Reproduce “Fast-Slow Thinking GRPO for Large Vision-Language Model 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
Wenyi Xiao, Leilei Gan
The FAST-GRPO framework enables Large Vision-Language Models to dynamically adapt their reasoning depth based on question complexity, achieving state-of-the-art accuracy with over 10% average improvement over base models while reducing token usage by 32.7% to 67.3%. This approach intelligently balances verbose and concise thinking for multimodal reasoning tasks.
Key insights
- Reasoning length and accuracy in Large Vision-Language Models are not strictly coupled, allowing for significant manipulation of response length with only modest changes in accuracy.
- Multimodal question difficulty implicitly influences reasoning length, providing a valuable signal for dynamically adapting an LVLM's cognitive effort.
- The inherent structure of the GRPO gradient's length normalization term naturally biases the model towards longer outputs for challenging or incorrect responses and shorter ones for simpler, correct answers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.