Artificial Impressions: Evaluating Large Language Model Behavior Through the Lens of Trait Impressions

TL;DR

The paper investigates artificial impressions in large language models through the Stereotype Content Model's warmth and competence dimensions. It demonstrates that LLM impressions can be reliably recovered from hidden-state representations using linear probes, even when prompts themselves fail to yield consistent self-reports. The study shows that these impression signals predict downstream behaviors such as response quality and hedging, and that prompt content, style, and dialect (notably AAL vs WME) modulate encoded impressions. This work highlights how prompt features shape model biases and behavior, raising ethical considerations about personalization and potential harms, while offering a concrete probing framework for understanding impression formation in LLMs.

Abstract

We introduce and study artificial impressions--patterns in LLMs' internal representations of prompts that resemble human impressions and stereotypes based on language. We fit linear probes on generated prompts to predict impressions according to the two-dimensional Stereotype Content Model (SCM). Using these probes, we study the relationship between impressions and downstream model behavior as well as prompt features that may inform such impressions. We find that LLMs inconsistently report impressions when prompted, but also that impressions are more consistently linearly decodable from their hidden representations. Additionally, we show that artificial impressions of prompts are predictive of the quality and use of hedging in model responses. We also investigate how particular content, stylistic, and dialectal features in prompts impact LLM impressions.

Figures (13)

• Figure 1: Overview of our approach. Because LLMs inconsistently report impressions of users, we fit probes to extract LLM artificial impressions of prompt authors according to the Stereotype Content Model.
• Figure 2: 1st and 3rd-Person setting prompts for evaluating LLM-reported impressions.
• Figure 3: Overview of our approach to generating ground truth data for training and evaluating impression probes.
• Figure 4: Prompt for generating synthetic prompts conditioned on an impression specification. The Impression Specification is composed of a single or pair of traits.
• Figure 5: F1 scores (y-axis) of trained impression probes against the input model layer (x-axis) for each LLM and impression dimension. Colors represent varying percentages of data used for training. Shaded regions reflect 95% confidence intervals across 5 folds. Maximum F1 score achieved for each variant is starred and dashed lines represent scores for BOW-classifier baselines.