Reproduce “Interpreting Latent CoT Reasoning as Dynamical Systems” 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
Sabari Iyyappan Duraipandian, Shreya Sanjay Boyane, Manju Nagesh, Jerome Francis, Archana Vaidheeswaran, Kevin Zhu
Researchers from Algoverse AI Research, San Jose State University, Worcester Polytechnic Institute, and George Mason University applied dynamical systems theory to interpret latent Chain-of-Thought (CoT) reasoning. The analysis revealed distinct dynamic profiles and geometric structures for CODI (convergent, classification-like) and COCONUT (exploratory, computation-like) models, demonstrating how supervised training enhances stability without altering fundamental dynamics.
Key insights
- Latent CoT models exhibit fundamentally different stability types and cognitive strategies: CODI operates as a stable attractor system (classification-like), while COCONUT demonstrates unstable, expanding dynamics (computation-like).
- SIM-CoT supervision enhances latent stability for both models by enforcing structured semantics, leading to deeper convergence in CODI and suppressed exploratory mid-chain divergences in COCONUT, without fundamentally altering their core dynamics.
- The geometric structures of latent trajectories, such as CODI's two-lobe patterns and COCONUT's 'butterfly' patterns, directly reflect the underlying training objectives and reasoning mechanisms of each model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.