Research map / Large language models / Reasoning and chain of thought
Training Chain-of-Thought via Latent-Variable Inference
Du Phan, Matthew Douglas Hoffman, david dohan, Sholto Douglas, Tuan Anh Le, Aaron T Parisi, Pavel Sountsov, Charles Sutton, Sharad Vikram, Rif A. Saurous
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 the control-variate gradient estimator to TRICE (ablation) | TRICE with bootstrapped initialization · BIG-Bench Hard (27-task average) | Greedy-decoding accuracy (%): 67.8 -> 72.8 | Improved |
| Add the control-variate gradient estimator to TRICE (ablation) | TRICE with 3-shot Q-R-A initialization · BIG-Bench Hard (27-task average) | Self-consistency accuracy (%): 75.2 -> 77.6 | Improved |
| Use the human-provided 3-shot Q-R-A initialization instead of bootstrapped initialization (ablation) | TRICE with CV · BIG-Bench Hard (27-task average) | Greedy-decoding accuracy (%): 72.8 -> 76.7 | Improved |
| Add the control-variate gradient estimator to TRICE (ablation) | TRICE with bootstrapped initialization · GSM8K test | Greedy-decoding accuracy (%): 72.8 -> 74.7 | Improved |
| Add the control-variate gradient estimator to TRICE (ablation) | TRICE with bootstrapped initialization · GSM8K test | Self-consistency accuracy (%): 81.5 -> 82.3 | Improved |
| Use non-bootstrapped rather than bootstrapped initialization (ablation) | TRICE with CV · GSM8K test | Greedy-decoding accuracy (%): 74.7 -> 77.7 | Improved |
| Use TRICE instead of rejection sampling | Prompt-tuned CoT model · GSM8K test | Self-consistency accuracy (%): 87.0 -> 86.6 | Hurt |
| Use TRICE instead of STaR | Prompt-tuned CoT model · BIG-Bench Hard (27-task average) | Greedy-decoding accuracy (%): 62.0 -> 72.8 | Improved |
| Use TRICE instead of direct prompt tuning | Prompt-tuned model · GSM8K test | Greedy-decoding accuracy (%): 19.4 -> 77.7 | Improved |
| Initialize the GSM8K rationale memory from a guide prompt containing the answer instead of sampling from the model (ablation) | TRICE · GSM8K | Convergence speed | Hurt |
Closest papers by meaning
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.89)
- Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.89)
- From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning (ICML 2024 · similarity 0.88)
- An Emulator for Fine-tuning Large Language Models using Small Language Models (ICLR 2024 · similarity 0.88)
- QFFT, Question-Free Fine-Tuning for Adaptive Reasoning (NeurIPS 2025 · similarity 0.87)
- Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis (ICLR 2025 · 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.
