SLVR：基于类人推理流的结构化潜在视觉推理
原标题：SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows
AI 摘要
该论文提出 SLVR 训练框架，将多模态推理组织为规划、定位、证据选择与推理整合等类型化潜在阶段，并用特殊分隔符 token 对每个阶段施加对应监督信号。SLVR 还引入 Visual Dependency Induction 对比训练，通过遮蔽答案提示文本并加入视觉干扰项，减少模型对语言先验的依赖。基于 Qwen2.5-VL-7B 构建的 SLVR 在 MMVP 上提升 9.4、BLINK Relation 上提升 14.2，并在 V*、MathVista、ChartQA 上也有改进，且无需在推理时生成文本思维链。
正文节选
SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows Abstract Multimodal large language models (MLLMs) often answer visual reasoning questions by relying on linguistic priors rather than task-relevant visual evidence. Textual chain-of-thought reasoning can partially mitigate this issue by encouraging models to decompose visual questions into intermediate evidence-seeking steps, but generating these steps autoregressively increases inference cost. Latent reasoning avoids explic