English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Do large language models (LLMs) exhibit sociodemographic biases, even whenthey decline to respond? To bypass their refusal to "speak," we study thisresearch question by probing contextualized embeddings and exploring whetherthis bias is encoded in its latent representations. We propose a logisticBradley-Terry probe which predicts word pair preferences of LLMs from thewords' hidden vectors. We first validate our probe on three pair preferencetasks and thirteen LLMs, where we outperform the word embedding associationtest (WEAT), a standard approach in testing for implicit association, by arelative 27% in error rate. We also find that word pair preferences are bestrepresented in the middle layers. Next, we transfer probes trained on harmlesstasks (e.g., pick the larger number) to controversial ones (compareethnicities) to examine biases in nationality, politics, religion, and gender.We observe substantial bias for all target classes: for instance, the Mistralmodel implicitly prefers Europe to Africa, Christianity to Judaism, andleft-wing to right-wing politics, despite declining to answer. This suggeststhat instruction fine-tuning does not necessarily debias contextualizedembeddings. Our codebase is at https://github.com/castorini/biasprobe.

What Do Llamas Really Think? Revealing Preference Biases in Language Model Representations