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.
Shengqin Wang, Jie Jin, Xiaofan Li, Wentao Yan, Yu Cheng, Qingsong HuYihang Chen, Weilin Luo, Jingyu Gong 等 11 位作者
Multimodal tool-using agents augment vision-language models with access to external information and capabilities, enabling them to solve knowledge-intensive and visually complex tasks that cannot be addressed using visual context alone. However, existing reinforcement learning pipelines largely train single-role agents to search and answer using terminal rewards, overlooking how humans inspect retrieved evidence, reformulate queries, and revise conclusions during search. Consequently, they fail to optimize this self-corrective process over complete interaction trajectories. We propose MRC-MMAgent, a multi-role collaborative reinforcement learning framework that incorporates trajectory assessment and continuation directly into the executable agent loop. First, the framework establishes a multi-role collaborative loop that integrates solving, inspection, and continuation, enabling the agent to acquire evidence through tools, evaluate the preceding trajectory, and resume problem solving when the current answer is insufficiently supported. Second, we introduce a role-coupled learning objective that combines task success, interaction validity, and successful correction with full-trajectory policy optimization across the decisions generated by all roles. Extensive experiments show that, using only the original baseline training set without constructing additional datasets, our method consistently outperforms the corresponding base models and achieves leading or highly competitive performance across benchmarks.
Andie Huang, Ying Pang, Qian Zhang, Yihang Cheng, Lin Ma
Many LLMs now emit an explicit reasoning trace before they write code, but benchmark results for code generation and bug fixing usually report correctness alone. It is therefore unclear how much the extra reasoning contributes relative to the tokens it consumes. We examine this question with 866 Python tasks drawn from HumanEval Pro, MBPP Pro, and BugsInPy under five model settings. We look at the accuracy gain and token overhead of reasoning, the behaviors that appear inside reasoning traces, and whether a shortened trace still leads to a correct solution. Turning reasoning on in Qwen3-32B over five runs improves mean pass@1 on HumanEval Pro from 71.1% to 76.5% and mean tests pass on BugsInPy from 3.5% to 7.9%. On MBPP Pro the improvement from 68.1% to 69.5% is not statistically significant, even though the model uses a median of over 2,000 thinking tokens per task. Across models, a longer median trace does not reliably mean higher accuracy. A qualitative analysis shows that traces mostly choose and check an implementation in code generation, while in bug fixing they mostly locate the fault and plan the repair. We then truncate stored traces offline and ask Qwen3-32B with reasoning off to produce a solution from what remains. Retaining the first half of a trace yields no statistically significant drop in correctness. Stopping when the trace becomes repetitive seldom triggers, and stopping once solution code first appears benefits some models but harms others. Neither rule beats plain fixed truncation. These results suggest that more reasoning is not always better and that studies should report reasoning cost alongside correctness.
Yihang Chen, Bin Tan, Yujun Shen, Yinghao Xu, Nan Xue
Recent multi-view Transformers (e.g., DUSt3R and VGGT) have advanced 3D reconstruction by scaling training data and model capacity from two views to many views and long sequences, subsuming the classical SfM/SLAM pipeline of feature extraction, matching, and pose estimation in a single feed-forward pass. Yet this subsumption hides the interfaces between pipeline stages---correspondences, geometry, and pose are entangled in a single forward pass, leaving supervised fine-tuning as the major path to improvement and offering no mechanism to correct predictions from geometric evidence at test time. We propose PosePlaner to close this gap by reviving the classical coupling between correspondences and pose via flow matching, treating pose refinement as a conditional generation problem that takes coarse geometric matches and initial pose as input condition and fuses the resulting pose hypotheses into a single refined estimate, providing a continuous, data-driven analog of RANSAC voting. Our method is trained efficiently on two-view data, yet serves as a periodic refiner for streaming n-view pose estimators with no retraining. Empirically, it corrects failures of feed-forward predictors, substantially improves correspondence-based solvers, and reduces accumulated drift in long streaming sequences.