foreground-background tokens and jointly selecting local and global semantics, LGS-HTP effectively reduces redundancy while preserving reasoning-critical tokens. Second, we introduce a lightweight Token Joint Source-Channel Coding (T-JSCC) module to further compress the pruned tokens and ensure their robust transmission over error-prone channels. By employing lightweight cascaded linear layers and Signal-to-Noise Ratio (SNR)-adaptive mechanism, this module adapts to fluctuating channel conditions while maintaining minimal computational overhead. Extensive experiments on ScanQA and SQA3D datasets indicate the exceptional efficiency and robustness of our approach. Specifically, on ScanQA validation set, the LGS-HTP module could achieve 90% pruning ratio while incurring less than a 1% CIDEr loss. Building on this, the T-JSCC further compresses the tokens by 18 times, resulting in an overall 180 times data volume reduction for robust transmission, while saving 90% in computational complexity for LVLM deployment. Our code is available at https://github.com/yuanhui0325/TP-JSCC.
Yunfei Wang, Bo Du, Zhe Yang, Xin Liu, Zhiyu Lin, Tianxin Xu, Ji-Zhe Zhou
With the rapid evolution of synthetic media, Image Manipulation Localization (IML) has emerged as a critical component in multimedia forensics for ensuring the integrity of digital content. However, generalization remains a core challenge, as existing discriminative methods typically learn a fixed decision boundary that tends to overfit to specific training artifacts and fails to adapt to unseen manipulation types. To address this, we propose DiffIML, a novel framework that introduces score-based generative modeling to IML. Diverging from the direct estimation of hard boundaries, DiffIML approximates the score function, the gradient of the log-likelihood, to capture the intrinsic geometric topology of mask distributions. This paradigm leverages structural priors to iteratively recover coherent masks from noise, thereby circumventing the brittleness associated with discriminative models. Under this formulation, diffusion models serve as an effective numerical solver for the learned score function.To ensure practicality, we respectively resolve the efficiency and stability bottlenecks of standard diffusion by: (1) utilizing a Lightweight Mask-Specific VAE for fast latent-space process and a decoupled architecture with a lightweight denoising UNet, (2) edge supervision and error prior to mitigate error accumulation during sampling. Extensive experiments of two distinct protocols on eight non-generative and three generative benchmarks demonstrate that DiffIML consistently outperforms state-of-the-art methods, yielding remarkable generalization improvements on diverse unseen datasets. The code is publicly available at https://github.com/scu-zjz/DiffIML.
World action models (WAMs) have emerged as an important class of generalist robot policies, aiming to transfer video prediction priors to action learning. However, we find that existing WAMs frequently produce incorrect motion predictions in open-ended environment, leading to erroneous actions. We attribute this limitation to two factors: (1) base video models are not optimized for manipulation, and (2) naively incorporating action components into video models can substantially degrade their generalization capabilities. We introduce Video Prediction Policy 2 (VPP2), a WAM that enables strong zero-shot generalization in both video prediction and action generation. First, we curate a large-scale, diverse dataset of manipulation videos to continue pretraining the base video foundation model. We annotate video clips with detailed captions and perform \textit{event-level} video pretraining to promote generalization across open-ended manipulation tasks. Second, we post-train and distill the video model into a single-step visual planner with fixed prediction horizon. Finally, we introduce action module via a mixture-of-transformers (MoT) architecture to learn implicit inverse dynamics model. Experiments demonstrate three key results: (1) VPP2-14B outperforms Cosmos3-64B by 11.0\% points in video prediction instruction-following success rate on open-ended tasks; (2) VPP2 surpasses the strongest baseline by 18.5\% points in success rate on real-world zero-shot ALOHA manipulation tasks; and (3) following benchmark-specific post-training, VPP2 achieves the highest success rates among evaluated methods on the challenging LIBERO-Pro, LIBERO-OOD, and RoboDojo benchmarks.
