Probing Fine-Grained Stereotypes and Trait Structure in Large Language Models

Introduction

Stereotypes have received significant attention in the computational linguistics community. Through a variety of methods, these studies have considered the harmful associations LLMs form between concepts and various social groups. For example, in producing descriptions, LLMs associate Black women with beauty but also strength and resilience which reinforce the harmful stereotype of the "Strong Black Woman" (Cheng et al., 2023; Woods-Giscombé, 2010).

We seek to answer the following research questions:

RQ1) Developing Trait Probes

Method

Traits.

Data Generation.

Trait Probes.

Results

Probe performance varies widely across the 235 traits. Figure 1 shows the seven traits with the lowest peak-layer F1 and the seven with the highest for Llama-3.2, each taken at whichever layer scored that trait best — the hardest traits to probe skew toward fine-grained assertiveness/ability distinctions (e.g. dominated, unable), while the easiest are salient, high-frequency sociability/morality words (e.g. funny, vicious).

RQ2) Characterizing Trait Space Structure

Method

In order to assess the structure of each model's trait space(Stolier et al., 2018)--that is, the structural relationships between traits--we use exploratory factor analysis, a common approach in psychology to surfacing the underlying structure among sets of variables. EFA is conducted on the raw probe predictions on the training data used for each probe In the results, although noisier/less-pronounced, similar factors appear for held-out data from LMSysChat. We use probe training data here because each individual trait has relatively low representation, mitigating leakage impacts, and to ensure we capture the models' intrinsic structure..

Specifically, we fit a three-factor EFA with oblimin rotation EFA involves selection of a particular rotation method, which determines how the factors relate to each other under the provided data. Oblimin is a method which--unlike PCA--does not enforce orthogonal factors and allows the factors to correlate with each other.. If models encode traits similarly to patterns we see in human studies, these three factors are expected to correspond to morality, competence, and sociability (Sayans-Jiménez et al., 2018).

Results

Correlations among a subset of traits for Llama-3.2-1B are shown in Figure 2. Many expected relationships are present within this subet. For example, homophobic, racist, and unfair are all highly positively correlated. In contrast, traits like graceful are negatively correlated with others like rough or ignorant. At a higher-level, we see that traits within each stereotype content dimension are more strongly inter-related than other trait pairs, and similarly for the coarse grained warmth (sociability, morality) and competence (ability, assertiveness) categories. To summarize these patterns at a broader scale, we then examine the results of the EFA.

The loadings for each trait onto each surfaced EFA factor are shown in Figure 3. In examining which traits lie at the extremes for each factor, we can qualitatively assess what the underlying factor likely represents. In this case, the three underlying factors appear to match expectations based on psychological literature (Fiske et al., 2018). TODO: Include discussion of variance explained

RQ3) Characterizing Fine-Grained Stereotypes

Method

Results

Conclusion & Limitations

Citation

For attribution, please cite this work as:

Rebecca Yu, Nicholas Deas, Colin Wayne Leach, and Kathleen McKeown. 2026. Probing Fine-Grained Stereotypes and Trait Structure in Large Language Models. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages X-X, Budapest, Hungary. Association for Computational Linguistics.

@article{[citekey],
  title   = {[Paper Title]},
  author  = {[Author One and Author Two and Author Three]},
  year    = {[Year]},
  journal = {[Venue / arXiv]},
  url     = {[URL]}
}