Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution
This addresses a key problem in preference learning for AI alignment by mitigating over-optimization without needing data distribution knowledge, though it is incremental as it builds on DPO-style methods.
The paper tackles the over-optimization issue in Direct Preference Optimization by introducing PEPO, a single-step algorithm that uses an ensemble of policies trained on disjoint data subsets and aggregates them pessimistically, achieving sample complexity guarantees that avoid all-policy concentrability and showing practical performance while retaining simplicity.
We introduce PEPO (Pessimistic Ensemble based Preference Optimization), a single-step Direct Preference Optimization (DPO)-like algorithm to mitigate the well-known over-optimization issue in preference learning without requiring the knowledge of the data-generating distribution or learning an explicit reward model. PEPO achieves pessimism via an ensemble of preference-optimized policies trained on disjoint data subsets and then aggregates them through a worst case construction that favors the agreement across models. In the tabular setting, PEPO achieves sample complexity guarantees depending only on a single-policy concentrability coefficient, thus avoiding the all-policy concentrability which affects the guarantees of algorithms prone to over-optimization, such as DPO. The theoretical findings are corroborated by a convincing practical performance, while retaining the simplicity and the practicality of DPO-style training.