Reproduce “CoT-Kinetics: A Theoretical Modeling Assessing LRM Reasoning Process” 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
Jinhe Bi, Danqi Yan, Yifan Wang, Wenke Huang, Haokun Chen, Guancheng Wan, Mang Ye, Xun Xiao and 3 more
Researchers from LMU Munich and collaborators introduce CoT-Kinetics, a theoretical framework that evaluates reasoning quality in Large Reasoning Models by modeling the reasoning process as particle motion through a force field, achieving superior performance across 7 open-source LRMs and 6 diverse benchmarks while enabling fine-grained assessment of reasoning trajectories beyond final answer correctness.
Key insights
- Reasoning quality can be modeled mathematically by treating token state transitions analogously to particle motion in physics
- Both first-order (momentum) and second-order (curvature) semantic transitions are important for capturing reasoning dynamics
- Incorporating model uncertainty through output logit entropy improves assessment accuracy
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.