Regularized Best-of-N Sampling to Mitigate Reward Hacking for Language Model Alignment

Best-of-N (BoN) sampling with a reward model has been shown to be an effective strategy for aligning Large Language Models (LLMs) to human preferences at the time of decoding. BoN sampling is susceptible to a problem known as reward hacking. Because the reward model is an imperfect proxy for the true objective, over-optimizing its value can compromise its performance on the true objective. A common solution to prevent reward hacking in preference learning techniques is to optimize a reward using proximity regularization (e.g., KL regularization), which ensures that the language model remains close to the reference model. In this research, we propose Regularized Best-of-N (RBoN), a variant of BoN that aims to mitigate reward hacking by incorporating a proximity term in response selection, similar to preference learning techniques. We evaluate two variants of RBoN on the AlpacaFarm dataset and find that they outperform BoN, especially when the proxy reward model has a low correlation with the true objective.

Scalable and Provably Fair Exposure Control for Large-Scale Recommender Systems

Typical recommendation and ranking methods aim to optimize the satisfaction of users, but they are often oblivious to their impact on the items (e.g., products, jobs, news, video) and their providers. However, there has been a growing understanding that the latter is crucial to consider for a wide range of applications, since it determines the utility of those being recommended. Prior approaches to fairness-aware recommendation optimize a regularized objective to balance user satisfaction and item fairness based on some notion such as exposure fairness. These existing methods have been shown to be effective in controlling fairness, however, most of them are computationally inefficient, limiting their applications to only unrealistically small-scale situations. This indeed implies that the literature does not yet provide a solution to enable a flexible control of exposure in the industry-scale recommender systems where millions of users and items exist. To enable a computationally efficient exposure control even for such large-scale systems, this work develops a scalable, fast, and fair method called \emph{\textbf{ex}posure-aware \textbf{ADMM} (\textbf{exADMM})}. exADMM is based on implicit alternating least squares (iALS), a conventional scalable algorithm for collaborative filtering, but optimizes a regularized objective to achieve a flexible control of accuracy-fairness tradeoff. A particular technical challenge in developing exADMM is the fact that the fairness regularizer destroys the separability of optimization subproblems for users and items, which is an essential property to ensure the scalability of iALS. Therefore, we develop a set of optimization tools to enable yet scalable fairness control with provable convergence guarantees as a basis of our algorithm.

Return-Aligned Decision Transformer

Traditional approaches in offline reinforcement learning aim to learn the optimal policy that maximizes the cumulative reward, also known as return. However, as applications broaden, it becomes increasingly crucial to train agents that not only maximize the returns, but align the actual return with a specified target return, giving control over the agent's performance. Decision Transformer (DT) optimizes a policy that generates actions conditioned on the target return through supervised learning and is equipped with a mechanism to control the agent using the target return. Despite being designed to align the actual return with the target return, we have empirically identified a discrepancy between the actual return and the target return in DT. In this paper, we propose Return-Aligned Decision Transformer (RADT), designed to effectively align the actual return with the target return. Our model decouples returns from the conventional input sequence, which typically consists of returns, states, and actions, to enhance the relationships between returns and states, as well as returns and actions. Extensive experiments show that RADT reduces the discrepancies between the actual return and the target return of DT-based methods.

Learning Fair Division from Bandit Feedback

This work addresses learning online fair division under uncertainty, where a central planner sequentially allocates items without precise knowledge of agents' values or utilities. Departing from conventional online algorithm, the planner here relies on noisy, estimated values obtained after allocating items. We introduce wrapper algorithms utilizing \textit{dual averaging}, enabling gradual learning of both the type distribution of arriving items and agents' values through bandit feedback. This approach enables the algorithms to asymptotically achieve optimal Nash social welfare in linear Fisher markets with agents having additive utilities. We establish regret bounds in Nash social welfare and empirically validate the superior performance of our proposed algorithms across synthetic and empirical datasets.

Model-Based Minimum Bayes Risk Decoding

Minimum Bayes Risk (MBR) decoding has been shown to be a powerful alternative to beam search decoding in a variety of text generation tasks. MBR decoding selects a hypothesis from a pool of hypotheses that has the least expected risk under a probability model according to a given utility function. Since it is impractical to compute the expected risk exactly over all possible hypotheses, two approximations are commonly used in MBR. First, it integrates over a sampled set of hypotheses rather than over all possible hypotheses. Second, it estimates the probability of each hypothesis using a Monte Carlo estimator. While the first approximation is necessary to make it computationally feasible, the second is not essential since we typically have access to the model probability at inference time. We propose Model-Based MBR (MBMBR), a variant of MBR that uses the model probability itself as the estimate of the probability distribution instead of the Monte Carlo estimate. We show analytically and empirically that the model-based estimate is more promising than the Monte Carlo estimate in text generation tasks. Our experiments show that MBMBR outperforms MBR in several text generation tasks, both with encoder-decoder models and with large language models.

