Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning

Problem
This preprint addresses the gap in understanding the reasoning capabilities of large language models (LLMs) compared to human reasoning. It challenges the prevailing notion that LLMs fail due to a lack of principled reasoning, positing instead that both humans and LLMs exhibit similar reasoning errors, indicating a shared reliance on pattern matching mechanisms. The authors aim to provide empirical evidence for this hypothesis by evaluating both human participants and LLMs in common-sense reasoning tasks.

Method
The study evaluates 25 LLMs alongside human participants on their ability to perform common-sense reasoning across various everyday scenarios. The authors analyze the attention heads of the LLMs to identify which components are responsible for generating responses. They employ a systematic approach to categorize reasoning errors, linking them to specific attention heads that exhibit pattern-matching behavior. The methodology includes a comparative analysis of error patterns in both humans and LLMs, focusing on how irrelevant prompt details can lead to similar reasoning failures.

Results
The findings reveal that both humans and LLMs exhibit comparable patterns of reasoning errors, suggesting a shared mechanism of pattern matching. The authors report that certain attention heads in LLMs can predict reasoning errors in humans, particularly when prompted with seemingly irrelevant details. While specific numerical results and effect sizes are not disclosed in the abstract, the implication is that the degree of similarity in error patterns is statistically significant, warranting further investigation into the underlying mechanisms.

Limitations
The authors acknowledge that their study is limited by the scope of the common-sense reasoning tasks selected and the specific LLMs evaluated. They do not address potential variations in reasoning across different cultural or contextual backgrounds in human participants, which could influence the generalizability of their findings. Additionally, the reliance on attention heads as a sole explanatory mechanism may overlook other factors contributing to reasoning errors in both humans and LLMs.

Why it matters
This research has significant implications for the understanding of reasoning in artificial intelligence and cognitive science. By demonstrating that both humans and LLMs may rely on similar pattern-matching strategies, it challenges the notion of LLMs as fundamentally different from human reasoning processes. This insight could inform the development of more robust AI systems that better mimic human reasoning or enhance interpretability in LLMs. The findings suggest a need for further exploration into the cognitive mechanisms underlying reasoning, as published in arXiv.