Abstract

Large language models (LLMs) are increasingly used as agents for monitoring and governing clinical AI systems, yet it remains unclear how model size affects fairness reasoning and governance in federated clinical AI. We present AgentFL-Bench, the first benchmark for evaluating fairnessgovernance agents in federated clinical AI. Built on a two-client federated linear regression task for paediatric heart-rate prediction with controlled demographic and institutional distribution shifts, AgentFL-Bench combines a deterministic LangGraph workflow, standardized fairnessanalysis tools, and mandatory human approval to enable reproducible comparison across language models. GPT-5 was used only to generate draft reference trajectories, which were manually reviewed and revised before evaluation. We evaluate four open-source Qwen2.5 models (3B–32B) on 100 governance scenarios spanning fairness analysis, hospital comparison, mitigation planning, report generation, adversarial safety, and multi-step reasoning. Performance improves consistently up to 14B parameters, with only marginal gains from larger models, while mitigation planning remains challenging across all models. These results demonstrate the value of standardized benchmarks for evaluating fairness-governance agents and show that medium-sized open-source LLMs can provide competitive performance for clinical AI governance. Our anonymous repository can be found at https://anonymous.4open.science/r/AgentFLBench-5C95/

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MedAgent_041.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=Sr37H8mV5C

BibTex

@InProceedings{GhaFar_AgentFLBench_MICCAISAT2026,
        author = { Ghadiri, Farnoosh AND Ramburn, Tirathraj AND Mazur, Alexander AND Jouvet, Philippe},
        title = { { AgentFL-Bench: Benchmarking Small Language Models for Fairness Governance in Federated Clinical AI } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17263},
        month = {pending},
        page = {pending}
}


back to top