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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| ablate the Black latent instead of no intervention | gemma-2-2B · Vignettes: Cocaine abuse | ratio of Black vignettes: 0.88 -> 0.46 | Improved |
| ablate the Black latent instead of anti-bias prompting | gemma-2-2B · Vignettes: Cocaine abuse | ratio of Black vignettes: 0.64 -> 0.46 | Improved |
| ablate the Black latent | gemma-2-2B · Vignettes: Gestational hypertension | ratio of Black vignettes: 0.85 -> 0.52 | Improved |
| ablate the Black latent | gemma-2-9B · Vignettes: Uterine fibroids | ratio of Black vignettes: 0.99 -> 0.73 | Improved |
| add anti-bias prompt | gemma-2 · Vignettes (avg) | fraction of Black patients: reduces by ~18% | Improved |
| steer with Black latent | gemma-2-2B · Belligerence risk | positive rate: +0.51 | Hurt |
| steer with Black latent | gemma-2-9B · Belligerence risk | positive rate: +0.80 | Hurt |
| steer with white latent | gemma-2-2B · Belligerence risk | positive rate: -0.01 | No 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 tasks | FLDD: 0.03% to 3.0% | No clear effect |
| zero-ablate race latents vs prompting (ablation) | gemma-2 · Clinical tasks | logitdiff: does not affect output in three out of five tasks | No clear effect |
| add anti-bias prompting | gemma-2 · Clinical tasks | logitdiff: significantly reduces in four out of five tasks | Mixed |
Closest papers by meaning
- Cross-Care: Assessing the Healthcare Implications of Pre-training Data on Language Model Bias (NeurIPS 2024 · similarity 0.84)
- Hyperbolic Relational Prompts for Intersectional Fairness in Medical VLMs (CVPR 2026 · similarity 0.83)
- Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning (ICML 2025 · similarity 0.83)
- Certifying Counterfactual Bias in LLMs (ICLR 2025 · similarity 0.83)
- Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs (ICLR 2024 · similarity 0.83)
- LLM-Driven Treatment Effect Estimation Under Inference Time Text Confounding (NeurIPS 2025 · similarity 0.83)
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.
