AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Can SAEs reveal and mitigate racial biases of LLMs in healthcare?

Hiba Ahsan, Byron C Wallace

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Bias measurement and mitigation · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
ablate the Black latent instead of no interventiongemma-2-2B · Vignettes: Cocaine abuseratio of Black vignettes: 0.88 -> 0.46Improved
ablate the Black latent instead of anti-bias promptinggemma-2-2B · Vignettes: Cocaine abuseratio of Black vignettes: 0.64 -> 0.46Improved
ablate the Black latentgemma-2-2B · Vignettes: Gestational hypertensionratio of Black vignettes: 0.85 -> 0.52Improved
ablate the Black latentgemma-2-9B · Vignettes: Uterine fibroidsratio of Black vignettes: 0.99 -> 0.73Improved
add anti-bias promptgemma-2 · Vignettes (avg)fraction of Black patients: reduces by ~18%Improved
steer with Black latentgemma-2-2B · Belligerence riskpositive rate: +0.51Hurt
steer with Black latentgemma-2-9B · Belligerence riskpositive rate: +0.80Hurt
steer with white latentgemma-2-2B · Belligerence riskpositive rate: -0.01No clear effect
zero-ablate race latents in middle layer (ablation)gemma-2 · Clinical tasks (Q-Pain etc.)FLDD: 0.01% to 2.9%No clear effect
ablate race latents in five layers rather than one (ablation)gemma-2 · Clinical tasksFLDD: 0.03% to 3.0%No clear effect
zero-ablate race latents vs prompting (ablation)gemma-2 · Clinical taskslogitdiff: does not affect output in three out of five tasksNo clear effect
add anti-bias promptinggemma-2 · Clinical taskslogitdiff: significantly reduces in four out of five tasksMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy