🤖 AI

Evidence of conceptual mastery in the application of rules by Large Language Models

This paper employs psychological methods and comparative experiments to demonstrate that Large Language Models exhibit conceptual mastery in applying rules, as evidenced by their ability to replicate diverse human decision-making patterns and context-dependent responses to time pressure.

Original authors: José Luiz Nunes, Guilherme FCF Almeida, Brian Flanagan

Published 2026-08-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: José Luiz Nunes, Guilherme FCF Almeida, Brian Flanagan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

For decades, the question of whether machines can truly understand has been a quiet hum in the background of artificial intelligence. When a computer writes a poem or solves a math problem, it is often unclear if it has grasped the meaning behind the words or if it is simply recalling patterns it has seen before, much like a student who memorizes the answers to a practice test without understanding the subject. This distinction between genuine understanding and rote memorization is the central puzzle for researchers studying large language models, the powerful computer systems that can converse, reason, and create. To find out if these machines possess a real grasp of concepts, scientists look for evidence that they can apply rules flexibly in new situations, rather than just repeating what they have been told. If a machine can navigate a complex rule by understanding its intent, even when the wording changes or the situation is unfamiliar, it suggests a deeper form of competence. This is not just an academic curiosity; as these tools begin to appear in legal settings and other high-stakes environments, knowing whether they are truly reasoning or merely mimicking becomes a matter of practical importance.

A team of researchers set out to test this very question by putting artificial intelligence through a series of rigorous tests designed to separate memorization from true understanding. They focused on how both humans and machines apply rules, specifically looking at situations where the literal text of a rule might conflict with its underlying purpose. Imagine a sign that says "No Dogs Allowed" in a restaurant, posted to keep the floor clean and prevent noise. If a person walks in with a noisy, unruly dog, both the text and the purpose of the rule say they broke the rule. But what if a blind person walks in with a well-trained guide dog? The text says no dogs, but the purpose of the rule is not violated. The researchers wanted to see if the machines could weigh the text against the purpose, just as humans do, and if they could do this even when the scenarios were completely new to them.

In the first phase of their work, the scientists created a set of brand-new scenarios that had never been published on the internet before the machines were trained. This was a crucial step to ensure the models were not simply reciting answers they had memorized from their training data. They asked both human participants and several different large language models to judge whether characters in these stories had broken the rules. The results were striking. The machines did not just guess; their judgments closely tracked those of the humans. When the new stories made the rule's text less important than its purpose, both the humans and the machines shifted their thinking in the same way, placing less weight on the strict wording and more on the reason behind the rule. This suggested that the models were not just recalling old examples but were actually applying a general understanding of how rules work.

To be sure this wasn't a fluke caused by the specific way the questions were asked, the researchers then changed the instructions given to the machines and even reversed the numbers on their answer scales. They asked the models to respond under different system prompts and with different labels for their choices. Despite these changes, the core pattern remained stable. The machines continued to balance the text and the purpose of the rules in a way that mirrored human intuition. This robustness indicated that their understanding was not fragile or dependent on a specific trick of the question, but rather a standing competence that held up even when the surface details of the task shifted.

The researchers then introduced a twist that humans experience but machines do not: time pressure. In human studies, when people are forced to answer quickly, they tend to rely more on the literal text of a rule and less on its purpose. When given more time to think, they are better able to consider the spirit of the rule. The scientists tried to replicate this by telling the machines they had only a few seconds to answer, even though the machines process information at the same speed regardless of the instruction. The results here were mixed and depended on the specific model. Some of the newer, smaller models seemed to mimic the human reaction to time pressure, changing their answers as if they were rushing. However, other models ignored the time constraint entirely, applying the rules consistently whether they were told to hurry or to wait. This divergence was telling; the models that ignored the irrelevant time pressure appeared to be demonstrating a more stable form of competence, one that did not waver based on a cue that had no real effect on their ability to think.

