Meta researchers distill System 2 thinking into LLMs, improving performance on complex reasoning
What are "System 2" techniques in LLMs?
"System 2" techniques in LLMs are prompting schemes that enhance the reasoning capabilities of LLMs by forcing them to generate intermediate steps toward solving a problem. These techniques, such as "Chain of Thought," require conscious effort and are used for complex problem-solving, making LLM applications slow and computationally expensive.
What is "System 2 distillation" as proposed by Meta FAIR?
System 2 distillation is a technique developed by Meta FAIR that enhances the reasoning capabilities of large language models (LLMs) on complex tasks without sacrificing speed or efficiency5. It distills the knowledge gained from LLMs' own System 2 reasoning capabilities into their fast-paced System 1 generation, allowing them to tackle complex tasks more effectively while maintaining speed and computational efficiency5.
How do System 1 and System 2 thinking differ?
System 1 thinking is fast, intuitive, and automatic, used for recognizing patterns, making quick judgments, and understanding familiar symbols1. System 2 thinking is slow, deliberate, and analytical, required for complex problem-solving, such as manipulating abstract symbols, solving mathematical equations, or planning a trip1.