KL Penalty Control via Perturbation for Direct Preference Optimization

Direct Preference Optimization (DPO) demonstrates the advantage of aligning a large language model with human preference using only an offline dataset. However, DPO has the limitation that the KL penalty, which prevents excessive deviation from the reference model, is static throughout the training process. Several methods try to turn this static KL penalty into a dynamic one, but no approach can adaptively assign different KL penalties for each preference pair. In this paper, we propose $\varepsilon$-Direct Preference Optimization ($\varepsilon$-DPO), which allows adaptive control of the KL penalty strength $\beta$ for each preference pair. Specifically, $\varepsilon$-DPO adaptively controls $\beta$ for each preference pair based on the monotonicity of logits as a preference model under the perturbation of $\beta$ during training by simply reusing the logit of the current policy and the reference policy. Experimental results show that $\varepsilon$-DPO outperforms existing direct alignment algorithms and KL penalty relaxation methods on general chatbot benchmarks, highlighting the significance of adaptive KL penalty relaxation at the instance-level in DPO.

External Knowledge Selection with Weighted Negative Sampling in Knowledge-grounded Task-oriented Dialogue Systems

Constructing a robust dialogue system on spoken conversations bring more challenge than written conversation. In this respect, DSTC10-Track2-Task2 is proposed, which aims to build a task-oriented dialogue (TOD) system incorporating unstructured external knowledge on a spoken conversation, extending DSTC9-Track1. This paper introduces our system containing four advanced methods: data construction, weighted negative sampling, post-training, and style transfer. We first automatically construct a large training data because DSTC10-Track2 does not release the official training set. For the knowledge selection task, we propose weighted negative sampling to train the model more fine-grained manner. We also employ post-training and style transfer for the response generation task to generate an appropriate response with a similar style to the target response. In the experiment, we investigate the effect of weighted negative sampling, post-training, and style transfer. Our model ranked 7 out of 16 teams in the objective evaluation and 6 in human evaluation.