推理“刹不住车”？新框架让DeepSeek-R1们告别过度思考，已开源

Core Viewpoint - The article discusses the emergence of a new method called Self-Braking Tuning (SBT) aimed at improving the efficiency of reasoning models by preventing them from overthinking and generating unnecessary outputs [3][4][10]. Group 1: Overthinking in Models - As reasoning models like DeepSeek-R1 and OpenAI o1 become more capable, they tend to overthink, which affects efficiency and can lead to errors due to the accumulation of small mistakes [2][3]. - A key challenge is to enable models to know when to stop reasoning, balancing thoroughness with efficiency [3][10]. Group 2: Self-Braking Tuning (SBT) Mechanism - SBT is a lightweight and universal tuning mechanism that integrates seamlessly into existing large models, focusing on reaching correct answers efficiently [4][5]. - The core design includes a braking signal mechanism and multi-task fine-tuning, allowing models to learn when to terminate reasoning without external intervention [5][6][7]. Group 3: Identifying Overthinking - The research team developed a reference standard answer evaluation system to identify redundant reasoning during the process [16]. - Reasoning is divided into two main phases: Foundation Solution (initial answer) and Evolution Solution (further thoughts and validations) [17][18]. - Two core metrics were proposed: reasoning efficiency ratio and overthinking marking ratio, which assess the redundancy in reasoning from structural and linguistic perspectives [20][21][22]. Group 4: Data Construction Strategies - The team created two complementary data construction strategies: SBT-E (Exact) and SBT-D (Dynamic) [24]. - SBT-E uses a unified truncation strategy to structure reasoning paths, helping models distinguish necessary from unnecessary reasoning [25][27]. - SBT-D employs a stepwise adaptive strategy to dynamically adjust reasoning length based on problem complexity, preventing overthinking [28][29]. Group 5: Self-Regulating Brake Strategy - A self-regulating brake strategy was introduced to enhance the model's self-control during reasoning [31]. - Redundant parts of the reasoning process are masked during training, allowing models to focus on key steps without being penalized for the masked content [32][35]. - Natural language prompts are also used to guide models to recognize when they have enough information, reducing unnecessary output [36][37]. Group 6: Performance Evaluation - Extensive experiments on mathematical reasoning benchmarks (AIME, AMC, MATH500, GSM8K) showed significant performance improvements with the SBT framework, particularly in reasoning efficiency [38][39]. - For instance, the Llama-3.1-8B-Instruct model reduced token generation by 62.8% while maintaining an accuracy of 94.1% after applying the SBT-E strategy [40]. - The method demonstrated robustness and generalizability across various model architectures and scales, confirming its value in eliminating redundant reasoning without compromising understanding [41][42].