Recent advances have enabled unified omni-modal models in understanding audio, vision, and language. However, existing benchmarks, training data, and learning methods largely treat the modalities independently, leaving the capability of audio-visual joint reasoning poorly evaluated and insufficiently elicited. We address this gap with a benchmark, data engine, and learning method. First, we introduce OmniReasoningBench, a benchmark where both audio and visual evidence are indispensable. It comprises 1,150 multiple-choice and open-ended questions across two tasks, reasoning over video and reasoning beyond video. Second, we develop a data engine OmniQA. It automatically constructs evidence-grounded QA pairs that explicitly necessitate audio-visual joint reasoning, together with time-stamped clue chains that guide the annotation of thinking process. Besides our benchmark, this engine produces training data OmniReasoning-SFT-112K and OmniReasoning-RL-19K. Finally, we propose an on-policy self-distillation method Modality-Factored Self-Distillation (MFSD). It evaluates each sampled response under modality-specific clue contexts, disentangling the contributions of individual clues and their cross-modal interactions for token-level credit assignment. With our training data and learning method, our model OmniReasoning-30B-A3B achieves 50.0% on OmniVideoBench and 42.5% on OmniReasoningBench, improving the base model Qwen3-Omni-30B-A3B-Thinking by 12.8 and 9.3 percentage points, respectively. Moreover, it delivers substantial gains on general and long-video benchmarks, including Video-MME-v2. We hope our work offers a solid step for facilitating future research in omni-modal joint reasoning.
Yang Liu, Jinxuan Cai, Yishen Li, Qi Meng Beijing University of Posts and Telecommunications, Zedi Liu, Xin Li Free Org., Chen Qian Shanghai Jiaotong University, Chuan Shi Beijing University of Post and Telecommunication 等 9 位作者
Large language model-based (LLM-based) multi-agent systems (MAS) are increasingly used to extend agentic problem solving via role specialization and collaboration. MAS workflows can be naturally modeled as directed computation graphs, where nodes execute agents or sub-workflows and edges encode dependencies and message passing. However, implementing complex graph workflows in current frameworks still requires substantial manual effort, offers limited reuse, and makes it difficult to integrate heterogeneous external context sources. To overcome these limitations, we present MASFactory, a graph-centric framework for orchestrating LLM-based MAS. It introduces Vibe Graphing, a human-in-the-loop approach that compiles natural-language intent into an editable workflow specification and then into an executable graph. In addition, the framework provides reusable components, skill support, multimodal message handling, and pluggable context integration, as well as a visualizer for topology preview, runtime tracing, and human-in-the-loop interaction. We evaluate MASFactory on seven public benchmarks, validating both reproduction consistency for representative MAS methods and the effectiveness of Vibe Graphing. Our code (https://github.com/BUPT-GAMMA/MASFactory, licensed under Apache-2.0) and video demonstration (https://youtu.be/ANynzVfY32k) are publicly available.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI · cs.MA · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Tingyu Song University of the Chinese Academy of Sciences, Yanzhao Zhang Alibaba Group, Mingxin Li Alibaba Group, Zhuoning Guo Hong Kong University of Science and Technology (Guangzhou), Dingkun Long, Pengjun Xie Alibaba Group, Siyue Zhang, Yilun Zhao Yale University 等 9 位作者
Composed Image Retrieval (CIR) is a pivotal and complex task in multimodal understanding. Current CIR benchmarks typically feature limited query categories and fail to capture the diverse requirements of real-world scenarios. To bridge this evaluation gap, we leverage image editing to achieve precise control over modification types and content, enabling a pipeline for synthesizing queries across a broad spectrum of categories. Using this pipeline, we construct EDIR, a novel fine-grained CIR benchmark. EDIR encompasses 5,000 high-quality queries structured across five main categories and fifteen subcategories. Our comprehensive evaluation of 13 multimodal embedding models reveals a significant capability gap; even state-of-the-art models (e.g., RzenEmbed and GME) struggle to perform consistently across all subcategories, highlighting the rigorous nature of our benchmark. Through comparative analysis, we further uncover inherent limitations in existing benchmarks, such as modality biases and insufficient categorical coverage. Furthermore, an in-domain training experiment demonstrates the feasibility of our benchmark. This experiment clarifies the task challenges by distinguishing between categories that are solvable with targeted data and those that expose intrinsic limitations of current model architectures.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhenyu Liu, Yunxin Li Harbin Institute of Technology, Xuanyu Zhang, Qixun Teng, Shenyuan Jiang Dayu Algorithms Team, Xinyu Chen, Haoyuan Shi, Haolan Chen Tencent 等 15 位作者
