Reproduce “Think 360°: Evaluating the Width-centric Reasoning Capability of MLLMs Beyond Depth” 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
Mingrui Chen, Hexiong Yang, Haogeng Liu, Huaibo Huang, Ran He
A new multimodal benchmark, Think360, assesses the "width-centric" reasoning of MLLMs, addressing a gap in evaluations that typically focus on sequential "reasoning depth." The research reveals that even top models like Gemini-2.5-pro exhibit limitations in broad exploration, constraint management, and backtracking, with the proposed Tree-of-Thought width score correlating more strongly with final accuracy than its depth counterpart.
Key insights
- "Reasoning width," defined as broad exploratory search and multi-constrained optimization, is a crucial dimension of intelligence distinct from sequential "reasoning depth."
- Advanced MLLMs, including top performers, demonstrate significant deficiencies in width-centric reasoning, often failing due to insufficient exploration space, inadequate constraint recognition, and poor backtracking.
- The proposed ToT-Width score shows a stronger correlation with final answer accuracy on the Think360 benchmark (ρ = 0.9218) than ToT-Depth, emphasizing the importance of broad exploration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.