The Guided Thought Reinforcement (GTR) framework was proposed by researchers from Tsinghua University, Tencent, and Peking University on July 11, 2025. The related research findings were published in a paper. GTR: Guided Thought Reinforcement Prevents Thought Collapse in RL-based VLM Agent Training .

GTR is a simple and scalable framework combining automatic error correction and reinforcement learning, primarily designed to address the "thinking breakdown" problem in Visual Language Model (VLM) agents making multi-step decisions in complex visual environments, which arises from relying solely on outcome rewards. This framework introduces an automatic error corrector to evaluate and improve the agent's reasoning at each step of reinforcement learning, enabling simultaneous training of reasoning and actions without intensive manual point-by-point annotation. Research results show that GTR effectively suppresses thinking breakdown and significantly enhances the performance and generalization ability of models (such as LLaVA-7B) in various visual environments; in complex scenarios such as the 24-point game and embodied tasks, it enables models to achieve a 3 to 5 times higher task success rate than existing state-of-the-art models with a smaller number of parameters.