Research map / Probabilistic and scientific modeling / Molecules and proteins
What Makes Effective Supervision in Latent Chain-of-Thought: An Information-Theoretic Analysis
Xinghao Chen, Chak Tou Leong, Guo Wenjin, Jian Wang, Wenjie Li, Xiaoyu Shen
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 |
|---|---|---|---|
| add generative reconstruction space supervision (ablation) | PS-Latent · GSM8K-Aug (GPT-2) | max accuracy (%): 31.2 -> 34.9 | Improved |
| add geometric compression (MSE) space supervision (ablation) | PS-Latent · GSM8K-Aug (GPT-2) | max accuracy (%): 31.2 -> 23.2 | Hurt |
| add progressive trajectory supervision with explicit suffix scaffold | OS-Progressive · GSM8K-Aug (GPT-2) | max accuracy (%): 18.7 -> 31.2 | Improved |
| add generative reconstruction under outcome supervision (ablation) | OS-Latent · GSM8K-Aug (GPT-2) | max accuracy (%): 18.3 -> 18.2 | No clear effect |
| add geometric compression under outcome supervision (ablation) | OS-Latent · GSM8K-Aug (GPT-2) | max accuracy (%): 18.3 -> 13.1 | Hurt |
| increase granularity from g=1 to token-level (ablation) | PS-GR · GSM8K-Aug (GPT-2) Stage 2 E5 | accuracy (%): 31.6 -> 42.4 | Improved |
| increase granularity from g=1 to g=3 (ablation) | PS-GR · GSM8K-Aug (GPT-2) Stage 2 E5 | accuracy (%): 31.6 -> 36.1 | Improved |
| increase granularity from g=1 to g=2 (ablation) | PS-GR · GSM8K-Aug (GPT-2) Stage 2 E5 | accuracy (%): 31.6 -> 35.1 | Improved |
| use extended granularity g=2 in PS-GR | PS-GR · GSM8K-Aug (GPT-2) | max accuracy (%): 34.9 -> 39.1 | Improved |
| reset optimizer at each stage transition (ablation) | PS-Latent · GSM8K-Aug (GPT-2) | gradient dynamics | Improved |
| add process supervision (PS-Latent) over outcome supervision | OS-Latent · ProntoQA (GPT-2) | accuracy (%): 78.60 -> 97.40 | Improved |
| add latent states under outcome supervision | OS-No-CoT · LLaMA-3.2-1B GSM8K-Aug | best accuracy (%): 32.98 -> 33.81 | No clear effect |
Closest papers by meaning
- Why Struggle with Continuous Latents? Interpretable Discrete Latent Reasoning via Rendered Compression (NeurIPS 2026 · similarity 0.90)
- Geometric Latent Reasoning Induces Shorter Generations in LLMs (NeurIPS 2026 · similarity 0.89)
- Tracing the Traces: Latent Temporal Signals for Efficient and Accurate Reasoning (ICLR 2026 · similarity 0.88)
- KaVa: Latent Reasoning via Compressed KV-Cache Distillation (ICLR 2026 · similarity 0.88)
- A Formal Comparison Between Chain of Thought and Latent Thought (ICML 2026 · similarity 0.88)
- 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.