Why Guided Dialog Policy Learning performs well? Understanding the role of adversarial learning and its alternative

Dialog policies, which determine a system's action based on the current state at each dialog turn, are crucial to the success of the dialog. In recent years, reinforcement learning (RL) has emerged as a promising option for dialog policy learning (DPL). In RL-based DPL, dialog policies are updated according to rewards. The manual construction of fine-grained rewards, such as state-action-based ones, to effectively guide the dialog policy is challenging in multi-domain task-oriented dialog scenarios with numerous state-action pair combinations. One way to estimate rewards from collected data is to train the reward estimator and dialog policy simultaneously using adversarial learning (AL). Although this method has demonstrated superior performance experimentally, it is fraught with the inherent problems of AL, such as mode collapse. This paper first identifies the role of AL in DPL through detailed analyses of the objective functions of dialog policy and reward estimator. Next, based on these analyses, we propose a method that eliminates AL from reward estimation and DPL while retaining its advantages. We evaluate our method using MultiWOZ, a multi-domain task-oriented dialog corpus.

A Slingshot Approach to Learning in Monotone Games

In this paper, we address the problem of computing equilibria in monotone games. The traditional Follow the Regularized Leader algorithms fail to converge to an equilibrium even in two-player zero-sum games. Although optimistic versions of these algorithms have been proposed with last-iterate convergence guarantees, they require noiseless gradient feedback. To overcome this limitation, we present a novel framework that achieves last-iterate convergence even in the presence of noise. Our key idea involves perturbing or regularizing the payoffs or utilities of the games. This perturbation serves to pull the current strategy to an anchored strategy, which we refer to as a {\it slingshot} strategy. First, we establish the convergence rates of our framework to a stationary point near an equilibrium, regardless of the presence or absence of noise. Next, we introduce an approach to periodically update the slingshot strategy with the current strategy. We interpret this approach as a proximal point method and demonstrate its last-iterate convergence. Our framework is comprehensive, incorporating existing payoff-regularized algorithms and enabling the development of new algorithms with last-iterate convergence properties. Finally, we show that our algorithms, based on this framework, empirically exhibit faster convergence.

Exploration of Unranked Items in Safe Online Learning to Re-Rank

Bandit algorithms for online learning to rank (OLTR) problems often aim to maximize long-term revenue by utilizing user feedback. From a practical point of view, however, such algorithms have a high risk of hurting user experience due to their aggressive exploration. Thus, there has been a rising demand for safe exploration in recent years. One approach to safe exploration is to gradually enhance the quality of an original ranking that is already guaranteed acceptable quality. In this paper, we propose a safe OLTR algorithm that efficiently exchanges one of the items in the current ranking with an item outside the ranking (i.e., an unranked item) to perform exploration. We select an unranked item optimistically to explore based on Kullback-Leibler upper confidence bounds (KL-UCB) and safely re-rank the items including the selected one. Through experiments, we demonstrate that the proposed algorithm improves long-term regret from baselines without any safety violation.

Fair Matrix Factorisation for Large-Scale Recommender Systems

Modern recommender systems are hedged with various requirements, such as ranking quality, optimisation efficiency, and item fairness. It is challenging to reconcile these requirements at a practical level. In this study, we argue that item fairness is particularly hard to optimise in a large-scale setting. The notion of item fairness requires controlling the opportunity of items (e.g. exposure) by considering the entire ranked lists for users. It hence breaks the independence of optimisation subproblems for users and items, which is the essential property for conventional scalable algorithms, such as implicit alternating least squares (iALS). This paper explores a collaborative filtering method for fairness-aware item recommendation, achieving computational efficiency comparable to iALS, the most efficient method for item recommendation.

Last-Iterate Convergence with Full- and Noisy-Information Feedback in Two-Player Zero-Sum Games

The theory of learning in games is prominent in the AI community, motivated by several rising applications such as multi-agent reinforcement learning and Generative Adversarial Networks. We propose Mutation-driven Multiplicative Weights Update (M2WU) for learning an equilibrium in two-player zero-sum normal-form games and prove that it exhibits the last-iterate convergence property in both full- and noisy-information feedback settings. In the full-information feedback setting, the players observe their exact gradient vectors of the utility functions. On the other hand, in the noisy-information feedback setting, they can only observe the noisy gradient vectors. Existing algorithms, including the well-known Multiplicative Weights Update (MWU) and Optimistic MWU (OMWU) algorithms, fail to converge to a Nash equilibrium with noisy-information feedback. In contrast, M2WU exhibits the last-iterate convergence to a stationary point near a Nash equilibrium in both of the feedback settings. We then prove that it converges to an exact Nash equilibrium by adapting the mutation term iteratively. We empirically confirm that M2WU outperforms MWU and OMWU in exploitability and convergence rates.