Using Artificial Populations to Study Psychological Phenomena in Neural Models

Jesse Roberts, Kyle Moore, Drew Wilenzick, Douglas H. FisherAAAI 2024 · DOI

Abstract

As Large Language Models (LLMs) are increasingly evaluated for human-like cognitive traits, researchers face a significant methodological challenge: the "sample size of one" problem. This paper introduces PopulationLM, an open-source framework that utilizes Stratified Monte Carlo (MC) dropout to generate "artificial populations" from single pre-trained models. By treating these variations as individual subjects in a psychological study, we demonstrate that single-model evaluations often overestimate the presence of human-like behaviors such as typicality effects and structural priming. Our findings emphasize the need for statistical rigor and population-based analysis in the study of AI behavior.

Introduction

In recent years, the field of AI research has increasingly adopted methodologies from human psychology to test LLMs for "Theory of Mind," logical reasoning, and linguistic biases. However, most of these studies draw conclusions from a single trained model, which may exhibit behaviors that are fragile artifacts of specific weights or training flukes. In human psychology, a conclusion about the species would never be drawn from a single individual; similarly, "LLM psychology" requires a way to study populations to find robust, statistically significant effects.

This paper addresses this lack of replicability by highlighting how small experimental tweaks can cause supposedly "human-like" behaviors to disappear. We propose that studying populations of models is essential for distinguishing core model understanding from statistical noise.

Methodology

The authors propose PopulationLM, a method to efficiently generate a population from a single pre-trained model using Stratified Monte Carlo (MC) Dropout. While dropout is typically disabled during inference, PopulationLM keeps it active, applying different random masks to create 50 distinct "individuals" from the same base model. This approach approximates a Bayesian representation of the model's uncertainty.

To ensure statistical rigor, the framework employs the Kolmogorov-Smirnov (KS) test to determine if a population's behavior is statistically distinct from the base model. Standard psychological metrics, including Pearson and Spearman correlations and Wilcoxon signed-rank tests, are used to measure effect sizes and significance across the artificial population.

Experimental Setup

The study validates the population-based approach using several established benchmarks:

  • Typicality Data: Human-ranked typicality for category items (e.g., "robin" vs. "penguin") from Rosch (1975).
  • Structural Priming Data: Sentence pairs from Sinclair et al. (2022) used to test syntactic mimicry.
  • Training Frequencies: BERT family training data frequencies to control for frequency as a confounding variable.
  • Models Tested: A diverse set of architectures including BERT (Base/Large), DistilBERT, GPT, GPT-2 (Base/Medium), and RoBERTa (Base/Large).

Results and Discussion

The Over-Optimism of Single Models

A primary finding of this research is that evaluating a single "base" model usually results in an overestimation of cognitive behaviors. As shown in Figure 1, the population-wide average typically exhibits a smaller, more realistic effect size compared to the single model baseline, which often appears as an outlier.

Probability vs Typicality Rank for single models vs populations.

Typicality and Training Frequency

While LLMs do exhibit typicality effects—showing higher certainty for typical category members—this behavior is strongly correlated with training frequency. In well-understood categories, typicality explains approximately 20% ($r^2 \approx 0.20$) of the variance in probability outputs. However, if a category is poorly represented in the training data, the typicality effect largely vanishes, suggesting it is a learned statistical frequency rather than an inherent cognitive structure.

Divergent Uncertainty Patterns: MLMs vs. CLMs

The study revealed distinct uncertainty patterns between Masked Language Models (MLMs) and Causal Language Models (CLMs). While MLMs like BERT become more certain with typical items, CLMs like GPT-2 can show higher uncertainty for typical pairings. This phenomenon is quantified in Figure 4, which shows the Spearman correlation of model uncertainty with typicality across different architectures.

Spearman Correlation of Model Normalized Uncertainty with Typicality.

The "Priming" Illusion

Previous studies suggested LLMs show "structural priming" (mimicking sentence structures). However, PopulationLM revealed that for the GPT-2 family, 96% of the change in probability attributed to structural priming was actually a general preference for any prompt. When controlled for this general preference, evidence for true structural priming became marginal.

Conclusion

PopulationLM provides a framework for making AI behavioral studies as rigorous as human psychology. While MC dropout is a mathematical proxy for true model diversity and is currently limited to open-source models, it offers a vital tool for identifying robust behaviors. Future work will apply this method to other phenomena, such as "basic level effects," and further investigate the unique uncertainty patterns of causal models. By moving from case studies to population studies, we can better understand the true cognitive capabilities of neural models.

Summary updated

BibTeX
@inproceedings{roberts2024using,
  author = {Jesse Roberts and Kyle Moore and Drew Wilenzick and Douglas H. Fisher},
  title = {{Using Artificial Populations to Study Psychological Phenomena in Neural Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i17.29856},
}
Figure