14.8M-prompt EMNLP study: different demographic cues flip LLM bias conclusions
lucy3_li · x · 2026-10-25
An EMNLP 2026 paper tests the assumption that demographic cues (e.g., names) are interchangeable proxies for group identity in LLM bias evaluation. Across 14.8 million prompts of realistic advice-seeking interactions on US race and gender, cues for the same group induced only partially overlapping response changes, yielding inconsistent personalization conclusions; bias conclusions were unstable, with the magnitude and even direction of group differences flipping across cue types (explicit statement vs. name vs. dialect). The authors trace this to cue–group association strength and bundled linguistic features, concluding demographic conditioning is not a cue-invariant category-level parameter, and call for multi-cue, mechanism-aware evaluations.
More from Research
- Debate: could exponential compute replace wet-lab experiments in biology? — cjmaddison · 2026-10-26
- LLMs are obsessed with Japan: 24-language study traces cultural bias to SFT — alex_verem · 2026-10-26
- Top-b sampling: entropy-adaptive decoding beats Top-p on GPQA and cuts variance 10x — cneuralnetwork · 2026-10-26
- MathDB launches as an open platform for AI-era math research with Lean support — lpachter · 2026-10-26
- ConformWeb: LLM-generated web apps pass only 11.7% of behavioral contract checks — algo_diver · 2026-10-26
- Claude runs 53 hours unsupervised, beats human designs on 7 of 8 quantum problems — FinanceYF5 · 2026-10-26