Faithful Reasoning Using Large Language Models

TL;DR

This paper introduces a method for large language models to perform faithful, multi-step reasoning by chaining reasoning steps with fine-tuned models, resulting in more accurate and interpretable answers.

Contribution

It presents a novel approach that mimics logical reasoning structures using chained calls to specialized models, improving reasoning transparency and accuracy.

Findings

Outperforms baselines on logical deduction tasks

Generates human-interpretable reasoning traces

Improves final answer accuracy in scientific QA

Abstract

Although contemporary large language models (LMs) demonstrate impressive question-answering capabilities, their answers are typically the product of a single call to the model. This entails an unwelcome degree of opacity and compromises performance, especially on problems that are inherently multi-step. To address these limitations, we show how LMs can be made to perform faithful multi-step reasoning via a process whose causal structure mirrors the underlying logical structure of the problem. Our approach works by chaining together reasoning steps, where each step results from calls to two fine-tuned LMs, one for selection and one for inference, to produce a valid reasoning trace. Our method carries out a beam search through the space of reasoning traces to improve reasoning quality. We demonstrate the effectiveness of our model on multi-step logical deduction and scientific question-answering, showing that it outperforms baselines on final answer accuracy, and generates humanly interpretable reasoning traces whose validity can be checked by the user.