mlsanityData to 14 Oct 2026

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

arXiv 2511.00177 ↗First posted 31 Oct 2025cs.LG cs.CLcited by 2

On the map: Mechanistic interpretability → Probing deception, bias and beliefs

Abstract (arXiv)

LLMs are increasingly being used in healthcare. This promises to free physicians from drudgery, enabling better care to be delivered at scale. But the use of LLMs in this space also brings risks; for example, such models may worsen existing biases. How can we spot when LLMs are (spuriously) relying on patient race to inform predictions? In this work we assess the degree to which Sparse Autoencoders (SAEs) can reveal (and control) associations the model has made between race and stigmatizing concepts. We first identify SAE latents in Gemma-2 models which appear to correlate with Black individuals. We find that this latent activates on reasonable input sequences (e.g., "African American") but also problematic words like "incarceration". We then show that we can use this latent to steer models to generate outputs about Black patients, and further that this can induce problematic associations in model outputs as a result. For example, activating the Black latent increases the risk assigned to the probability that a patient will become "belligerent". We evaluate the degree to which such steering via latents might be useful for mitigating bias. We find that this offers improvements in simple settings, but is less successful for more realistic and complex clinical tasks. Overall, our results suggest that: SAEs may offer a useful tool in clinical applications of LLMs to identify problematic reliance on demographics but mitigating bias via SAE steering appears to be of marginal utility for realistic tasks.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Probing deception, bias and beliefs, one of 13 directions in Mechanistic interpretability. Builds on 40 of its 61 references that are in the map. First posted on arXiv 31 Oct 2025.

Builds on

40 of 61 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering20 Apr
  2. 2025Elucidating Mechanisms of Demographic Bias in LLMs for Healthcare18 Feb

Other references · 38

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  3. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  4. 2023Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Promptingcited by 1,122
  5. 2024Scaling and evaluating sparse autoencoderscited by 553
  6. 2020Null It Out: Guarding Protected Attributes by Iterative Nullspace Projectioncited by 455
  7. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  8. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  9. 2019Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphologycited by 200
  10. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 179
  11. 2020Towards Debiasing Sentence Representationscited by 171
  12. 2023Evaluating and Mitigating Discrimination in Language Model Decisionscited by 124
26 more
  1. 2024Automatically Interpreting Millions of Features in Large Language Modelscited by 113
  2. 2023Is This the Subspace You Are Looking for? An Interpretability Illusion for Subspace Activation Patchingcited by 80
  3. 2020Hurtful Words: Quantifying Biases in Clinical Contextual Word Embeddingscited by 79
  4. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  5. 2022A Systematic Study of Bias Amplificationcited by 65
  6. 2024Self-Debiasing Large Language Models: Zero-Shot Recognition and Reduction of Stereotypescited by 59
  7. 2024Applying sparse autoencoders to unlearn knowledge in language modelscited by 59
  8. 2024Thinking Fair and Slow: On the Efficacy of Structured Prompts for Debiasing Language Modelscited by 45
  9. 2025Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuningcited by 39
  10. 2024Unmasking and Quantifying Racial Bias of Large Language Models in Medical Report Generationcited by 31
  11. 2024Bias patterns in the application of LLMs for clinical decision support: A comprehensive studycited by 27
  12. 2025Sparse Autoencoders for Hypothesis Generationcited by 26
  13. 2018Creating Fair Models of Atherosclerotic Cardiovascular Disease Riskcited by 23
  14. 2025SAEs Can Improve Unlearning: Dynamic Sparse Autoencoder Guardrails for Precision Unlearning in LLMscited by 22
  15. 2025Sparse Autoencoder Features for Classifications and Transferabilitycited by 21
  16. 2022Write It Like You See It: Detectable Differences in Clinical Notes By Race Lead To Differential Model Recommendationscited by 20
  17. 2025Position: Use Sparse Autoencoders to Discover Unknownscited by 19
  18. 2024BiasDPO: Mitigating Bias in Language Models through Direct Preference Optimizationcited by 18
  19. 2025Robustly Improving LLM Fairness in Realistic Settings via Interpretabilitycited by 18
  20. 2021Q-Pain: A Question Answering Dataset to Measure Social Bias in Pain Managementcited by 16
  21. 2025CRISP: Persistent Concept Unlearning via Sparse Autoencoderscited by 11
  22. 2021On the Impact of Random Seeds on the Fairness of Clinical Classifierscited by 8
  23. 2023Retrieving Evidence from EHRs with LLMs: Possibilities and Challengescited by 8
  24. 2024A Data-Centric Approach To Generate Faithful and High Quality Patient Summaries with Large Language Modelscited by 6
  25. 2025On the Effectiveness and Generalization of Race Representations for Debiasing High-Stakes Decisionscited by 5
  26. 2025Insights into a radiology-specialised multimodal large language model with sparse autoencoderscited by 2

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026Tracing the Latent Threads: A Mechanistic Study of How LLMs Represent and Operationalize Race and Ethnicity Cuescited by 1
  2. 2026ClasSAE: Class-Aligned Sparse Autoencoders via Differentiable Feature-Class Affinitycited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026First, Do No Harm (With LLMs): Mitigating Racial Bias via Agentic Workflows
  2. 2025Evaluating the Bias in LLMs for Surveying Opinion and Decision Making in Healthcare
  3. 2024Mitigating the Risk of Health Inequity Exacerbated by Large Language Models
  4. 2024BiasDora: Exploring Hidden Biased Associations in Vision-Language Models
  5. 2025Robust or Suggestible? Exploring Non-Clinical Induction in LLM Drug-Safety Decisions
  6. 2026AgentFairBench: Do LLM Agents Discriminate When They Act?
  7. 2025Activation Steering for Bias Mitigation: An Interpretable Approach to Safer LLMs
  8. 2025Can Finetuing LLMs on Small Human Samples Increase Heterogeneity, Alignment, and Belief-Action Coherence?