Can Small Language Models Help Large Language Models Reason Better?: LM-Guided Chain-of-Thought

TL;DR

This paper presents LM-Guided CoT, a resource-efficient framework where a small language model guides a large one in reasoning tasks, improving accuracy and rationale quality through knowledge distillation and reinforcement learning.

Contribution

The novel framework uses a lightweight LM to guide a large LM in reasoning, optimizing with distillation and RL, and demonstrates improved performance on multi-hop QA benchmarks.

Findings

Outperforms baselines in answer accuracy

Reinforcement learning enhances rationale quality

Effective resource-efficient reasoning guidance

Abstract

We introduce a novel framework, LM-Guided CoT, that leverages a lightweight (i.e., <1B) language model (LM) for guiding a black-box large (i.e., >10B) LM in reasoning tasks. Specifically, the lightweight LM first generates a rationale for each input instance. The Frozen large LM is then prompted to predict a task output based on the rationale generated by the lightweight LM. Our approach is resource-efficient in the sense that it only requires training the lightweight LM. We optimize the model through 1) knowledge distillation and 2) reinforcement learning from rationale-oriented and task-oriented reward signals. We assess our method with multi-hop extractive question answering (QA) benchmarks, HotpotQA, and 2WikiMultiHopQA. Experimental results show that our approach outperforms all baselines regarding answer prediction accuracy. We also find that reinforcement learning helps the model to produce higher-quality rationales with improved QA performance.