Finally, the team tested whether giving the machines more time to "think" by allowing them to generate longer chains of reasoning would change their answers. For most of the models, increasing the effort they spent on reasoning did not alter their judgments. They applied the rules with the same fluency whether they were asked to think quickly or to deliberate deeply. This mirrors how humans often apply familiar concepts; we do not need to re-derive the logic of a rule every time we encounter it. The findings suggest that for these models, the concept of a rule is already mastered and ready for use, rather than something they have to reconstruct from scratch every time.

The study concludes that while these machines are not perfect copies of human thought, they do show signs of genuine conceptual mastery. They can generalize their understanding to new situations, remain stable when the task changes, and apply rules without needing to overthink every detail. The evidence points away from the idea that they are simply memorizing patterns and toward the possibility that they have developed a flexible, semantic competence. This does not mean they are human, nor does it mean they are infallible, but it does suggest that their ability to reason about rules is more than just a sophisticated form of mimicry. As these tools become more integrated into society, understanding the nature of this competence becomes essential for knowing how to trust and use them.

Technical Summary: Evidence of Concept Mastery in Large Language Models

Problem Statement

The central challenge addressed by this research is distinguishing between conceptual mastery and rote memorization in Large Language Models (LLMs). While LLMs have demonstrated performance matching or exceeding humans in various domains, critics argue this may result from the reproduction of memorized patterns from pre-training data rather than genuine understanding. Previous studies comparing human and LLM judgments on rule violation often utilized stimuli available on the open web prior to model training cut-offs, leaving open the possibility that models simply recalled human responses. Furthermore, LLMs exhibit "diminished diversity of thought" (lower response variance than humans) and sensitivity to incidental task features (e.g., prompt wording, scale anchors), which critics cite as evidence against true conceptual competence. The paper seeks to determine if LLMs possess a generalizable, standing semantic competence in applying rules, specifically in scenarios where a rule's text and purpose diverge.

Methodology

The authors conducted five experiments across 13 different LLMs, utilizing a framework of Experimental Jurisprudence involving rule violation judgments. The methodology focused on four key manipulations to test robustness and generalization:

Novelty of Stimuli (Studies 1A & 1B):
- Study 1A: Compared LLM judgments against newly collected human data ( $n=115$ ) using both published "old" vignettes and newly created vignettes that were not available online until after the models' training cut-offs. This tested if models could generalize to unseen data.
- Study 1B: Replicated the findings with a newer cohort of smaller, more efficient models (Qwen 3, Claude 4.5 Haiku, Gpt-5 Mini, Gemini 3 Flash) while manipulating system prompts (short vs. long instructions) and numerical scale anchors (standard vs. reversed Likert scales) to test robustness against incidental features.
Time Pressure Manipulation (Studies 2A & 2B):
- Leveraged the known human cognitive phenomenon where time pressure shifts judgment from rule purpose to rule text.
- Mechanism: Humans are constrained by cognitive resources; LLMs are not. The study instructed models to simulate "speeded" (4-second) or "delayed" (15-second) conditions. Since the actual token generation time for LLMs is unaffected by these instructions, a model exhibiting human-like shifts in response based solely on these cues would suggest memorization of human patterns rather than conceptual understanding.
- Study 2A: Used the original model cohort (Llama 3.2 90b, Gemini Pro, Claude 3 Opus, GPT-4o).
- Study 2B: Applied the same manipulation to the newer, smaller model cohort with prompt and scale variations.
Reasoning Effort (Study 3):
- Investigated whether increasing "reasoning effort" (a parameter in reasoning models that expands chain-of-thought deliberation) alters rule application.
- Hypothesis: If rule application relies on expanded deliberation, higher effort should change outcomes. If it relies on standing semantic competence (fluency), effort levels should have little effect.
- Models: Tested GPT-5.4 Mini, GPT-5.5, Gemini 3.5 Flash, Claude Sonnet 5, and GPT-OSS 120B.

