Reproduce “Learn to Think: Improving Multimodal Reasoning through Vision-Aware Self-Improvement Training” 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
Qihuang Zhong, Liang Ding, Wenjie Xuan, Juhua Liu, Bo Du, Dacheng Tao
Wuhan University and collaborators developed VISTA, a self-improvement training framework for Multimodal Large Language Models (MLLMs), which enhances multimodal reasoning by mitigating data imbalance and language prior bias. VISTA achieves an average performance gain of up to +13.66% on reasoning tasks and improves out-of-distribution generalization by addressing visual hallucination.
Key insights
- Reusing partially correct reasoning prefixes, identified through a critical token detection mechanism, is more effective than simple re-sampling or providing more trials for generating solutions to challenging queries.
- Quantifying attention allocation to visual tokens can reliably identify and filter out visually hallucinated reasoning steps, even when the final answer is correct.
- Balancing data collection for hard samples and ensuring visual grounding in reasoning trajectories is crucial for achieving sustained performance improvements and reducing model collapse during iterative self-improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.