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

Accordion-Thinking: Self-Regulated Step Summaries for Efficient and Readable LLM Reasoning

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.

Trains LLMs with SFT plus RL to periodically write step summaries and discard earlier detailed reasoning (Fold mode), closing the accuracy gap with full-context reasoning while cutting peak KV-cache tokens by about half.

Problem

Long chain-of-thought reasoning causes linear KV-cache growth and quadratic attention cost, and prior compression methods rely on heuristic eviction or fixed chunking that hurts readability and is not learned.

Approach

The model is trained to emit <step>-delimited detailed reasoning followed by a summary; in Fold mode, detailed segments are dropped from context after each summary, and RL with a correctness-only reward (shared across steps) incentivizes lossless summaries. A mixed Fold/Unfold RL regime is used.

Key results

• On Qwen2.5-Math-7B, Mix-RL in Fold mode reaches 52.8 macro Avg@32, matching Unfold mode (52.8) and above Unfold-RL (52.2).
• Fold-RL and Mix-RL reduce peak token length by roughly 50-60% versus Unfold-RL on five math benchmarks (e.g., AIME24 10.3k to 4.4k on Qwen2.5-Math-7B).
• On a 48GB single-GPU vLLM setup, Mix-RL on the 4B model reaches 5888 tokens/s vs 1483 for Unfold-RL (about 3x).

Surprising findings

• Fold mode initially lags Unfold by about 27-42 points during RL, but the gap vanishes as training proceeds.
• SFT-only Cold-Start model drops 2.8 (7B) and 4.9 (4B) points when switching to Fold mode.

Limitations

• Experiments limited to two base models and mainly math benchmarks, with OOD results only on ARC-C and GPQA-D.
• Relies on a teacher LLM (DeepSeek-V3.2) to rewrite traces and as judge; human evaluation covers only 20 summaries.

Experimental setup

Datasets: OpenR1-Math-220k openr1-math-46k OpenR1-45K ARC-Challenge GPQA-Diamond

Benchmarks: MATH-500 OlympiadBench Minerva Math AIME24 AIME25 AMC ARC-Challenge GPQA-Diamond

Metrics: Pass@1 (Avg@32) peak token length throughput (tokens/s) accuracy gap

Baselines: Zero-RL Cold-Start Unfold-RL H2O LightThinker DeleThink

Builds on: GRPO Dr. GRPO InftyThink LightThinker DeleThink

Closest papers by meaning

• Making Slow Thinking Faster: Compressing LLM Chain-of-Thought via Step Entropy (ICLR 2026 · similarity 0.91)
• Think Silently, Think Fast: Dynamic Latent Compression of LLM Reasoning Chains (NeurIPS 2025 · similarity 0.90)
• StreamingThinker: Large Language Models Can Think While Reading (ICLR 2026 · similarity 0.89)
• A State-Transition Framework for Efficient LLM Reasoning (ICLR 2026 · similarity 0.89)
• MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning (NeurIPS 2026 · similarity 0.89)
• Efficient Reasoning with Hidden Thinking (ICML 2026 · similarity 0.89)