Recently, a research group hailing from the Institute of Software at the Chinese Academy of Sciences has put forward an innovative reinforcement fine-tuning framework. This framework, termed Learning to Think (L2T), is grounded in information theory and is specifically designed to optimize large language models (LLMs) when tackling complex reasoning tasks. Within this framework, a dense process reward mechanism is introduced. This mechanism serves to assess the information gain attained in each reasoning round. Additionally, the framework adopts an enhanced GRPO algorithm strategy for model optimization, with the overarching goal of striking a balance between the effectiveness and efficiency of reasoning.

Experimental results clearly show that L2T consistently delivers stable performance enhancements across foundational models of varying scales. When compared to outcome-based reward methods, L2T leads to an accuracy boost of over 3.2% and doubles the token efficiency.