Counterfactual Fairness Audits of Multi-Step Clinical LLM Agents Require a Measured Per-Action Instability Floor

SummaryAI generated

For anti-factual fairness auditing of clinical multi-step large language model (LLM) agents, it is necessary to introduce a measured lower limit of single-action instability. Relying solely on the flipping rate cannot provide interpretability. After running the same conditions ten times in sixteen case scenarios, agent actions changed in 8.7% of cells. The instability difference between different actions (such as ICU upgrades and controlled operations) was as high as eight times, ranging from 0.022 to 0.179.

Related eventsRELATED EVENTS

2026-09-09 05:29OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’
2026-09-09 06:11OpenAI claims to have solved the Navier-Stokes equation problem, but its originality is questioned by the academic community
2026-09-08 22:00Google DeepMind Maps 9 Billion Possible DNA Variants

Key entitiesKEY ENTITIES

FairMedAgent

SignalsSIGNALS

Keyword heat

FairMedAgent1

All reports (1)SOURCES

↗

A arXiv cs.LG en 2026-09-07 12:00

Counterfactual audits for clinical LLM agents require a measured per-action instability floor because flip rates alone are uninterpretable. Re-running identical conditions ten times across sixteen vignettes moved agent actions in 8.7% of cells, with instability varying by a factor of eight (0.022 to 0.179) across actions like ICU escalation and controlled-s…