Using Artificial Populations to Study Psychological Phenomena in Neural Models

TL;DR

PopulationLM introduces a population-based framework for studying cognitive phenomena in neural language processing using stratified Monte Carlo dropout to form artificial populations. The approach enables uncertainty-aware, replication-friendly analyses and is demonstrated on typicality effects and structural priming, revealing robust typicality across models but weak structural priming, while single-model analyses tend to overstate cognitive behaviors. Across model families, typicality correlates with category representation in training data and population uncertainty, whereas structural priming shows limited and model-specific evidence. The work highlights the value of population-level methods for robust psycholinguistic inference and points to future research directions in prompt pattern effects and additional cognitive phenomena.

Abstract

The recent proliferation of research into transformer based natural language processing has led to a number of studies which attempt to detect the presence of human-like cognitive behavior in the models. We contend that, as is true of human psychology, the investigation of cognitive behavior in language models must be conducted in an appropriate population of an appropriate size for the results to be meaningful. We leverage work in uncertainty estimation in a novel approach to efficiently construct experimental populations. The resultant tool, PopulationLM, has been made open source. We provide theoretical grounding in the uncertainty estimation literature and motivation from current cognitive work regarding language models. We discuss the methodological lessons from other scientific communities and attempt to demonstrate their application to two artificial population studies. Through population based experimentation we find that language models exhibit behavior consistent with typicality effects among categories highly represented in training. However, we find that language models don't tend to exhibit structural priming effects. Generally, our results show that single models tend to over estimate the presence of cognitive behaviors in neural models.

Figures (4)

• Figure 1: Single model regression vs population of models probability-typicality regression for RoBERTa Large. Rank is inversely related to typicality. 95% confidence intervals shown for both with very narrow bounds on the population.
• Figure 2: For each model the, colored bars show within category Pearson correlation (p$<$0.03). For each model the total Pearson correlation (p$<$0.03) is shown as the gray background bar. The total Pearson correlation (p$<$0.01) for well understood categories (categories with an average item frequency $>$ 60000 in training data for Bert) is shown as the light blue bar. In well understood categories, typicality of the item may explain up to $r^2 \approx$ 20% of the category probability volatility.
• Figure 3: Within category and total Spearman correlation (p$<$0.08) is shown. The total Spearman correlation (p$<$0.01) for categories with an average item frequency $>$ 60000 in training data for Bert is also shown.
• Figure 4: Emergence of within category behavior consistent with typicality in BERT is strongly predicted by within category item frequency in training data.