arXiv:2608.30457 · 多模态
Learning Where Outcomes Change:Credit-Addressable Reasoning for Multimodal Geometry
学习结果变化的位置：面向多模态几何的可寻址信用推理
Learning Where Outcomes Change:Credit-Addressable Reasoning for Multimodal Geometry
- 类型：arxiv
- 标识：2608.30457
- 链接：https://arxiv.org/abs/2608.30457
- 主分类：multimodal
- 形态：method
- TLDR：Multimodal geometry reasoning requires VLMs to extract precise visual relations and preserve them through multi-step deduction. Existing free-form traces obscure the decisions that determine the answer, and trajectory-level reinforcement learning distributes a single terminal signal across the entire response. We introduce credit-addressable reasoning, in which the semantic units exposed during inference also define where learning compares alternatives and assigns credit. We instantiate this principle with Code-CoT, which retains the diagram, represents visual relations as line-addressable exe
- 待LLM分类：否
- 标题中文：学习结果变化的位置：面向多模态几何的可寻址信用推理
- TLDR中文：多模态几何推理要求 VLM 提取精确的视觉关系，并在多步演绎过程中保持这些关系。现有自由形式的推理轨迹掩盖了决定答案的关键决策，而轨迹级强化学习将单一的终端信号分散到整个响应中。我们提出可寻址推理（credit-addressable reasoning），即推理时暴露的语义单元同时也定义了学习阶段比较替代方案并分配信用的位置。我们以 Code-CoT 实例化该原则，它保留图表，将视觉关系表示为可寻址行
- 来源文件：
- /inbox/tom/_candidates/2026-09-03-agent-rag-longcontext-candidates.json