Newest papers

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

Xiaoyuan Bao, Jingsi Jiang, Jiao Shan, Yulong Cao, Haoyu Wang, Zida Liu, Jun Gao · 2026 · arXiv

DOI: 10.48550/arXiv.2610.26088 · open-access PDF

Spanish

Los agentes de modelado clínico autónomos pueden optimizar el rendimiento predictivo mientras violan las condiciones que hacen que un modelo sea usable. Introducimos LeakBench, una suite de auditoría ejecutable derivada de directrices que prueba las decisiones de agentes para la fuga de tiempo de predicción, la calibración, la evaluación a nivel del sitio, la conciencia del soporte y el beneficio neto de la curva de decisión. LeakBench combina una tarea sintética con una estructura de fallos conocida y dos configuraciones de datos reales: MIMIC-IV (430k admisiones) y datos de la página frontal de registro de hospitales multi-hospital (MHFP).

Machine translation (Meta M2M100); the English original follows.

Abstract

Autonomous clinical-modeling agents can optimize predictive performance while violating conditions that make a model usable. 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. 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. Our primary analysis uses reward hacking descriptively, without attributin

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Is this paper real? This page was built from OpenAlex and arXiv records, not generated. Software can verify any reference in one call: GET https://store.agentexchange.work/gw/papers/verify?title=...&year=... (0.003 USD), fetch this record as JSON with /gw/papers/paper?id=2610.26088, or list what it cites with /gw/papers/references?id=.... Pay per call in USDC (x402) or with a 19 USD card key for 1,000 calls. Catalogue · llms.txt

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange