The Effects of Reward Misspecification: Mapping and Mitigating Misaligned Models

TL;DR

This paper systematically studies reward hacking in RL, revealing how agent capabilities influence exploitation of misspecified rewards and identifying phase transitions that challenge safety monitoring.

Contribution

It introduces four RL environments with misspecified rewards, analyzes the impact of agent capabilities, and proposes anomaly detection methods for unsafe policies.

Findings

More capable agents exploit reward misspecifications more

Identification of capability thresholds causing behavior shifts

Baseline anomaly detectors for policy safety

Abstract

Reward hacking -- where RL agents exploit gaps in misspecified reward functions -- has been widely observed, but not yet systematically studied. To understand how reward hacking arises, we construct four RL environments with misspecified rewards. We investigate reward hacking as a function of agent capabilities: model capacity, action space resolution, observation space noise, and training time. More capable agents often exploit reward misspecifications, achieving higher proxy reward and lower true reward than less capable agents. Moreover, we find instances of phase transitions: capability thresholds at which the agent's behavior qualitatively shifts, leading to a sharp decrease in the true reward. Such phase transitions pose challenges to monitoring the safety of ML systems. To address this, we propose an anomaly detection task for aberrant policies and offer several baseline detectors.