Recent advances in unified multimodal models indicate a clear trend towards comprehensive content generation. However, the auditory domain remains a significant challenge, with music and speech often developed in isolation, hindering progress towards universal audio synthesis. This separation stems from inherent task conflicts between semantic speech and structural music modeling, and severe data imbalances, which impede the development of a truly unified model. To address these challenges, we propose UniMoE-Audio a unified speech and music generation model built upon a novel Dynamic-Capacity Mix-of-Experts (DCMoE) framework. Architecturally, UniMoE-Audio extends the conventional MoE paradigm by introducing a Top-P routing strategy for adaptive capacity allocation. To tackle data imbalance, we introduce a three-stage training curriculum: 1) Independent Specialist Training leverages original datasets to instill domain-specific knowledge into each specialists without interference; 2) MoE Integration and Warmup incorporates these specialists into the UniMoE-Audio architecture, warming up the gate module and shared expert using a subset of balanced dataset; and 3) Synergistic Joint Training trains the entire model end-to-end on the fully balanced dataset, fostering enhanced cross-domain synergy. Extensive experiments show that UniMoE-Audio not only achieves state-of-the-art performance on major speech and music generation benchmarks, but also demonstrates superior synergistic learning, mitigating the performance degradation typically seen in naive joint training. Our findings highlight the substantial potential of specialized MoE architecture and curated training strategies in advancing universal audio generation.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhiyu Xu Peking University, Lean Wang DeepSeek and Peking University, Yuanxin Liu Peking University, Lei Li, Hao Zhou Tencent, Fandong Meng WeChat AI, Tencent Inc., Jie Zhou, Xu Sun Peking University
Vision-Language Models (VLMs) have demonstrated remarkable proficiency in general multi-modal understanding; yet they struggle to efficiently acquire continually evolving domain-specific skills. Conventional approaches to enhancing VLM capabilities, such as Supervised Fine-Tuning (SFT), require extensive dataset curation and substantial computational resources. Model merging has emerged as an efficient alternative that enables the transfer of domain-specific expertise from Large Language Models (LLMs) to VLMs without incurring additional training data requirements or significant computational overhead. Unlike conventional merging of homogeneous LLMs, which mainly aggregates existing capabilities, cross-modal skill injection aims to induce emergent cross-modal capabilities by integrating a domain-expert LLM into a VLM. However, existing research lacks a systematic analysis of the applicability and methodology of cross-modal skill injection. In this study, we investigate cross-modal skill injection across three main aspects: scenarios, methods, and hyperparameters. For scenarios, we find that cross-modal skill injection generally performs well in instruction-following and cross-lingual settings, yet struggles with mathematical reasoning. For methods, we find that classic approaches such as TA and DARE consistently achieve superior performance over alternative merging methods. We also provide a systematic and quantitative analysis of the hyperparameter tuning that these classic methods critically depend on.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Jie Huang, Xin Liao, Junjie Wang Institute of Software, Chinese Academy of Sciences, Mingyang Li Institute of Software Chinese Academy of Sciences, Wenshuo Wang ISCAS, Ziyou Jiang Institute of Software Chinese Academy of Sciences, Shoubin Li, Qing Wang
With the rise of short-video platforms, hate speech has evolved from static text and memes into more covert and aggressive hateful video formats, profoundly impacting social dynamics and public sentiment. Existing detection methods typically rely on multimodal feature fusion, which blurs the distinct boundaries of modality-specific information. This leads to the feature dilution problem, where dominant benign modalities often overwhelm sparse, localized hateful cues. To address this, we propose SAGE (Synergistic Adaptive Gating of Experts), a novel framework that shifts the paradigm from blind feature mixing to decision-level arbitration. Mimicking human cognitive processes, SAGE instantiates disentangled experts to rigorously preserve modality-specific semantics, facilitates global expert deliberation for context-aware refinement, and convenes an instance-level tribunal to dynamically arbitrate the final verdict based on evidentiary salience. Extensive experiments on HateMM and MultiHateClip benchmarks demonstrate that SAGE significantly outperforms state-of-the-art methods, achieving accuracy gains of 6.37% to 21.23% and macro-F1 score gains of 6.77% to 28.01%.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表