Reproduce “Unsupervised Visual Chain-of-Thought Reasoning via Preference Optimization” 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
Kesen Zhao, Beier Zhu, Qianru Sun, Hanwang Zhang
A framework for Unsupervised Visual Chain-of-Thought (UV-CoT) reasoning enables Multimodal Large Language Models (MLLMs) to perform image-level reasoning by automatically generating preference data and using a novel Score-DPO, surpassing supervised methods and demonstrating robust generalization across diverse visual tasks.
Key insights
- Explicit image-level visual CoT is crucial for complex multimodal reasoning, consistently outperforming text-only CoT approaches by dynamically integrating visual cues.
- Unsupervised generation of visual preference data, combined with a refined preference optimization technique (sDPO), can match or exceed the performance of supervised visual CoT methods.
- Evaluating visual regions based on their cumulative impact on the entire reasoning chain, not just immediate relevance, is vital for training MLLMs to construct effective visual CoT.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.