SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning

2026-05-29

AI Summary

This work addresses the excessive computational cost of large language models in chain-of-thought reasoning, which often stems from structural redundancy. Existing compression approaches lack fine-grained awareness of reasoning passages and frequently discard useful content. To overcome this limitation, the paper proposes a reinforcement learning–based, segment-level adaptive pruning framework that, for the first time, theoretically characterizes the suboptimality of reasoning segments. It identifies high-probability redundant segments with low marginal utility and selectively suppresses them. Moving beyond conventional token-level uniform compression, the method achieves an improved trade-off between reasoning length and accuracy. Experiments demonstrate that the approach reduces reasoning length by 50% on standard benchmarks while maintaining competitive accuracy, significantly outperforming existing methods along the accuracy–efficiency Pareto frontier.

Abstract

Recent advances in Large Reasoning Models have significantly improved chain-of-thought (CoT) capabilities via reinforcement learning (RL). However, generated reasoning chains frequently suffer from structural redundancy (i.e., \emph{overthinking}), incurring high computational overhead without improving answer correctness. Existing mitigation strategies typically rely on token-uniform length penalties, which provide coarse, segment-agnostic pressure toward shorter outputs and can inadvertently suppress useful reasoning alongside redundancy. To address this, we demonstrate that inefficiency concentrates in high-probability segments with low marginal utility. We derive a theoretical characterization of segment suboptimality under the correctness-length trade-off objective and propose \textsc{SLAT} (Segment-Level Adaptive Trimming), an RL framework that selectively suppresses redundant segments based on this criterion. Empirical results on standard benchmarks indicate that \textsc{SLAT} establishes a superior accuracy-efficiency Pareto frontier, reducing reasoning length by $50\%$ relative to uncompressed baselines while maintaining competitive accuracy. Overall, our results suggest that theoretically grounded, segment-aware trimming is a promising direction for efficient CoT reasoning in large language models.

Problem

• chain-of-thought reasoning
• structural redundancy
• computational overhead
• length penalty
• reasoning efficiency

Innovation

• Segment-Level Adaptive Trimming
• Chain-of-Thought Reasoning
• Reinforcement Learning
• Efficiency-Accuracy Trade-off
• Redundancy Suppression