Abstract

The reasoning abilities of Large Language Models (LLMs) remain a topic of considerable interest and debate. Among the original papers arguing for emergent reasoning abilities of LLMs, ReAct became particularly popular by claiming to tease out LLM reasoning abilities with special prompting involving “interleaving reasoning trace with action execution". In this paper, we critically examine the claims of ReAct style prompting for planning and sequential decision-making problems. By introducing systematic variations to the input prompt, we perform a sensitivity analysis along the original claims of ReAct. Our experiments in AlfWorld and WebShop, domains that were used in the original ReAct work, show that the performance is minimally influenced by the interleaved reasoning trace or by the content of these generated reasoning traces. Instead, the performance of LLMs is primarily driven by the unreasonably high degree of similarity between input example tasks and queries, with shockingly little ability to generalize. In addition to raising questions on claims about reasoning abilities, this lack of generalization also implicitly forces the prompt designer to provide instance-specific examples, significantly increasing the cognitive burden on the human. Our empirical results show that the perceived reasoning abilities of LLMs stem from the exemplar-query similarity and approximate retrieval rather than any inherent reasoning abilities, thereby leading to severe lack of generalization beyond the few-shot examples given in the prompts. Our code and prompt settings can be found here on GitHub.