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

SLAT: Segment-Level Adaptive Trimming for Efficient CoT 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.

Proposes SLAT, an RL (GRPO-based) method that penalizes long high-probability sliding-window segments in correct reasoning traces, cutting CoT length by about 50% on math benchmarks with comparable accuracy.

Problem

Reasoning models produce overlong chain-of-thought traces (overthinking), and token-uniform length penalties in RL suppress useful reasoning along with redundancy.

Approach

Derives a sufficient condition under which a segment is suboptimal under a correctness-length Lagrangian, which favors long, high-probability segments, and operationalizes it as a sliding-window count of high-joint-probability windows used as a correctness-gated penalty added to the GRPO reward.

Key results

• On DeepSeek-R1-Distill-Qwen-7B, SLAT reaches 66.4 average accuracy vs 66.0 for the original model while cutting average length from 8404 to 4176 tokens (-50%).
• On Qwen3-14B, GRPO+SLAT reduces average length from 5024 to 2677 tokens (-47%) with average accuracy 60.6 vs 60.9 for GRPO.
• On Qwen2.5-Math-7B, SLAT variants trace a better accuracy-length frontier than Truncation, Length-MaxMin, and Length-Mean, with the largest gains on AIME24/25.

Surprising findings

• L1-Max reduces length to about 3144 tokens but drops average accuracy to 61.2, and SLAT outperforms it by 5.2 points at moderate extra length.
• Accuracy on Qwen2.5-Math-7B with the SLAT w=7, lambda=0.001 setting is 46.2 vs 48.8 for plain GRPO, showing some accuracy cost in that setup.

Limitations

• Relies on a sliding-window threshold heuristic (w, tau, lambda) and the paper admits this is not fully adaptive.
• Evaluation is limited to mathematical reasoning benchmarks plus MMLU-STEM and GPQA; experiments are on 7B and 14B models only.
• Baseline numbers for Length-MaxMin and Length-Mean were tuned over only two lambda values due to compute limits.

Experimental setup

Datasets: DAPO-Math-17k

Benchmarks: MATH500 OlympiadBench AMC23 AIME24 AIME25 MMLU-STEM GPQA-Diamond

Metrics: avg@k accuracy average CoT token length

Baselines: LC-R1 TLMRE AdaptThink DAST L1-Max Vanilla Truncation Length-MaxMin Length-Mean

Builds on: GRPO DeepSeek-R1 DAPO Kimi-k1.5 length reward TLMRE

Closest papers by meaning

• Demystifying Long Chain-of-Thought Reasoning (ICML 2025 · similarity 0.91)
• Ada-R1: Hybrid-CoT via Bi-Level Adaptive Reasoning Optimization (NeurIPS 2025 · similarity 0.90)
• Short Chains, Deep Thoughts: Balancing Reasoning Efficiency and Intra-Segment Capability via Split-Merge Optimization (ICML 2026 · similarity 0.89)
• OS-Pruner: Pruning Chains-of-Thought of Reasoning Models via Optimal Stopping (NeurIPS 2026 · similarity 0.89)
• Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.89)
• ShorterBetter: Guiding Reasoning Models to Find Optimal Inference Length for Efficient Reasoning (NeurIPS 2025 · similarity 0.89)