Calibration: To address the "diminished diversity" critique, the authors calibrated model temperature (and reasoning effort for specific models) per model to match the response variance of the human sample in the "old vignettes" condition.

Key Results

Generalization to Novel Stimuli (Study 1A): LLMs tracked human judgments closely on both old and new vignettes. Crucially, both humans and LLMs showed a reduced reliance on rule text when evaluating the new vignettes compared to the old ones. This unanticipated, parallel shift suggests LLMs are sensitive to subtle, systematic differences in the stimuli that are difficult to articulate, supporting generalization over memorization.
Robustness to Incidental Features (Study 1B): Despite significant interactions with system prompts and scale anchors, the core effects of rule text and rule purpose remained robust and significant across all models and conditions. The correlation between LLM and human judgments remained high ( $r \approx 0.77 - 0.93$ ) regardless of prompt variations.
Time Pressure Sensitivity (Studies 2A & 2B): Results were mixed and model-specific:
- GPT-4o and Llama 3.2 90b largely ignored the time pressure instructions, maintaining consistent reliance on text and purpose regardless of the "speeded" or "delayed" cue. This suggests they are not merely mimicking human time-pressure responses.
- Gemini Pro and Claude 3 (Study 2A) and Claude 4.5 Haiku and Gemini 3 Flash (Study 2B) showed sensitivity to the time pressure manipulation, mimicking human patterns in some cases (e.g., increased textualism under delay). However, they also diverged from humans in other cases (e.g., Claude's response to purpose-only violations).
- Parameter Count: The replication of human-like time-pressure sensitivity was more apparent in models with fewer parameters, suggesting a potential link between model size, memorization, and alignment with human heuristics.
Reasoning Effort (Study 3): Increasing reasoning effort produced no detectable change in rule application for most models. Sensitivity to text and purpose remained stable. A significant "purpositivist" trend (increased focus on purpose) was observed only in higher-effort settings for GPT-OSS and Claude Sonnet 5, but overall, the concept appeared to be applied fluently without expanded deliberation.
Response Variance: Despite temperature calibration, LLMs consistently exhibited lower response variance than humans, particularly for Claude 3 and GPT-4o.

Key Contributions

Evidence of Generalization: Demonstrated that LLMs apply rules to novel, post-training stimuli in ways that parallel human intuition, including subtle shifts in reliance on text vs. purpose, which challenges the "rote memorization" hypothesis.
Differentiation of Competence vs. Alignment: Showed that "conceptual competence" (robust application of rules) does not always equate to "human alignment" (mimicking human heuristics like time-pressure responses). Some models (e.g., GPT-4o) displayed competence by ignoring irrelevant time constraints, while others mimicked human biases.
Fluency of Rule Application: Provided evidence that rule application in LLMs (and humans) relies on standing semantic competence rather than expanded deliberative processing, as increasing reasoning effort did not significantly alter outcomes for most models.
Methodological Rigor: Established a protocol for calibrating LLM temperature to human variance and using novel stimuli to control for pre-training data contamination.

Significance and Claims

The paper concludes that the findings are inconsistent with rote memorization or prompt-specific responding alone. Instead, the results suggest that LLM rule application reflects a generalizable, standing semantic competence that does not typically depend on expanded deliberation.

The authors argue that while LLMs may not perfectly align with human responses in every manipulation (particularly regarding time pressure), their ability to track human intuition in novel cases and remain robust to incidental task features indicates a genuine acquisition of conceptual competence. This has implications for the deployment of LLMs in legal reasoning, suggesting that the "absence of a human touch" may not necessarily degrade the quality of rule violation judgments, as models can apply concepts fluently and robustly.

The paper modestly notes that the "diminished diversity of thought" in LLMs remains a persistent issue resistant to temperature tuning, and that the relationship between model size, memorization, and conceptual mastery warrants further investigation.