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

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

Problem Statement
The prevailing view in cognitive science and AI safety posits that human reasoning relies on abstract, structured "world models" that preserve causal structure, compositionality, and role-filler independence. In contrast, Large Language Models (LLMs) are often criticized for failing to generalize or making haphazard errors, leading to the conclusion that they merely perform shallow "pattern matching" based on statistical associations rather than true reasoning. However, human reasoning is also known to be highly content-sensitive and prone to errors that violate strict logical compositionality (e.g., the Wason selection task, or the "798 is less even" phenomenon). This paper investigates whether the reasoning failures observed in LLMs are unique to their architecture or if they reflect a shared cognitive mechanism with humans. Specifically, the authors ask: Is everyday causal reasoning in both humans and LLMs better explained by abstract world models or by a form of context-dependent pattern matching?

Methodology
The study employs a multi-pronged approach combining behavioral experiments with humans and LLMs, followed by mechanistic interpretability analysis.

1. Behavioral Evaluation:

   • Participants: 142 human participants and 25 LLMs (including open-source models like Gemma and Llama, and proprietary models like GPT-4.1 and DeepSeek-R1).
   • Task: A modular causal reasoning task involving 11 categories of everyday scenarios (e.g., geocentric-near, egocentric-distant, state-follows, action-nonfollows).
   • Design: Prompts were presented in two framings: target (inferring an outcome given a cause) and context (inferring a cause given an outcome). Crucially, the study included minimal-difference pairs where only ostensibly irrelevant content words changed (e.g., "The soup is on the table" vs. "The stew is on the table") while the ground-truth causal outcome remained identical.
   • Metrics: Human accuracy was measured by binary choice. LLM accuracy was measured via logit differences between correct and incorrect tokens (for open weights) or binary correctness (for API models).

2. Mechanistic Interpretability:

   • Model: Focus was placed on gemma-2-27b-it due to its strong alignment with human behavioral patterns.
   • Causal Identification: The authors performed ablation experiments to identify the top-$k$ attention heads most causally responsible for model outputs (i.e., heads whose removal caused the largest change in logit differences).
   • Activation Patching: To determine what information these heads attend to, the authors patched activations from "critical" prompts (where content changes the ground-truth outcome) and "non-critical" prompts (where content changes but the outcome remains the same) into the base prompt.
   • Prediction: The entropy of these causally critical attention heads was used as a feature to predict human accuracy on held-out prompts and scenarios.

Key Results

1. Shared Failure Profiles: Both humans and LLMs exhibited "graded accuracy," where performance was highly sensitive to small, surface-level changes in prompt content.

   • Humans and LLMs were more accurate on geocentric relations involving distant references (e.g., cities) than nearby objects (e.g., paintings), and more accurate on egocentric relations for nearby objects than distant ones.
   • Minor alterations to prompts (e.g., changing "soup" to "stew" in a scenario where the stove is turned on) induced similar "fail states" in both groups, even when the logical outcome was unchanged.
   • LLMs with higher overall accuracy (e.g., DeepSeek-R1) were not necessarily more human-like; in some cases, smaller models (like Gemma-3-27b-it) aligned better with human error patterns.

2. Mechanism of LLM Responses:

   • Content Sensitivity: Activation patching revealed that the top-$k$ causally important attention heads were significantly more sensitive to non-critical content substitutions (e.g., changing "painting" to "statue") than to critical structural changes (e.g., changing "North" to "South"). This supports the hypothesis that these heads implement a form of pattern matching based on surface content rather than abstract causal structure.
   • Predicting Human Errors: The entropy of these specific attention heads successfully predicted human accuracy. A linear model using these head activations explained significant variance in human responses ($R^2 \approx 0.65$ for held-out prompts), even after controlling for category, prompt format, and n-gram frequency.
   • Minimal-Difference Scenarios: In a follow-up evaluation with 96 prompts differing only by a single content word, the activations of these pattern-matching heads predicted human accuracy with high fidelity, explaining up to 76% of the variance.

Key Contributions

• Empirical Alignment: The paper provides robust evidence that human and LLM reasoning failures are not distinct phenomena but share a common profile of content sensitivity and brittleness.
• Mechanistic Explanation: It moves beyond behavioral correlation to identify the specific internal circuit (a subset of attention heads) in LLMs that drives these behaviors.
• Reframing Reasoning: The authors demonstrate that the "pattern matching" mechanism, often dismissed as a failure of LLMs, is actually the same mechanism that allows these models to predict human errors. This suggests that human reasoning itself may rely heavily on pattern matching rather than purely abstract world models.

Significance and Claims
The paper argues that the distinction between "reasoning" (abstract world models) and "pattern matching" (shallow associations) may be a false dichotomy for everyday causal reasoning. The authors claim that:

1. Human Reasoning is Context-Dependent: Human reasoning is not fully compositional or abstract; it is deeply influenced by the specific content and context of a problem. The "fail states" seen in LLMs are not evidence of a lack of intelligence but rather a reflection of how human cognition actually operates.
2. LLMs as a Tool for Cognitive Science: Because LLMs allow for causal manipulation of internal mechanisms (unlike humans), they serve as a unique tool for reverse-engineering human cognition. The fact that LLM attention heads can predict human errors suggests that the "messiness" of human reasoning is a feature, not a bug, likely evolved to handle the open-ended, context-rich nature of real-world problems.
3. Pattern Matching as a Valid Mechanism: The study posits that pattern matching is not merely a fallback or a shallow heuristic but a fundamental way neural networks (both biological and artificial) solve problems. The ability to track "ostensibly irrelevant" details may be adaptive in naturalistic settings where the specific entities involved (e.g., a specific type of food or a specific city) carry implicit contextual information that rigid abstract models might discard.

The authors conclude that reasoning about everyday causal scenarios is, for both people and LLMs, a form of pattern matching that is highly sensitive to the full content of a prompt, rather than a process driven solely by abstract, structure-preserving world models.