Reproduce “Dynamics Within Latent Chain-of-Thought: An Empirical Study of Causal Structure” 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
Zirui Li, Xuefeng Bai, Kehai Chen, Yizhi Li, Jian Yang, Chenghua Lin, Min Zhang
The paper studies hidden reasoning by treating internal states as variables in a structural causal model and overwriting one state at a time to measure effects on later outputs. It compares Coconut and CODI with explicit chain-of-thought baselines in six-step runs, finding less-local latent influence, non-adjacent effects, and readout-dependent commitment, limited to the studied models, tasks, and protocol.
Key insights
- Latent steps are not homogenous; they exhibit heterogeneous causal leverage, with specific steps contributing disproportionately to the final decision and exhibiting varying decision instability.
- Information flow in latent CoT is often non-local and skip-dominant, diverging from the sequential, chain-like propagation characteristic of explicit CoT.
- A "readout gap" exists where the model's output-level commitment (bias towards a single answer) can occur earlier than representational commitment (retention of alternative modes in richer latent states).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.