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

Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Long CoT dynamics and efficiency · 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 Extra-CoT, a three-stage framework (question-aware LongFormer CoT compressor, mixed-ratio SFT with control tokens, and CHRPO hierarchical-reward RL) for extreme-ratio chain-of-thought compression on math reasoning.

Problem

Existing controllable CoT compression methods lose logical fidelity and collapse in accuracy at high compression ratios (20-30% of tokens), limiting efficient reasoning in LLMs.

Approach

A question-aware token-classification compressor (Longformer with global attention on the question) generates formula-atomic, high-fidelity compressed CoT supervision; an SFT model learns ratio control tokens and a policy token, then RL (CHRPO) with hierarchical main and control-head rewards optimizes ratio selection for accuracy at low budgets.

Key results

• On MATH-500 with Qwen3-1.7B, Extra-CoT (CHRPO policy) reaches 64.8% accuracy at 0.27 ActRatio (452 tokens), versus Thinkless at 63.6% with 0.53 ActRatio.
• At fixed ratio 0.2 on MATH-500, Extra-CoT achieves 47.8 Acc@all versus TokenSkip's 23.4 (+24.4 points).
• End-to-end latency on GSM8K drops from 0.7298 s (base) to 0.2254 s (Extra-CoT), a 3.24x speedup.

Surprising findings

• Removing the fixed-ratio SFT cohort roughly doubles token usage on GSM8K (210 to 425), showing the mixed-ratio data is needed for compression control.
• TokenSkip is slower than the base model on MATH-500 in latency despite fewer tokens at its smallest ratio.

Limitations

• Evaluation is limited to mathematical reasoning benchmarks; generalization to other domains is not established beyond MMLU STEM.
• Reliance on GPT-4o and Qwen3-32B for supervision and judging introduces dependence on proprietary or large models.
• Several reported comparisons use the authors' own re-implementations of TokenSkip and Thinkless.

Experimental setup

Datasets: MetaMathQA-395K CAMEL GSM8K MATH-500 AMC2023 SVAMP MultiArith MMLU

Benchmarks: GSM8K MATH-500 AMC2023 SVAMP MultiArith MetaMath-1k MMLU (STEM)

Metrics: Acc@all ActRatio token count latency

Baselines: TokenSkip Thinkless (DeGRPO) LC-Prompt Truncation LLMLingua-2

Builds on: TokenSkip LLMLingua-2 Thinkless GRPO

Closest papers by meaning

• Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.94)
• C3oT: Generating Shorter Chain-of-Thought Without Compromising Effectiveness (AAAI 2025 · similarity 0.93)
• Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · similarity 0.93)
• Efficient Reasoning with Hidden Thinking (ICML 2026 · similarity 0.91)
• Markovian Transformers for Informative Language Modeling (ICLR 2026 · similarity 0.91)
• Reasoning Models Can be Accurately Pruned Via Chain-of-Thought Reconstruction (ICLR 2026 · similarity 0.91)