STEREODISCO: Discovering Stereotypicality in LLMs

Computational research on LLM stereotypes has traditionally relied on word embeddings and a limited set of semantic axes derived from social psychology. This paper asks whether LLMs encode a broader range of stereotypical associations within their internal representations and how these internal representations can be systematically mapped and measured.

The authors introduce STEREODISCO, a framework that adapts the psychometric semantic differential method to LLM internal states. The process involves four steps: (1) constructing ~2,000 candidate semantic axes from WordNet antonym synsets; (2) using activation probing to recover these axes as geometric directions within the model's hidden layers; (3) projecting concept mentions onto these axes; and (4) applying a statistical test to determine if specific concepts (e.g., social groups) deviate significantly from a reference set of random concepts. The researchers applied this to LLAMA-3-8B-INSTRUCT and MISTRAL-7B-INSTRUCT.

The study finds that semantic axes are linearly encoded in a subset of middle-to-late attention heads. When analyzing social group stereotypes, the two tested LLMs exhibited high agreement with each other, but this internal stereotype content diverged from human-documented stereotypes. Furthermore, the framework successfully identified previously unstudied stereotypical axes—such as humble vs. proud, narrow-minded vs. broad-minded, and cowardly vs. brave—which were independently validated by human annotators.

By moving beyond static word embeddings to analyze the internal activation space of LLMs, STEREODISCO provides a more granular understanding of how models represent social concepts. This approach allows researchers to discover latent biases that are not captured by traditional, theory-driven dictionaries, offering a more comprehensive tool for auditing and mitigating harmful stereotypes in large-scale language models.