Repairing Reward Functions with Feedback to Mitigate Reward Hacking
Fuente:
arXiv
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866910004010811392 |
|---|---|
| author | Hatgis-Kessell, Stephane Bhamidipaty, Logan Mondal Brunskill, Emma |
| author_facet | Hatgis-Kessell, Stephane Bhamidipaty, Logan Mondal Brunskill, Emma |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2510_13036 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Repairing Reward Functions with Feedback to Mitigate Reward Hacking Hatgis-Kessell, Stephane Bhamidipaty, Logan Mondal Brunskill, Emma Artificial Intelligence Machine Learning 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. |
| title | Repairing Reward Functions with Feedback to Mitigate Reward Hacking |
| topic | Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2510.13036 |