Can Large Language Models Reason Robustly with Noisy Rationales?

Zhanke Zhou, Rong Tao, Jianing Zhu, Yiwen Luo, Zengmao Wang, Bo Han

Keywords: Large Language Models, Noisy Rationales, Chain-of-Thought.

TL;DR: We investigate the under-explored robustness of LLMs' reasoning with noisy rationales—irrelevant or inaccurate reasoning steps.

Abstract: This paper investigates an under-explored challenge in large language models (LLMs): their reasoning with noisy rationales—irrelevant or inaccurate reasoning steps—despite advancements in in-context learning and chain-of-thought strategies. We construct the NoRa dataset, specifically designed to evaluate LLMs' robustness to noisy rationales, based on which we reveal a widespread vulnerability among LLMs to such noise, with limited efficacy from existing robust methods.

Submission Number: 56