Meta AI researchers are advancing a new technique for Large Language Models (LLMs) called “System 2 distillation,” which improves the reasoning capabilities of these models without requiring intermediate steps. The finding holds implications for making models faster and more computationally efficient.
System 1 and System 2 thinking in cognitive science and LLMs: The article draws a parallel between the two modes of thinking in humans – fast and intuitive System 1, and slow and analytical System 2 – and how they relate to LLMs:
- LLMs are usually considered analogous to System 1 thinking, as they can generate text quickly but struggle with tasks requiring deliberate reasoning and planning.
- AI researchers have shown that LLMs can mimic System 2 thinking by prompting them to generate intermediate reasoning steps before providing their final answer, leading to more accurate results for logical reasoning tasks.
Introducing System 2 distillation: Meta AI researchers have developed a technique called “System 2 distillation” that teaches LLMs complex tasks without requiring intermediate steps:
- The process involves prompting the LLM to solve a problem using System 2 techniques, verifying the responses for correctness, discarding the intermediate steps, and fine-tuning the model on the initial question and answer.
- This allows the model to skip the reasoning steps and jump straight to the answer, making the process faster and less computationally expensive.
Evaluating System 2 distillation: The researchers evaluated their method on various reasoning tasks and System 2 prompting techniques using the Llama-2-70B model:
- The results show that System 2 distillation can significantly improve the performance of LLMs on complex reasoning tasks, often matching or exceeding the accuracy of the original System 2 methods while generating responses much faster and with less compute.
- However, the researchers found that LLMs can’t distill all types of reasoning skills into their fast-paced inference mechanism, suggesting that some tasks might always require deliberate reasoning.
Looking ahead: While more research is needed to fully understand the potential and limitations of System 2 distillation, the technique is expected to be a powerful optimization tool for mature LLM pipelines that perform specific tasks at each step:
- Future systems that can distill useful tasks will have more time to spend on reasoning about the tasks they cannot yet do well, just as humans do.
- Distillation will likely play a significant role in making LLMs more efficient and effective in handling complex reasoning tasks.
Meta researchers distill System 2 thinking into LLMs, improving performance on complex reasoning