Research map / Large language models / Reasoning and chain of thought

Rethinking Chain-of-Thought from the Perspective of Self-Training

ICML 2025 · Large language models › Reasoning and chain of thought · approach: Chain-of-thought training and analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 a CoT framework inspired by self-training's entropy-reduction view, combining a task-specific prompt selected by semantic entropy with an adaptive reasoning iteration module (entropy-based stopping, Jaccard-diversified re-prompting), evaluated with GPT-3.5-turbo on 10 reasoning datasets.

Problem

Chain-of-thought reasoning suffers from over-reasoning (correct early answers flipping under further iterations) and high similarity between consecutive reasoning rounds, while initial prompts are generic.

Approach

Frames iterative CoT as uncertainty (semantic entropy) minimization analogous to self-training, and uses it to (i) pick a task-specific zero-shot prompt with lowest mean semantic entropy, (ii) stop iterating once entropy falls below a threshold, and (iii) switch to a divergence-promoting prompt with Jaccard-based resampling when iterating further.

Key results

• Zero-shot setting: TSP+ARI reaches 83.7% average accuracy over 10 datasets versus 79.5% for Zero-Shot-CoT + SC (+4.2 points).
• Few-shot setting on MultiArith and GSM8K: adding ARI raises Few-Shot-CoT + SC average from 85.5% to 87.4%.
• On AQuA, adaptive iteration reaches 71.3% by iteration 5 while fixed iteration peaks at 67.3% and drops to 62.6%; adaptive run takes 4h18m vs 6h29m at iteration 5.

Measured effects reported in the paper

Surprising findings

• Fixed iterative CoT degrades accuracy with more rounds on AQuA (peak 67.3% then 62.6%), which the authors attribute to over-reasoning.
• Commonsense tasks gain little from deeper reasoning, performing near the zero-shot level, which the authors attribute to dependence on pretrained knowledge.

Limitations

• Relies on a single model (GPT-3.5-turbo-0125) for all experiments.
• Gains on commonsense tasks are weak, and the paper attributes this to reliance on prior knowledge.
• Theoretical analysis of self-training uses a simplified 2D Gaussian mixture model and does not directly prove CoT behavior.

Experimental setup

Datasets: MultiArith GSM8K SingleEq AddSub AQuA SVAMP StrategyQA CommonsenseQA LastLetters CoinFlip

Benchmarks: MultiArith GSM8K SingleEq AddSub AQuA SVAMP StrategyQA CommonsenseQA LastLetters CoinFlip

Metrics: accuracy time cost semantic entropy

Baselines: Zero-Shot Zero-Shot-CoT Zero-Shot-CoT + SC RE2 RE2 + SC Contrastive-CoT Few-Shot-CoT + SC Fixed iterative reasoning

Builds on: zero-shot chain-of-thought self-consistency semantic entropy RE2 Contrastive-CoT

Closest papers by meaning

• Unlocking the Capabilities of Thought: A Reasoning Boundary Framework to Quantify and Optimize Chain-of-Thought (NeurIPS 2024 · similarity 0.92)
• Understanding Chain-of-Thought in LLMs through Information Theory (ICML 2025 · similarity 0.91)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)
• SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.91)
• CoT Vectors: Transferring and Probing the Reasoning Mechanisms of LLMs (ICLR 2026 · similarity 0.91)
• The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.91)