Get Started
Home
Topics
Search
Library
Research questionHow can fairness audits of multi-step clinical LLM agents separate demographic disparity from stochastic action instability?A counterfactual flip can arise because an agent responds differently to a demographic descriptor or because it produces different actions under identical inputs. In multi-step clinical trajectories, this baseline instability varies by action, making an unadjusted flip rate difficult to interpret.
AI
AI Agents
Alignment & Safety
Evaluation & Benchmarks
Health
Latest papersRecent research connected to this question, newest first.Counterfactual Fairness Audits of Multi-Step Clinical LLM Agents Require a Measured Per-Action Instability FloorThe evidence uses FairMedAgent with synthetic six-stage clinical trajectories containing five model-facing decisions and fixed-form vignettes varying race, sex, age, insurance, English proficiency, and their intersections. The measured estimand counts only within-range flips between actions admitted by a published decision rule and clinician adjudication; band adjudication is still underway, so no demographic disparity result is claimed. Repeated identical-condition runs estimate a per-action instability floor, and results include a second model and five-draw majority voting; the harness, protocol, and analysis scripts are released.research paper · Sep 4, 2026
Related questions
How can we detect unfairness hidden in LLM multi-agent hiring decisions when final hire rates appear balanced?How can LLM agents make safe primary-care decisions as the action space grows?How should repeated-query audits determine whether LLM brand recommendations are reliable amid sampling and other sources of variation?How can long-horizon LLM agents learn when to group actions without overcommitting?