Large language models show human-like grasp of rules, study finds

A new study from arXiv suggests that large language models may possess genuine conceptual understanding of how to apply rules in decision-making, matching human behavior in ways that span different contexts and training periods.

The experimental approach#

Researchers borrowed techniques from cognitive psychology to examine whether LLMs grasp the underlying concept of rules, rather than simply pattern-matching. They designed a novel procedure to calibrate the diversity of responses generated by AI models against responses from human participants. This methodological bridge allowed direct comparison of rule-application strategies between the two groups.

What the experiments revealed#

The first study tested whether LLMs replicated human patterns when given scenarios, regardless of whether those scenarios existed before or after the models' training cut-off dates. The findings showed consistent alignment: all models examined replicated human decision patterns across both sets of scenarios.

An unexpected result emerged when researchers discovered that humans themselves responded differently to scenarios created before versus after the training period. Striking to the researchers, the LLMs replicated this human variation as well—suggesting more than mechanical rule application.

The second experiment explored a nuanced aspect of human reasoning: when people face time pressure, they tend to rely more heavily on the literal text of a rule than on its underlying purpose or intent. The results here diverged by model. Gemini Pro and Claude 3 responded in human-like ways to prompts about both forced delays and time pressure. In contrast, GPT-4o and Llama 3.2 90b did not exhibit the same pattern.

Implications#

The researchers argue that their findings point to conceptual mastery of rules within LLMs—a capability with consequences for how these systems might be applied in legal decision-making and raises philosophical questions about the nature of understanding itself.

Read the full paper on arXiv.