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

**Authors:** Nicholas Deas, Kathleen McKeown  
**arXiv:** [2510.08915](https://arxiv.org/abs/2510.08915)  
**Date:** 2025-10-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.08915v1/

## Key points

- Authors Nicholas Deas and Kathleen McKeown introduce the concept of artificial impressions within Large Language Models (LLMs)
- Artificial impressions are patterns in LLMs' internal representations of prompts that mirror human impressions and stereotypes based on language
- Linear probes are used to predict these impressions according to the Stereotype Content Model (SCM)
- The study explores the connection between artificial impressions and downstream model behavior
- Inconsistency in LLMs' reporting of impressions when prompted, contrasting with reliable linear decodability from hidden representations
- Artificial impressions derived from prompts can forecast quality and utilization of hedging in model responses
- Analysis of how specific content, stylistic elements, and dialectal characteristics in prompts impact LLM impressions

## Summary

In their paper titled "Artificial Impressions: Evaluating Large Language Model Behavior Through the Lens of Trait Impressions," authors Nicholas Deas and Kathleen McKeown introduce and explore the concept of artificial impressions within Large Language Models (LLMs). These artificial impressions are patterns in LLMs' internal representations of prompts that mirror human impressions and stereotypes based on language. The authors employ linear probes on generated prompts to predict these impressions according to the two-dimensional Stereotype Content Model (SCM). Through their research, they delve into the connection between these impressions and downstream model behavior, as well as investigate prompt features that may influence such impressions. The study reveals intriguing findings, including the inconsistency in LLMs' reporting of impressions when prompted, contrasting with the more reliable linear decodability of impressions from their hidden representations. Moreover, the authors demonstrate that artificial impressions derived from prompts can forecast the quality and utilization of hedging in model responses. Additionally, they analyze how specific content, stylistic elements, and dialectal characteristics in prompts impact LLM impressions. This comprehensive exploration sheds light on the intricate interplay between language models, human perceptions, stereotypes, and linguistic nuances. By delving into artificial impressions within LLMs, Deas and McKeown provide valuable insights into understanding and evaluating large language model behavior through a unique lens of trait impressions.