On the Optimal Reasoning Length for RL-Trained Language Models
This work addresses efficiency and performance trade-offs in RL-trained language models, which is an incremental improvement for researchers and practitioners in AI.
The study investigated the optimal reasoning length for RL-trained language models to balance efficiency and performance, finding that length penalties can hinder reasoning acquisition while properly tuned length control improves efficiency for models with strong prior reasoning.
Reinforcement learning substantially improves reasoning in large language models, but it also tends to lengthen chain of thought outputs and increase computational cost during both training and inference. Though length control methods have been proposed, it remains unclear what the optimal output length is for balancing efficiency and performance. In this work, we compare several length control methods on two models, Qwen3-1.7B Base and DeepSeek-R1-Distill-Qwen-1.5B. Our results indicate that length penalties may hinder reasoning acquisition, while properly tuned length control can improve efficiency for models with strong prior reasoning. By extending prior work to RL trained policies, we identify two failure modes, 1) long outputs increase dispersion, and 2) short outputs lead to under-thinking.