SHAPE of Chain-of-Thought in Math Reasoning

Recent research introduces SHAPE and SCAFFOLD to enhance LLM reasoning. SHAPE is a framework analyzing Chain-of-Thought trajectories in mathematical reasoning through semantic spaces and heuristics, focusing on mathematical interpretations and specific actions. Meanwhile, SCAFFOLD provides a large-scale structured dataset of computer science research figures. It pairs architecture drawings and flowcharts with captions, questions, and Chain-of-Thought reasoning traces, aiming to train vision-language models to better understand complex technical diagrams and their associated context.