Bi-mode annealing was proposed by Tencent Hunyuan team and Chinese Academy of Sciences Automation in August 2025. The relevant research results were published in the paper "R-4B: Incentivizing General-Purpose Auto-Thinking Capability in MLLMs via Bi-Mode Annealing and Reinforce Learning".

Dual-mode annealing aims to train a model that is naturally capable of both thinking and thinking outside of the model in a general domain. After the annealing phase, the model's subsequent automatic thinking training in the general domain will lay a solid foundation.