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

Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency

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 step entropy to identify redundant CoT steps, prunes 80% of low-entropy steps into [SKIP] tokens, and trains models via SFT plus GRPO to generate compressed reasoning autonomously.

Problem

Long chain-of-thought reasoning in large reasoning models is verbose and redundant, raising inference cost, and there is no principled way to identify which reasoning steps can be removed.

Approach

Define step entropy as the length-normalized sum of token-level entropies within a reasoning step, prune the lowest-entropy steps by replacing them with [SKIP] tokens, and train the model with SFT followed by GRPO using a composite reward (correctness, skip ratio, skip number, response length) so it learns to skip steps at inference.

Key results

• Pruning 80% of lowest-entropy steps for DeepSeek-R1-7B keeps Math500 accuracy at 88.17% while cutting thinking tokens by 29.7%, and AIME 2024 tokens by 36.3%.
• SFT+GRPO on DeepSeek-R1-7B reduces AIME 2024 tokens by 57% (63.33% to 57.14% accuracy) and GSM8K accuracy rises from 78.54% to 79.15% with 44% fewer tokens.
• Compared with prior compression methods on Math500, the method matches R1-Compress's 3.2% accuracy drop while reducing tokens by 35.0% versus 20.3%.

Surprising findings

• Pruning low-entropy steps while keeping high-entropy ones preserves accuracy, whereas pruning high-entropy steps degrades it immediately.
• Using only correctness and skip-ratio rewards in GRPO collapses performance (Math500 accuracy drops to 51%) and increases token usage.

Limitations

• Step segmentation relies on newline delimiters generated by the model, which may limit generality.
• Validation is mainly on math and select MMLU domains; no multimodal or open-ended evaluation.
• Fixed 80% pruning ratio may not be optimal across models and problem types.

Experimental setup

Datasets: DeepScaleR OpenR1-Math GSM8K MATH500 AIME 2024 AIME 2025 MMLU

Benchmarks: GSM8K MATH500 AIME 2024 AIME 2025 MMLU

Metrics: Pass@1 accuracy average thinking tokens

Baselines: Prompt-BeConcise LC-Prompt CoT-Valve TokenSkip R1-Compress

Builds on: GRPO CoT-Valve TokenSkip R1-Compress

Closest papers by meaning

• Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression (ICML 2026 · similarity 0.94)
• Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · similarity 0.92)
• When More is Less: Understanding Chain-of-Thought Length in LLMs (ICLR 2026 · similarity 0.92)
• Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.92)
• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)
• C3oT: Generating Shorter Chain-of-Thought Without Compromising Effectiveness (AAAI 2025 · similarity 0.91)