What is Auto Think
Auto Think is the KwaiCoder-AutoThink-preview automatic thinking large model open-sourced by Kuaishou's Kwaipilot team. The model conducts in-depth research on the "overthinking" problem existing in deep thinking large models and proposes a brand-new automatic thinking model training paradigm. Based on the traditional reinforcement learning algorithm (GRPO), it proposes a reinforcement learning method with process supervision, Step-SRPO, to further improve the model's performance in complex tasks. The model integrates "thinking" and "non-thinking" capabilities and has the ability to automatically switch thinking modes according to question difficulty. Through this thinking mode training, the model achieves performance improvements on multiple "thinking" and "non-thinking" evaluation benchmarks. Among them, on some code and math tasks, the model score under automatic thinking mode increases by as much as about 20 points.
Main Features of Auto Think
- Automatic switching of thinking modes: The model integrates "thinking" and "non-thinking" capabilities and automatically switches thinking modes according to question difficulty. For simple questions, the model adopts a "fast thinking" mode and directly gives the answer, avoiding unnecessary complex reasoning processes; for complex questions, it switches to a "slow thinking" mode for deep reasoning and analysis, solving the problem more accurately.
-
Improved efficiency and performance: The ability to automatically switch thinking modes enables the model to achieve performance improvements on multiple "thinking" and "non-thinking" evaluation benchmarks. On some code and math tasks, the model score under automatic thinking mode increases by as much as about 20 points.
Technical Principles of Auto Think
- Minimal prompt intervention: Through an Ellipsis Prompt with an added ellipsis, the model's ability to randomly switch thinking modes is activated. This prompt structure is simple and effective, can guide the model to switch between different thinking modes, and provides a foundation for subsequent reinforcement learning training.
- Multi-stage reinforcement learning
-
First stage: Enable the model to stably exhibit two thinking modes, fast and slow. Among them, "fast thinking" is used to solve simple problems, while complex problems use "slow thinking". The goal of this stage is to enable the model to preliminarily judge according to question difficulty and select the corresponding thinking mode.
-
Second stage: Optimize fast and slow thinking behavior and improve the ability to answer correctly under both modes. Through training at this stage, the model can handle problems more accurately under different thinking modes, improving its overall performance.
-
Third stage: Refine the chain-of-thought output of fast and slow thinking. After training at this stage, the model no longer randomly decides whether to think deeply, and can autonomously choose the thinking mode according to question difficulty, achieving a more efficient and more precise reasoning process.
-
Project Address of Auto Think
- HuggingFace Model Library：https://huggingface.co/Kwaipilot/KwaiCoder-AutoThink-preview
Application Scenarios of Auto Think
-
Video generation: Auto Think's automatic thinking capability can further optimize the video generation process, making the generation of video content better fit the needs of different difficulty and complexity.
-
Copywriting: Auto Think can automatically switch thinking modes according to question difficulty, providing more efficient and more precise ideas and methods for copywriting.
-
Intelligent customer service: Auto Think's automatic thinking capability enables it to quickly and accurately respond based on the complexity of the question when interacting with users, improving the user experience.
-
Precise Search: Auto Think's automatic thinking capability can further optimize search results, providing more precise information that better meets user needs.
-
Personalized Recommendations: Auto Think can automatically switch thinking modes based on users' personalized needs, providing more precise recommendation results.