Reinforcement learning from human feedback (RLHF) is a widely used method for aligning large language models with human preferences. However, RLHF often suffers from reward model overoptimisation, in which models overfit to the reward function, resulting in non-generalisable policies that exploit the idiosyncrasies and peculiarities of the reward function. A common mitigation is iterated RLHF, in which reward models are repeatedly retrained with updated human feedback and policies are re-optimised. Despite its increasing adoption, the dynamics of overoptimisation in this setting remain poorly understood. In this work, we present the first comprehensive study of overoptimisation in iterated RLHF. We systematically analyse key design choices - how reward model training data is transferred across iterations, which reward function is used for optimisation, and how policies are initialised. Using the controlled AlpacaFarm benchmark, we observe that overoptimisation tends to decrease over successive iterations, as reward models increasingly approximate ground-truth preferences. However, performance gains diminish over time, and while reinitialising from the base policy is robust, it limits optimisation flexibility. Other initialisation strategies often fail to recover from early overoptimisation. These findings offer actionable insights for building more stable and generalisable RLHF pipelines.

A simple yet effective method for inference-time alignment of generative models is Best-of-$N$ (BoN), where $N$ outcomes are sampled from a reference policy, evaluated using a proxy reward model, and the highest-scoring one is selected. While prior work argues that BoN is almost optimal in reward vs KL tradeoffs, the effectiveness of BoN depends critically on the quality of the proxy reward model used for selection. For this purpose, we study BoN through a smooth version known as Soft Best-of-N (SBoN) and develop a theoretical framework to address this gap. We analyze the scaling behaviour of BoN by providing bounds on the KL divergence between the SBoN policy and the reference policy, offering insights into how performance varies with the number of samples. We also study the regret gap, i.e., the gap between the expected true reward under the optimal policy and the SBoN policy. Our theoretical and empirical findings show that smoothing helps SBoN mitigate reward overoptimization, especially when the quality of the proxy reward is low.

Large language models (LLMs) are powerful tools but are often expensive to deploy at scale. LLM query routing mitigates this by dynamically assigning queries to models of varying cost and quality to obtain a desired trade-off. Prior query routing approaches generate only one response from the selected model and a single response from a small (inexpensive) model was often not good enough to beat a response from a large (expensive) model due to which they end up overusing the large model and missing out on potential cost savings. However, it is well known that for small models, generating multiple responses and selecting the best can enhance quality while remaining cheaper than a single large-model response. We leverage this idea to propose BEST-Route, a novel routing framework that chooses a model and the number of responses to sample from it based on query difficulty and the quality thresholds. Experiments on real-world datasets demonstrate that our method reduces costs by up to 60% with less than 1% performance drop.

This work proposes `PET', a novel pessimistic reward fine-tuning method, to learn a pessimistic reward model robust against reward hacking in offline reinforcement learning from human feedback (RLHF). Traditional reward modeling techniques in RLHF train an imperfect reward model, on which a KL regularization plays a pivotal role in mitigating reward hacking when optimizing a policy. Such an intuition-based method still suffers from reward hacking, and the policies with large KL divergence from the dataset distribution are excluded during learning. In contrast, we show that when optimizing a policy on a pessimistic reward model fine-tuned through PET, reward hacking can be prevented without relying on any regularization. We test our methods on the standard TL;DR summarization dataset. We find that one can learn a high-quality policy on our pessimistic reward without using any regularization. Such a policy has a high KL divergence from the dataset distribution while having high performance in practice. In summary, our work shows the feasibility of learning a pessimistic reward model against reward hacking. The agent can greedily search for the policy with a high pessimistic reward without suffering from reward hacking.

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) with human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking. Since the reward model is an imperfect proxy for the true objective, an excessive focus on optimizing its value can lead to a compromise of its performance on the true objective. Previous work proposes Regularized BoN sampling (RBoN), a BoN sampling with regularization to the objective, and shows that it outperforms BoN sampling so that it mitigates reward hacking and empirically (Jinnai et al., 2024). However, Jinnai et al. (2024) introduce RBoN based on a heuristic and they lack the analysis of why such regularization strategy improves the performance of BoN sampling. The aim of this study is to analyze the effect of BoN sampling on regularization strategies. Using the regularization strategies corresponds to robust optimization, which maximizes the worst case over a set of possible perturbations in the proxy reward. Although the theoretical guarantees are not directly applicable to RBoN, RBoN corresponds to a practical implementation. This paper proposes an extension of the RBoN framework, called Stochastic RBoN sampling (SRBoN), which is a theoretically guaranteed approach to worst-case RBoN in proxy reward. We then perform an empirical evaluation using the AlpacaFarm and Anthropic's hh-rlhf datasets to evaluate which factors of the regularization strategies contribute to the improvement of the true proxy reward. In addition, we also propose another simple RBoN method, the Sentence Length Regularized BoN, which has a better performance in the experiment as compared to the previous methods.

To ensure that large language model (LLM) responses are helpful and non-toxic, a reward model trained on human preference data is usually used. LLM responses with high rewards are then selected through best-of-$n$ (BoN) sampling or the LLM is further optimized to produce responses with high rewards through reinforcement learning from human feedback (RLHF). However, these processes are susceptible to reward overoptimization or `hacking', where responses receive high rewards due to imperfections in the reward model rather than true preference, particularly as prompts or responses deviate from the training data. To address these challenges, we propose to train a Bayesian reward model, which signals higher uncertainty further from the training data distribution. We trained Bayesian reward models using Laplace approximation on LoRA weights, and found that the resulting uncertainty estimates can effectively mitigate reward overoptimization in BoN sampling.