让多模态推理更稳定，通过迭代聚合改善答案一致性。
CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation
- 用多条推理路径迭代聚合，结合视觉验证过滤错误步骤。
- 在ScienceQA上提升26.2个百分点，EgoSchema上提升9.1个百分点。
- 适合需要可靠多步推理的视觉理解任务，如视频问答。
视觉语言模型在多模态理解与推理任务中表现优异，但其多步推理过程仍不稳定。相同输入重复采样常导致推理路径发散、最终预测不一致。为此，我们提出两种互补方法：(1) CASHEW，一种推理时框架，通过迭代聚合多个候选推理路径生成更高质量的推理轨迹，并通过显式视觉验证过滤幻觉步骤，确保推理基于视觉证据；(2) CASHEW-RL，一种内化聚合行为的可学习变体，使用分组序列策略优化（GSPO）训练，采用复合奖励函数鼓励基于最小且充分视觉证据得出正确答案，并根据任务难度自适应分配推理资源。该训练目标使模型在推理时具备鲁棒的自我聚合能力。在13个图像理解、视频理解与视频推理基准上的实验显示显著性能提升，ScienceQA最高提升26.2个百分点，EgoSchema提升9.1个百分点。
原文摘要 · Abstract (English)
Vision-language models achieve strong performance across a wide range of multimodal understanding and reasoning tasks, yet their multi-step reasoning remains unstable. Repeated sampling over the same input often produces divergent reasoning trajectories and inconsistent final predictions. To address this, we introduce two complementary approaches inspired by test-time scaling: (1) CASHEW, an inference-time framework that stabilizes reasoning by iteratively aggregating multiple candidate trajectories into higher-quality reasoning traces, with explicit visual verification filtering hallucinated steps and grounding reasoning in visual evidence, and (2) CASHEW-RL, a learned variant that internalizes this aggregation behavior within a single model. CASHEW-RL is trained using Group Sequence Policy Optimization (GSPO) with a composite reward that encourages correct answers grounded in minimal yet sufficient visual evidence, while adaptively allocating reasoning effort based on task difficulty. This training objective enables robust self-aggregation at inference. Extensive experiments on 13 image understanding, video understanding, and video reasoning benchmarks show significant performance improvements, including gains of up to +26.2 percentage points on ScienceQA and +9.1 percentage points on EgoSchema.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。