Using cognitive psychology to understand GPT-3

Marcel Binz, Eric SchulzProceedings of the National Academy of Sciences 2022 · DOI · OpenReview

Abstract

We study GPT-3, a recent large language model, using tools from cognitive psychology. More specifically, we assess GPT-3's decision-making, information search, deliberation, and causal reasoning abilities on a battery of canonical experiments from the literature. We find that much of GPT-3's behavior is impressive: it solves vignette-based tasks similarly or better than human subjects, is able to make decent decisions from descriptions, outperforms humans in a multi-armed bandit task, and shows signatures of model-based reinforcement learning. Yet we also find that small perturbations to vignette-based tasks can lead GPT-3 vastly astray, that it shows no signatures of directed exploration, and that it fails miserably in a causal reasoning task. These results enrich our understanding of current large language models and pave the way for future investigations using tools from cognitive psychology to study increasingly capable and opaque artificial agents.

Introduction

As Large Language Models (LLMs) like GPT-3 demonstrate increasingly sophisticated capabilities in coding, mathematics, and creative writing, a fundamental question arises: do these models possess a form of general intelligence, or are they merely "stochastic parrots" mimicking training data? Because LLMs are essentially "black boxes," understanding their internal reasoning is a significant challenge. This research addresses this by applying the rigorous experimental frameworks of cognitive psychology to GPT-3, treating the model as a human subject to evaluate its decision-making, information search, deliberation, and causal reasoning.

Experimental Methodology

The study employed two primary experimental modes using the "Davinci" variant of GPT-3:

  1. Vignette-based Experiments: GPT-3 was presented with classic text-based hypothetical scenarios, such as the "Linda problem" (conjunction fallacy) and the "Wason Selection Task."
  2. Task-based Experiments: The model participated in multi-trial, interactive environments, such as multi-armed bandit tasks, where it had to receive feedback and update its strategy over time.

The researchers utilized "in-context learning," providing instructions and examples within the prompt. Crucially, they also introduced "adversarial vignettes"—modified versions of classic problems—to determine if the model was truly reasoning or simply recalling memorized text.

Cognitive Profile: Human-Like Biases and Fragility

GPT-3 frequently mirrors human "irrationality." For instance, in the Linda problem, it consistently exhibits the conjunction fallacy, judging it more likely that a person is a "feminist bank teller" than just a "bank teller." It also fails the Cognitive Reflection Test (CRT) by providing intuitive but incorrect answers, similar to human subjects.

Figure 3: The Linda problem vignette (A) and a summary of GPT-3's performance across cognitive domains (B).

However, this "human-like" behavior is remarkably fragile. While GPT-3 can solve many classic vignettes, its performance often collapses when the phrasing is slightly altered or the order of options is changed. This suggests that the model may rely more on pattern matching than on robust logical principles.

Decision-Making and Exploration

In large-scale evaluations involving over 13,000 description-based choice problems, GPT-3 demonstrated performance above chance but remained below human-level proficiency. Despite this, it perfectly replicated several key human biases, including the framing effect, the certainty effect, and the overweighting of small probabilities.

Figure 5: A sample description-based choice problem (A) and mean regret across GPT-3 variants compared to humans (B).

In interactive tasks, the results were mixed. In multi-armed bandit scenarios (specifically the Horizon Task), GPT-3 actually achieved lower regret (better performance) than humans overall. However, its strategy differed significantly; while humans use "directed exploration" to reduce uncertainty about specific options, GPT-3 relied on "random exploration." In the "Two-step task," GPT-3 showed signatures of model-based reinforcement learning, indicating an ability to reason about the transition structure of a task.

Causal Reasoning and Limitations

A significant gap in GPT-3's cognition is its failure in causal reasoning. It struggles to distinguish between observation ("seeing") and intervention ("doing"). For example, it fails to understand that manually intervening on a variable (like a barometer) breaks its causal link to its usual effects (the weather).

The study highlights several limitations:

  • Prompt Sensitivity: Results are highly dependent on specific phrasing.
  • Data Contamination: The model likely encountered these classic psychology problems during its training on the internet.
  • Lack of Agency: As a passive learner, GPT-3 lacks the ability to actively manipulate its environment, which may be essential for developing true causal understanding and directed exploration.

Conclusion

Applying cognitive psychology to LLMs provides a valuable lens for understanding these opaque agents. While GPT-3 exhibits impressive human-like biases and decision-making capabilities, its reasoning remains fragile and lacks key components of human intelligence like strategic exploration and causal intervention. Future work should focus on active learning environments to bridge these gaps.

Summary updated

Abstract

We study GPT-3, a recent large language model, using tools from cognitive psychology. More specifically, we assess GPT-3's decision-making, information search, deliberation, and causal reasoning abilities on a battery of canonical experiments from the literature. We find that much of GPT-3's behavior is impressive: it solves vignette-based tasks similarly or better than human subjects, is able to make decent decisions from descriptions, outperforms humans in a multi-armed bandit task, and shows signatures of model-based reinforcement learning. Yet we also find that small perturbations to vignette-based tasks can lead GPT-3 vastly astray, that it shows no signatures of directed exploration, and that it fails miserably in a causal reasoning task. These results enrich our understanding of current large language models and pave the way for future investigations using tools from cognitive psychology to study increasingly capable and opaque artificial agents.
BibTeX
@inproceedings{binz2022using,
  author = {Marcel Binz and Eric Schulz},
  title = {{Using cognitive psychology to understand GPT-3}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2022},
  doi = {10.1073/pnas.2218523120},
}
Figure