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

TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs

NeurIPS 2025 · 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 TokenSqueeze, a self-generated-data Long2Short method that combines adaptive-quantile depth selection, KL-constrained step-level rewriting, and a length-aware DPO+SFT objective to cut reasoning token usage while preserving accuracy.

Problem

Long chain-of-thought reasoning LLMs produce very long traces, raising latency and memory; existing Long2Short methods reduce length but often sacrifice accuracy (reasoning oversimplification).

Approach

Builds long/short preference pairs from the model's own samples using an adaptive quantile q = alpha*(1-p) to match reasoning depth to difficulty, rewrites each reasoning step into shorter candidates under a KL constraint, and trains with a composite SFT + length-aware DPO (DPO-L) objective.

Key results

• On MATH500, DeepSeek-R1-Distill-Qwen-7B with TokenSqueeze keeps accuracy at 92.4% (baseline 92.8%) while cutting correct-response length by 51.3% (3638 to 1773 tokens).
• On AIME24 with the 7B model, TokenSqueeze reaches 57.5% accuracy vs 55.5% baseline with 31.6% shorter correct responses.
• On LiveCodeBench with the 7B model, accuracy rises from 31.3% to 35.0% while correct-response length falls 19.2%.

Measured effects reported in the paper

Surprising findings

• Pure DPO substantially degrades accuracy (AIME24 48.3% vs 55.5% baseline) despite reducing length, motivating the SFT anchor.
• Adding correct longer traces as negatives (Q-DYN w/ extra pos) reduced accuracy, suggesting such pairs encourage over-compression.

Limitations

• KL threshold epsilon and other hyperparameters chosen heuristically without systematic tuning.
• Purely offline preference optimization without online feedback.
• Kimi-k1.5 (DPO) and DAST results partly reproduced or taken from other sources, with some table cells missing.

Experimental setup

Datasets: MATH500 AIME24 AIME25 LiveCodeBench

Benchmarks: AIME24 MATH500 AIME25 LiveCodeBench

Metrics: Answer accuracy Len-T Len-A AUC under 32K token budget

Baselines: Kimi-k1.5 (DPO) DAST TrainEffi TokenSkip GPT-4o-mini rewrite DPO SFT

Builds on: DPO Kimi-k1.5 TokenSkip DAST GRPO

Closest papers by meaning

• Reasoning Path Compression: Compressing Generation Trajectories for Efficient LLM Reasoning (NeurIPS 2025 · similarity 0.90)
• Token Assorted: Mixing Latent and Text Tokens for Improved Language Model Reasoning (ICML 2025 · similarity 0.89)
• Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression (ICML 2026 · similarity 0.89)
• Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (ICML 2026 · similarity 0.89)
• VeriThinker: Learning to Verify Makes Reasoning Model Efficient (NeurIPS 2025 · similarity 0.89)
• Efficient Reasoning with Hidden Thinking (ICML 2026 · similarity 0.89)