Reproduce “What MLLMs Learn about When they Learn about Multimodal Reasoning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiwan Chung, Neel Joshi, Pratyusha Sharma, Youngjae Yu, Vibhav Vineet
MathLens separates diagram-based perception, text-based reasoning, and combined-task failures using aligned problem versions and visual-fact probes. Across public models, textual supervised fine-tuning is associated with reasoning gains, while multimodal reinforcement learning is associated with improved perceptual grounding and consistency. For VL-Rethinker-7B, consistency rises from 31.0% to 37.3% after multimodal reinforcement learning. Comparisons are noncausal and geometry-focused.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.