arXiv: TRiMS — RL-Based Dynamic Reasoning Length Compression Cuts Chain-of-Thought Token Overhead

arXiv·medium signal

TRiMS (Real-Time Tracking of Minimal Sufficient Length) uses reinforcement learning to enforce minimal-length reasoning chains at inference time, dynamically halting token generation once sufficient reasoning is completed. Unlike static distillation approaches, TRiMS adapts per-query and eliminates 'reasoning theater' — unnecessarily long CoT sequences that inflate cost without improving outputs. For agent pipelines running reasoning models at scale, this directly targets the highest per-query token cost driver.