D2PO: Discriminator-Guided DPO with Response Evaluation Models
This work addresses the challenge of improving alignment methods for language models, offering incremental gains in data efficiency and output quality for tasks like chat settings.
The paper tackled the problem of aligning language models by proposing D2PO, a discriminator-guided DPO method that uses a response evaluation model to silver-label synthetic data for policy training, resulting in higher-quality outputs and greater efficiency in preference data requirements compared to DPO.
Varied approaches for aligning language models have been proposed, including supervised fine-tuning, RLHF, and direct optimization methods such as DPO. Although DPO has rapidly gained popularity due to its straightforward training process and competitive results, there is an open question of whether there remain practical advantages of using a discriminator, like a reward model, to evaluate responses. We propose D2PO, discriminator-guided DPO, an approach for the online setting where preferences are being collected throughout learning. As we collect gold preferences, we use these not only to train our policy, but to train a discriminative response evaluation model to silver-label even more synthetic data for policy training. We explore this approach across a set of diverse tasks, including a realistic chat setting, we find that our approach leads to higher-quality outputs compared to DPO with the same data budget, and greater efficiency in terms of preference data requirements. Furthermore, we show conditions under which silver labeling is most helpful: it is most effective when training the policy with DPO, outperforming traditional PPO, and benefits from maintaining a separate discriminator from the policy model.