Research map / Large language models / Reasoning and chain of thought
One-shot Entropy Minimization for Language Model Reasoning
Zitian Gao, Yilong Chen, Haoming Luo, Joey Zhou, Bryan Dai
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes One-shot Entropy Minimization, a fully unsupervised post-training method that minimizes token entropy on a single unlabeled example for about ten gradient steps to improve LM reasoning.
Problem
Improving language model reasoning and generation without labels, and avoiding data contamination when evaluating unsupervised post-training gains.
Approach
Applies entropy minimization using only a single unlabeled example and roughly ten gradient steps, and pretrains a 7B model from scratch on strictly decontaminated data to isolate the effect from pretraining contamination.
Key results
- One-shot EM with a single unlabeled example and about ten gradient steps significantly improves reasoning and generation performance (abstract; specific numbers not stated in available text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add 1-shot entropy minimization | Qwen2.5-Math-7B · 6-benchmark average | average score: 29.2 -> 51.1 | Improved |
| add 1-shot entropy minimization | Deconta-7B-SFT · 6-benchmark average | average score: 52.1 -> 56.6 | Improved |
| add 1-shot entropy minimization on top of RL | Deconta-7B-RL · 6-benchmark average | average score: 59.5 -> 62.3 | Improved |
| add 1-shot entropy minimization | Qwen3-8B-Base · 6-benchmark average | average score: 29.7 -> 34.6 | Improved |
| add 1-shot entropy minimization | Llama-3.1-Nemotron-Nano-8B-v1 · 6-benchmark average | average score: 40.6 -> 44.8 | Improved |
| add variance-based data selection (ablation) | 1-shot EM on Deconta-7B-SFT · 6-benchmark average | average score: 55.0 -> 56.6 | Improved |
| add variance-based data selection (ablation) | 1-shot EM on Deconta-7B-RL · 6-benchmark average | average score: 61.7 -> 62.3 | Improved |
| add variance-based data selection (ablation) | 1-shot EM on Qwen3-8B-Base · 6-benchmark average | average score: 33.9 -> 34.6 | Improved |
| add variance-based data selection (ablation) | 1-shot EM on Qwen2.5-Math-7B · 6-benchmark average | average score: 49.5 -> 51.1 | Improved |
| add variance-based data selection (ablation) | 1-shot EM on Llama-3.1-Nemotron-Nano-8B-v1 · 6-benchmark average | average score: 44.1 -> 44.8 | Improved |
| train EM beyond about 10 steps (ablation) | 1-shot EM · math reasoning | average score | Hurt |
| apply entropy minimization instead of RL | logit distribution (Qwen2.5-Math-7B) · flattened logits | skewness: 0.11 -> 1.54 | Mixed |
Experimental setup
Datasets: single unlabeled example
Closest papers by meaning
- The Unreasonable Effectiveness of Entropy Minimization in LLM Reasoning (NeurIPS 2025 · similarity 0.92)
- Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.89)
- Right Question is Already Half the Answer: Fully Unsupervised LLM Reasoning Incentivization (NeurIPS 2025 · similarity 0.89)
- Break the Block: Dynamic-size Reasoning Blocks for Diffusion Large Language Models via Monotonic Entropy Descent with Reinforcement Learning (ICML 2026 · similarity 0.88)
- DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference (ICLR 2026 · similarity 0.88)
- Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.88)
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.
