Bufan Gao

Excited to present at #EMNLP2025

Really appreciate @elisakreiss.bsky.social’s kind guidance and encouragement throughout this work 🙏

11.09.2025 16:01 — 👍 2    🔁 0    💬 0    📌 0

Measuring Bias or Measuring the Task: Understanding the Brittle Nature of LLM Gender Biases As LLMs are increasingly applied in socially impactful settings, concerns about gender bias have prompted growing efforts both to measure and mitigate such bias. These efforts often rely on evaluation...

👉 Our results highlight the brittleness of current bias evaluations: small prompt changes can reverse conclusions.

📄 Paper: arxiv.org/abs/2509.04373
💻 Code: github.com/jouisseuse/B...

11.09.2025 16:01 — 👍 2    🔁 0    💬 1    📌 0

Pronoun-Specific Shift Probabilities across models. Bars show the mean shift in token probability for each pronoun (he, she, they) across prompt conditions and attributes. Prompts with instructions and gender references increase preference for “they” and decrease preference for “he,” while “she” varies between models.

When prompts contain cues typical of gender bias evaluation setups, models shift pronoun use: fewer “he,” more “they.”

This suggests that LLM benchmark behavior may generalize less and less to non-benchmark settings, raising new concerns about ecological validity.

11.09.2025 16:01 — 👍 2    🔁 0    💬 1    📌 0

Violin plots of Probability of Pronoun Shift. Models show significant sensitivity to prompt changes: when prompts highlight gender evaluation, pronoun use shifts, with decreased “he” and increased “they” use.

🚨 New #EMNLP2025 paper!

Do LLMs exhibit distinct behavior when the prompt looks similar to common evaluation prompts? 👀

We show that prompts that signal bias evaluation can flip the measured bias. See below ⬇️

11.09.2025 16:01 — 👍 3    🔁 1    💬 1    📌 2