Logan Mondal Bhamidipaty

Logan Mondal Bhamidipaty, Esmeralda S. Whitammer, David Abel et al.

We propose a novel definition of model exploitation in reinforcement learning. Informally, a world model is exploitable if it implies that one policy should be strictly preferred over another while the environment's true transition model implies the reverse. We analogize our definition with a prior characterization of reward hacking but show that the associated proof of inevitability does not transfer to exploitation. To overcome this obstruction, we develop a general theory of reward hacking and model exploitation that proves that exploitation is essentially unavoidable on large policy sets and yields the corresponding claim for hacking as a special case. Unfortunately, we also find that the conditions that guarantee unhackability in finite policy sets have no counterpart that precludes exploitation. Consequently, we introduce a relaxed notion of exploitation and derive a safe horizon within which it can be avoided. Taken together, our results establish a formal bridge between reward hacking and model exploitation and elucidate the limits of safe planning in world models.

Stephane Hatgis-Kessell, Logan Mondal Bhamidipaty, Emma Brunskill

Human-designed reward functions for reinforcement learning (RL) agents are frequently misaligned with the humans' true, unobservable objectives, and thus act only as proxies. Optimizing for a misspecified proxy reward function often induces reward hacking, resulting in a policy misaligned with the human's true objectives. An alternative is to perform RL from human feedback, which involves learning a reward function from scratch by collecting human preferences over pairs of trajectories. However, building such datasets is costly. To address the limitations of both approaches, we propose Preference-Based Reward Repair (PBRR): an automated iterative framework that repairs a human-specified proxy reward function by learning an additive, transition-dependent correction term from preferences. A manually specified reward function can yield policies that are highly suboptimal under the ground-truth objective, yet corrections on only a few transitions may suffice to recover optimal performance. To identify and correct for those transitions, PBRR uses a targeted exploration strategy and a new preference-learning objective. We prove in tabular domains PBRR has a cumulative regret that matches, up to constants, that of prior preference-based RL methods. In addition, on a suite of reward-hacking benchmarks, PBRR consistently outperforms baselines that learn a reward function from scratch from preferences or modify the proxy reward function using other approaches, requiring substantially fewer preferences to learn high performing policies.