Q*: Improving Multi-step Reasoning for LLMs with Deliberative Planning

The article discusses the challenges faced by Large Language Models (LLMs) in performing multi-step reasoning due to their auto-regressive generation process, which often leads to errors and inconsistencies. To address this, the authors introduce Q*, a versatile framework that guides the decoding process of LLMs with deliberative planning. This is achieved by learning a Q-value model as a heuristic function, which helps LLMs select the most promising next step without the need for fine-tuning for each task, thus avoiding computational overhead and potential performance degradation.

The authors validate the effectiveness of their method through extensive experiments on GSM8K, MATH, and MBPP. They claim that their approach not only improves the performance of LLMs but also reduces the computational load and risk of performance degeneration across different tasks. The Q* framework is presented as a general, versatile, and agile solution for guiding the decoding process in LLMs.

Key takeaways

The paper discusses the issues with Large Language Models (LLMs) such as producing errors, hallucinations, and inconsistent statements during multi-step reasoning.
It introduces Q*, a new framework that guides LLMs decoding process with deliberative planning to mitigate these issues.
Q* uses a Q-value model as a heuristic function to guide LLMs in selecting the most promising next step without needing to fine-tune LLMs for each task.
Experiments on GSM8K, MATH and MBPP have confirmed the effectiveness of the Q* method.

Q*: Improving Multi-step Reasoning for LLMs with Deliberative Planning

Key takeaways

Discussion (0)