<ul><li>Large Reasoning Models (LRMs) show impressive performance in complex tasks by using extended Chain-of-Thought reasoning.</li><li>Test-time scaling methods like prolonging CoT can enhance LRMs' accuracy but lead to decoding overhead due to redundant thinking CoTs.</li><li>TrimR is proposed as a verifier-based, training-free framework for dynamic CoT compression to improve reasoning and scaling efficiency without fine-tuning LRMs or verifiers.</li><li>Empirical evaluations demonstrate TrimR's effectiveness in enhancing inference efficiency on various benchmarks without compromising accuracy, especially for large-batch workloads.</li></ul>

TrimR: Verifier-based Training-Free Thinking Compression for Efficient Test-Time Scaling

Discover more