LeakBench: Auditing Reward Hacking through Prediction-Time Leakage in Clinical Modeling Agents

Key points

• Autonomous clinical-modeling agents can optimize predictive performance while violating conditions that make a model usable.1
• LeakBench combines a synthetic task with known failure structure and two real-data settings: MIMIC-IV (430k admissions) and governed multi-hospital inpatient record front-page (MHFP) data.1
• Our primary analysis uses reward hacking descriptively, without attributing intent: an agent selects prediction-time-invalid inputs because they improve its validation reward.1
• A deterministic optimizer and every evaluated unconstrained local-model configuration select a fatal leaking feature.1

Sources (1)

• [1]LeakBench: Auditing Reward Hacking through Prediction-Time Leakage in Clinical Modeling Agents
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 09:50 AM

  We introduce LeakBench, an executable, guideline-derived audit suite that tests agent decisions for prediction-time leakage, calibration, site-level evaluation, support awareness, and decision-curve net benefit.

  Autonomous clinical-modeling agents can optimize predictive performance while violating conditions that make a model usable.

Extractive summary: sentences quoted from the sources.