Reinforcement learning from human feedback (RLHF) is a key driver of quality and safety in state-of-the-art large language models. Yet, a surprisingly simple and strong inference-time strategy is Best-of-N sampling that selects the best generation among N candidates. In this paper, we propose Best-of-N Distillation (BOND), a novel RLHF algorithm that seeks to emulate Best-of-N but without its significant computational overhead at inference time. Specifically, BOND is a distribution matching algorithm that forces the distribution of generations from the policy to get closer to the Best-of-N distribution. We use the Jeffreys divergence (a linear combination of forward and backward KL) to balance between mode-covering and mode-seeking behavior, and derive an iterative formulation that utilizes a moving anchor for efficiency. We demonstrate the effectiveness of our approach and several design choices through experiments on abstractive summarization and Gemma models.

Reinforcement learning from human feedback (RLHF) has emerged as the main paradigm for aligning large language models (LLMs) with human preferences. Traditionally, RLHF involves the initial step of learning a reward model from pairwise human feedback, i.e., expressed as preferences between pairs of text generations. Subsequently, the LLM's policy is fine-tuned to maximize the reward through a reinforcement learning algorithm. In this study, we introduce an alternative pipeline for the fine-tuning of LLMs using pairwise human feedback. Our approach entails the initial learning of a pairwise preference model, which is conditioned on two inputs (instead of a single input in the case of a reward model) given a prompt, followed by the pursuit of a policy that consistently generates responses preferred over those generated by any competing policy, thus defining the Nash equilibrium of this preference model. We term this approach Nash learning from human feedback (NLHF). In the context of a tabular policy representation, we present a novel algorithmic solution, Nash-MD, founded on the principles of mirror descent. This algorithm produces a sequence of policies, with the last iteration converging to the regularized Nash equilibrium. Additionally, we explore parametric representations of policies and introduce gradient descent algorithms for deep-learning architectures. We illustrate the effectiveness of our approach by presenting experimental results on a text summarization task. We believe NLHF offers a compelling avenue for fine-tuning LLMs and enhancing the alignment of LLMs with human preferences.

The task of obtaining meaningful annotations is a tedious work, incurring considerable costs and time consumption. Dynamic active learning and cooperative learning are recently proposed approaches to reduce human effort of annotating data with subjective phenomena. In this paper, we introduce a novel generic annotation framework, with the aim to achieve the optimal tradeoff between label reliability and cost reduction by making efficient use of human and machine work force. To this end, we use dropout to assess model uncertainty and thereby to decide which instances can be automatically labeled by the machine and which ones require human inspection. In addition, we propose an early stopping criterion based on inter-rater agreement in order to focus human resources on those ambiguous instances that are difficult to label. In contrast to the existing algorithms, the new confidence measures are not only applicable to binary classification tasks but also regression problems. The proposed method is evaluated on the benchmark datasets for non-native English prosody estimation, provided in the Interspeech computational paralinguistics challenge. In the result, the novel dynamic cooperative learning algorithm yields 0.424 Spearman's correlation coefficient compared to 0.413 with passive learning, while reducing the amount of human annotations by 74%.