GeoReform: Reflective Formalization Evolution for Multimodal Geometry Problem Solving
GeoReform raises Qwen3VL-2B’s Geometry3K accuracy from 42.0% to 56.0% by treating diagram formalization as something to optimize.
The paper says multimodal models often fail because extracted geometric facts are redundant, ambiguous, or poorly organized. Its framework runs the full reasoning pipeline, studies failed attempts, and mutates how entities, relations, constraints, and targets are represented. The authors argue the gain comes less from extracting more structure than from presenting the right structure for downstream reasoning. ArXiv · AI/CL/LG's note
The paper says multimodal models often fail because extracted geometric facts are redundant, ambiguous, or poorly organized. Its framework runs the full reasoning pipeline, studies failed attempts, and mutates how entities, relations, constraints, and targets are represented. The authors argue the gain comes less from extracting more structure than from presenting the right structure for downstream reasoning. ArXiv · AI/CL/LG's note
score 4