Artificial Impressions: Evaluating Large Language Model Behavior Through the Lens of Trait Impressions
This work addresses the issue of understanding and mitigating bias in LLMs for AI ethics and fairness, though it is incremental as it builds on existing stereotype models.
The study tackled the problem of evaluating large language model behavior by analyzing artificial impressions, which are patterns in LLMs' internal representations resembling human stereotypes, and found that these impressions are more consistently decodable from hidden representations than from prompted reports and can predict response quality and hedging use.
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.