This paper introduces credit-addressable reasoning for multimodal geometry tasks, aligning the semantic units used during inference with the learning mechanism that trains them. Code-CoT represents visual relations as executable code while preserving diagrams, and CE-GRPO, a reinforcement learning algorithm, assigns credit at fine-grained event boundaries rather than at the trajectory level. Across nine geometry benchmarks the approach improves substantially over baselines like Qwen3-VL-8B, with gains growing as problem complexity increases.
