SLAT: Segment-Level Adaptive Trimming for Efficient CoT Reasoning

A new reinforcement learning framework called SLAT (Segment-Level Adaptive Trimming) has been proposed to reduce computational overhead in chain-of-thought reasoning. Large reasoning models often generate redundant reasoning steps, a phenomenon known as overthinking, which increases costs without improving accuracy. Existing methods apply uniform length penalties that can suppress useful reasoning. SLAT identifies inefficiency in high-probability segments with low marginal utility and selectively suppresses them. The framework is based on a theoretical characterization of segment suboptimality under a correctness-length trade-off. Empirical results on standard benchmarks show that SLAT reduces reasoning length while maintaining or improving accuracy. The research is detailed in arXiv paper 2605.30832.