Phase-Conditioned Reward Shaping for Token-Efficient LLM Reasoning

Extended reasoning can help language models solve complex problems, but long responses often contain both useful computations and redund