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 (Cornell University)

DOI: 10.48550/arxiv.2610.26088 · cited by 0 works (OpenAlex) · open-access PDF

French

Le LeakBench introduit le LeakBench, une suite d'audit vérifiable, dérivée de lignes directrices qui teste les décisions de rétroaction des agents pour les fuites de temps de prédiction, l'étalonnage, l'évaluation au niveau du site, la prise de conscience de la récompense et le bénéfice net de la courbe de décision. LeakBench combine une propriété synthétique avec une structure de défaillance connue et deux paramètres de données réelles: MIMIC-IV (430k entrées de boucle de filtre) et chaque configuration locale non restreinte choisie par le médecin. Notre analyse primaire utilise le piratage des récompenses de manière descriptive, sans attribuer l'intention: un agent

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 attributing intent: an agent selects prediction-time-invalid inputs because they improve its validation reward. A deterministic optimizer and every evaluated unconstrained local-model configuration select a fatal leaking feature. In fully logged September 2026 A/B runs, DeepSeek Flash and GPT-6-astra each leak in 10/10 unconstrained runs and 0/10 runs with an explicit admission-time constraint. This contrast is version-sensitive: a blinded Claude/Fable-5.1 chat run does not leak in either condition, while explicit-constraint violations appear only in a historical August 2026 slice lacking raw trajectories. Replacing AUROC with AUPRC does not remove leakage selection. Renaming or disguising the leaking inputs changes neither selection nor provenance-based detection; a name-based filter fails. A deterministic provenance verifier-in-loop removes the leak, making the LLM repair loop defense-in-depth. Specific verifier feedback clears fatal leakage in a hosted agent, and a three-seed ablation finds principle-only feedback clears decision-level defects, though these mitigation results are limited. LeakBench treats trustworthy clinical prediction as a property of an agent's modeling process, not merely its score; it is an audit testbed with known-failure controls, not yet an expert-validated benchmark.

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

Topics: Computer science, Audit, Hacker, Task (project management), Leakage (economics), Leak

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