Guardado en:
| Autores principales: | , , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2026
|
| Materias: | |
| Acceso en línea: | https://arxiv.org/abs/2605.15960 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866910231550754816 |
|---|---|
| author | Bhamidipaty, Logan Mondal Whitammer, Esmeralda S. Abel, David Kochenderfer, Mykel J. Ramamoorthy, Subramanian |
| author_facet | Bhamidipaty, Logan Mondal Whitammer, Esmeralda S. Abel, David Kochenderfer, Mykel J. Ramamoorthy, Subramanian |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_15960 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Imperfect World Models are Exploitable Bhamidipaty, Logan Mondal Whitammer, Esmeralda S. Abel, David Kochenderfer, Mykel J. Ramamoorthy, Subramanian Artificial Intelligence Machine Learning 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. |
| title | Imperfect World Models are Exploitable |
| topic | Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2605.15960 |