Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences?
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866916766605639680 |
|---|---|
| author | Gölz, Paul Haghtalab, Nika Yang, Kunhe |
| author_facet | Gölz, Paul Haghtalab, Nika Yang, Kunhe |
| contents | After pre-training, large language models are aligned with human preferences based on pairwise comparisons. State-of-the-art alignment methods (such as PPO-based RLHF and DPO) are built on the assumption of aligning with a single preference model, despite being deployed in settings where users have diverse preferences. As a result, it is not even clear that these alignment methods produce models that satisfy users on average -- a minimal requirement for pluralistic alignment. Drawing on social choice theory and modeling users' comparisons through individual Bradley-Terry (BT) models, we introduce an alignment method's distortion: the worst-case ratio between the optimal achievable average utility, and the average utility of the learned policy.
The notion of distortion helps draw sharp distinctions between alignment methods: Nash Learning from Human Feedback achieves the minimax optimal distortion of $(\frac{1}{2} + o(1)) \cdot β$ (for the BT temperature $β$), robustly across utility distributions, distributions of comparison pairs, and permissible KL divergences from the reference policy. RLHF and DPO, by contrast, suffer $\geq (1 - o(1)) \cdot β$ distortion already without a KL constraint, and $e^{Ω(β)}$ or even unbounded distortion in the full setting, depending on how comparison pairs are sampled. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_23749 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences? Gölz, Paul Haghtalab, Nika Yang, Kunhe Machine Learning Computer Science and Game Theory After pre-training, large language models are aligned with human preferences based on pairwise comparisons. State-of-the-art alignment methods (such as PPO-based RLHF and DPO) are built on the assumption of aligning with a single preference model, despite being deployed in settings where users have diverse preferences. As a result, it is not even clear that these alignment methods produce models that satisfy users on average -- a minimal requirement for pluralistic alignment. Drawing on social choice theory and modeling users' comparisons through individual Bradley-Terry (BT) models, we introduce an alignment method's distortion: the worst-case ratio between the optimal achievable average utility, and the average utility of the learned policy. The notion of distortion helps draw sharp distinctions between alignment methods: Nash Learning from Human Feedback achieves the minimax optimal distortion of $(\frac{1}{2} + o(1)) \cdot β$ (for the BT temperature $β$), robustly across utility distributions, distributions of comparison pairs, and permissible KL divergences from the reference policy. RLHF and DPO, by contrast, suffer $\geq (1 - o(1)) \cdot β$ distortion already without a KL constraint, and $e^{Ω(β)}$ or even unbounded distortion in the full setting, depending on how comparison pairs are sampled. |
| title | Distortion of AI Alignment: Does Preference Optimization Optimize for Preferences? |
| topic | Machine Learning Computer Science and Game Theory |
| url | https://arxiv.org/abs/2505.23749 |