Generative models are transforming creative domains such as music generation, with inference-time strategies like Classifier-Free Guidance (CFG) playing a crucial role. However, CFG doubles inference cost while limiting originality and diversity across generated contents. In this paper, we introduce diversity-rewarded CFG distillation, a novel finetuning procedure that distills the strengths of CFG while addressing its limitations. Our approach optimises two training objectives: (1) a distillation objective, encouraging the model alone (without CFG) to imitate the CFG-augmented predictions, and (2) an RL objective with a diversity reward, promoting the generation of diverse outputs for a given prompt. By finetuning, we learn model weights with the ability to generate high-quality and diverse outputs, without any inference overhead. This also unlocks the potential of weight-based model merging strategies: by interpolating between the weights of two models (the first focusing on quality, the second on diversity), we can control the quality-diversity trade-off at deployment time, and even further boost performance. We conduct extensive experiments on the MusicLM text-to-music generative model, where our approach surpasses CFG in terms of quality-diversity Pareto optimality. According to human evaluators, our finetuned-then-merged model generates samples with higher quality-diversity than the base model augmented with CFG. Explore our generations at https://musicdiversity.github.io/.

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.

Post-training of language models (LMs) increasingly relies on the following two stages: (i) knowledge distillation, where the LM is trained to imitate a larger teacher LM, and (ii) reinforcement learning from human feedback (RLHF), where the LM is aligned by optimizing a reward model. In the second RLHF stage, a well-known challenge is reward hacking, where the LM over-optimizes the reward model, leading to degraded performance on the true objective, in line with Goodhart’s law. In this paper, we investigate whether a similar phenomenon, that we call teacher hacking, can occur during knowledge distillation. This could arise because the teacher LM is itself an imperfect approximation of the true distribution. To study this, we propose a controlled experimental setup involving: (i) an oracle LM representing the ground-truth distribution, (ii) a teacher LM distilled from the oracle, and (iii) a student LM distilled from the teacher. Our experiments reveal the following insights. When using a fixed offline dataset for distillation, teacher hacking occurs; moreover, we can detect it by observing when the optimization process deviates from polynomial convergence laws. In contrast, employing online data generation techniques effectively mitigates teacher hacking. More precisely, we identify data diversity as the key factor in preventing hacking. Overall, our findings provide a deeper understanding of the benefits and limitations of distillation for building robust LMs.

Reinforcement learning from human feedback (RLHF) has proven effective in aligning large language models (LLMs) with human preferences, but gathering high-quality preference labels is expensive. RL from AI Feedback (RLAIF), introduced in Bai et al. (2022b), offers a promising alternative that trains the reward model (RM) on preferences generated by an off-the-shelf LLM. Across the tasks of summarization, helpful dialogue generation, and harmless dialogue generation, we show that RLAIF achieves comparable performance to RLHF. Furthermore, we take a step towards "self-improvement" by demonstrating that RLAIF can outperform a supervised fine-tuned baseline even when the AI labeler is the same size as the policy, or even the exact same checkpoint as the initial policy. Finally, we introduce direct-RLAIF (d-RLAIF) - a technique that circumvents RM training by obtaining rewards directly from an off-the-shelf LLM during RL, which achieves superior performance to canonical RLAIF. Our results suggest that RLAIF can achieve performance on-par with using human feedback, offering a potential solution to the scalability limitations of RLHF.

Aligning large language models (LLMs) with human preferences through reinforcement learning (RLHF) can lead to reward hacking, where LLMs exploit failures in the reward model (RM) to achieve seemingly high rewards without meeting the underlying objectives. We identify two primary challenges when designing RMs to mitigate reward hacking: distribution shifts during the RL process and inconsistencies in human preferences. As a solution, we propose Weight Averaged Reward Models (WARM), first fine-tuning multiple RMs, then averaging them in the weight space. This strategy follows the observation that fine-tuned weights remain linearly mode connected when sharing the same pre-training. By averaging weights, WARM improves efficiency compared to the traditional ensembling of predictions, while improving reliability under distribution shifts and robustness to preference inconsistencies. Our experiments on summarization tasks, using best-of-N and RL methods, shows that WARM improves the overall quality and alignment of LLM predictions; for example, a policy RL fine-tuned with WARM has a 79.4% win rate against a policy RL fine-tuned with a single RM.

Paul Roit, Johan Ferret, Lior Shani, Roee Aharoni, Geoffrey Cideron, Robert Dadashi, Matthieu Geist, Sertan Girgin, Leonard Hussenot, Orgad Keller, Nikola Momchev, Sabela Ramos Garea, Piotr Stanczyk, Nino Vieillard, Olivier Bachem, Gal Elidan, Avinatan Hassidim, Olivier Pietquin, Idan Szpektor. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

We propose to learn to distinguish reversible from irreversible actions for better informed decision-making in Reinforcement Learning (RL). From theoretical considerations, we show that approximate reversibility can be learned through a simple surrogate task: ranking randomly sampled trajectory events in chronological order. Intuitively, pairs of events that are always observed in the same order are likely to be separated by an irreversible sequence of actions. Conveniently, learning the temporal order of events can be done in a fully self-supervised way, which we use to estimate the reversibility of actions from experience, without any priors.We propose two different strategies that incorporate reversibility in RL agents, one strategy for exploration (RAE) and one strategy for control (RAC). We demonstrate the potential of reversibility-aware agents in several environments, including the challenging Sokoban game. In synthetic tasks, we show that we can learn control policies that never fail and reduce to zero the side-effects of interactions, even without access to the reward function.

The ability to transfer knowledge to novel environments and tasks is a sensible desiderata for general learning agents. Despite the apparent promises, transfer in RL is still an open and little exploited research area. In this paper, we take a brand-new perspective about transfer: we suggest that the ability to assign credit unveils structural invariants in the tasks that can be transferred to make RL more sample-efficient. Our main contribution is SECRET, a novel approach to transfer learning for RL that uses a backward-view credit assignment mechanism based on a self-attentive architecture. Two aspects are key to its generality: it learns to assign credit as a separate offline supervised process and exclusively modifies the reward function. Consequently, it can be supplemented by transfer methods that do not modify the reward function and it can be plugged on top of any RL algorithm.