Research map / Probabilistic and scientific modeling / Causal discovery
Dynamics Within Latent Chain-of-Thought: An Empirical Study of Causal Structure
Zirui Li, Xuefeng Bai, Kehai Chen, Yizhi LI, Jian Yang, Chenghua Lin, Min Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| intervene on a single latent step by zeroing it (varying which step) (ablation) | Coconut and CODI · GSM8K and CommonsenseQA | flip rate | Mixed |
| switch dataset from CommonsenseQA to GSM8K (ablation) | Coconut and CODI · flip rate under zero intervention | flip rate: below ~0.1 -> around 0.1-0.2 or higher | Hurt |
| use Coconut instead of CODI at matched backbone | latent CoT · GSM8K | flip rate | Hurt |
| use a stronger backbone (ablation) | Coconut and CODI · GSM8K and CommonsenseQA | flip rate | Improved |
| use latent steps instead of explicit CoT steps | influence structure · GSM8K | locality of influence: CoT-SFT locality all >= 0.6 | Hurt |
| add more latent steps (increase k up to 6) (ablation) | Coconut and CODI · GSM8K | cumulative solved fraction S(k) | Improved |
| replace zero intervention with other operators (mean, Gaussian) (ablation) | CODI Llama3-1B · GSM8K and CommonsenseQA | step-wise flip profile | No clear effect |
| use step-aligned supervision (Sim-CoT) instead of Coconut/CODI | latent CoT · GSM8K | influence graph connectivity | Improved |
| use probe readout instead of teacher-forced readout (ablation) | latent CoT · StrategyQA | superposition score | Improved |
Closest papers by meaning
- A Formal Comparison Between Chain of Thought and Latent Thought (ICML 2026 · similarity 0.90)
- Thought Branches: Interpreting LLM Reasoning Requires Resampling (ICLR 2026 · similarity 0.90)
- Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.88)
- Bounds of Chain-of-Thought Robustness: Reasoning Steps, Embed Norms, and Beyond (ICLR 2026 · similarity 0.88)
- Unveiling the Entropy Dynamics of Chain-of-Thought Reasoning (ICML 2026 · similarity 0.87)
- How Far Ahead Do LLMs Plan? Uncovering the Latent Horizon in Chain-of-Thought Reasoning (ICML 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
