Reward shaping cuts reasoning trace length by 65% in hybrid language models
Researchers found that tailored reinforcement-learning rewards can reactivate a dormant reasoning pathway in hybrid language models, shortening chain-of-thought traces by roughly two-thirds. The reshaped reward signal reportedly restores an internal reasoning mode the models already possessed, rather than teaching entirely new behavior. At the 9B scale, trace termination rose from 45.9% to 93.9% and mean length fell from 2,494 to 868 tokens, while answer accuracy was largely preserved. The results were demonstrated only on specific Qwen3 and StartLux-Decision architectures and their benchmarks, so transfer to other models and domains remains unclear. Publicly released LoRA adapters for the Qwen3 models reportedly implement the approach without retraining the base model.
