AI Paper Atlas

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

NeurIPS 2023 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 2 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add the control-variate gradient estimator to TRICE (ablation)TRICE with bootstrapped initialization · BIG-Bench Hard (27-task average)Greedy-decoding accuracy (%): 67.8 -> 72.8Improved
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 -> 77.6Improved
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 -> 76.7Improved
Add the control-variate gradient estimator to TRICE (ablation)TRICE with bootstrapped initialization · GSM8K testGreedy-decoding accuracy (%): 72.8 -> 74.7Improved
Add the control-variate gradient estimator to TRICE (ablation)TRICE with bootstrapped initialization · GSM8K testSelf-consistency accuracy (%): 81.5 -> 82.3Improved
Use non-bootstrapped rather than bootstrapped initialization (ablation)TRICE with CV · GSM8K testGreedy-decoding accuracy (%): 74.7 -> 77.7Improved
Use TRICE instead of rejection samplingPrompt-tuned CoT model · GSM8K testSelf-consistency accuracy (%): 87.0 -> 86.6Hurt
Use TRICE instead of STaRPrompt-tuned CoT model · BIG-Bench Hard (27-task average)Greedy-decoding accuracy (%): 62.0 -> 72.8Improved
Use TRICE instead of direct prompt tuningPrompt-tuned model · GSM8K testGreedy-decoding accuracy (%): 19.4 -> 77.7Improved
Initialize the GSM8K rationale memory from a guide prompt containing the answer instead of sampling from the model (ablation)TRICE · GSM8KConvergence speedHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy