Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning
Zach Studdiford, Gary Lupyan
Why It Matters
What makes this one worth your time
Understanding the similarities in reasoning errors between humans and LLMs can inform the development of more robust AI systems and improve our understanding of human cognition.
The paper suggests that both human and LLM reasoning may rely on pattern-matching rather than abstract world models.
Summary
The paper evaluates human participants and 25 large language models (LLMs) on common-sense reasoning tasks, finding similar error patterns between humans and LLMs. It identifies attention heads in LLMs that perform pattern-matching, suggesting that both human and LLM reasoning may rely more on pattern-matching than abstract world models.
Key contributions
- Comparison of reasoning errors between humans and LLMs.
- Identification of attention heads in LLMs that perform pattern-matching.
Notable insights
- Attention heads in LLMs can predict reasoning errors caused by irrelevant prompt details.
- Human reasoning errors may be more aligned with pattern-matching than previously thought.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2606.13607v4 Announce Type: replace Abstract: When large language models (LLMs) fail to generalize or make content-sensitive errors in reasoning, it is often taken as evidence that LLMs are not truly reasoning, but rather performing a kind of pattern matching. The implication is that human behavior does not exhibit the same types of failures because human reasoning relies on principled and content-invariant world models. We test this assumption by first evaluating humans and LLMs on their ability to engage in common-sense reasoning about a variety of everyday situations. Our results reveal convergent patterns of reasoning across 46 LLMs and two cohorts of human participants. We then ask whether this behavioral convergence is due to LLMs having acquired content-invariant world models or a set of pattern-matching heuristics by characterizing the roles of content-invariant and content-sensitive model neurons in producing human-like responses. We find that while LLMs encode both content-invariant and content-sensitive representations, it is content-sensitive mechanisms which are causally responsible for aligning models with humans. Taken together, our results suggest that everyday causal reasoning in people and LLMs makes heavy use of pattern-matching.