Extragradient Preference Optimization (EGPO): Beyond Last-Iterate Convergence for Nash Learning from Human Feedback

Reinforcement learning from human feedback (RLHF) has become essential for improving language model capabilities, but traditional approaches rely on the assumption that human preferences follow a transitive Bradley-Terry model. This assumption fails to capture the non-transitive nature of populational human preferences. Nash learning from human feedback (NLHF), targeting non-transitive preferences, is a problem of computing the Nash equilibrium (NE) of the two-player constant-sum game defined by the human preference. We introduce Extragradient preference optimization (EGPO), a novel algorithm for NLHF achieving last-iterate linear convergence to the NE of KL-regularized games and polynomial convergence to the NE of original games, while being robust to noise. Unlike previous approaches that rely on nested optimization, we derive an equivalent implementation using gradients of an online variant of the identity preference optimization (IPO) loss, enabling more faithful implementation for neural networks. Our empirical evaluations demonstrate EGPO's superior performance over baseline methods when training for the same number of epochs, as measured by pairwise win-rates using the ground truth preference. These results validate both the theoretical strengths and practical advantages of EGPO for language model alignment with non-transitive human preferences.

Keeping up with dynamic attackers: Certifying robustness to adaptive online data poisoning

The rise of foundation models fine-tuned on human feedback from potentially untrusted users has increased the risk of adversarial data poisoning, necessitating the study of robustness of learning algorithms against such attacks. Existing research on provable certified robustness against data poisoning attacks primarily focuses on certifying robustness for static adversaries who modify a fraction of the dataset used to train the model before the training algorithm is applied. In practice, particularly when learning from human feedback in an online sense, adversaries can observe and react to the learning process and inject poisoned samples that optimize adversarial objectives better than when they are restricted to poisoning a static dataset once, before the learning algorithm is applied. Indeed, it has been shown in prior work that online dynamic adversaries can be significantly more powerful than static ones. We present a novel framework for computing certified bounds on the impact of dynamic poisoning, and use these certificates to design robust learning algorithms. We give an illustration of the framework for the mean estimation and binary classification problems and outline directions for extending this in further work. The code to implement our certificates and replicate our results is available at https://github.com/Avinandan22/Certified-Robustness.

Finite Sample Identification of Partially Observed Bilinear Dynamical Systems

We consider the problem of learning a realization of a partially observed bilinear dynamical system (BLDS) from noisy input-output data. Given a single trajectory of input-output samples, we provide a finite time analysis for learning the system's Markov-like parameters, from which a balanced realization of the bilinear system can be obtained. Our bilinear system identification algorithm learns the system's Markov-like parameters by regressing the outputs to highly correlated, nonlinear, and heavy-tailed covariates. Moreover, the stability of BLDS depends on the sequence of inputs used to excite the system. These properties, unique to partially observed bilinear dynamical systems, pose significant challenges to the analysis of our algorithm for learning the unknown dynamics. We address these challenges and provide high probability error bounds on our identification algorithm under a uniform stability assumption. Our analysis provides insights into system theoretic quantities that affect learning accuracy and sample complexity. Lastly, we perform numerical experiments with synthetic data to reinforce these insights.

Hybrid Preference Optimization for Alignment: Provably Faster Convergence Rates by Combining Offline Preferences with Online Exploration

Reinforcement Learning from Human Feedback (RLHF) is currently the leading approach for aligning large language models with human preferences. Typically, these models rely on extensive offline preference datasets for training. However, offline algorithms impose strict concentrability requirements, which are often difficult to satisfy. On the other hand, while online algorithms can avoid the concentrability issue, pure online exploration could be expensive due to the active preference query cost and real-time implementation overhead. In this paper, we propose a novel approach: Hybrid Preference Optimization (HPO) which combines online exploration with existing offline preferences by relaxing the stringent concentrability conditions for offline exploration, as well as significantly improving the sample efficiency for its online counterpart. We give the first provably optimal theoretical bound for Hybrid RLHF with preference feedback, providing sample complexity bounds for policy optimization with matching lower bounds. Our results yield improved sample efficiency of hybrid RLHF over pure offline and online exploration.

Dual Approximation Policy Optimization

We propose Dual Approximation Policy Optimization (DAPO), a framework that incorporates general function approximation into policy mirror descent methods. In contrast to the popular approach of using the $L_2$-norm to measure function approximation errors, DAPO uses the dual Bregman divergence induced by the mirror map for policy projection. This duality framework has both theoretical and practical implications: not only does it achieve fast linear convergence with general function approximation, but it also includes several well-known practical methods as special cases, immediately providing strong convergence guarantees.

Toward Global Convergence of Gradient EM for Over-Parameterized Gaussian Mixture Models

We study the gradient Expectation-Maximization (EM) algorithm for Gaussian Mixture Models (GMM) in the over-parameterized setting, where a general GMM with $n>1$ components learns from data that are generated by a single ground truth Gaussian distribution. While results for the special case of 2-Gaussian mixtures are well-known, a general global convergence analysis for arbitrary $n$ remains unresolved and faces several new technical barriers since the convergence becomes sub-linear and non-monotonic. To address these challenges, we construct a novel likelihood-based convergence analysis framework and rigorously prove that gradient EM converges globally with a sublinear rate $O(1/\sqrt{t})$. This is the first global convergence result for Gaussian mixtures with more than $2$ components. The sublinear convergence rate is due to the algorithmic nature of learning over-parameterized GMM with gradient EM. We also identify a new emerging technical challenge for learning general over-parameterized GMM: the existence of bad local regions that can trap gradient EM for an exponential number of steps.

Offline Multi-task Transfer RL with Representational Penalization

We study the problem of representation transfer in offline Reinforcement Learning (RL), where a learner has access to episodic data from a number of source tasks collected a priori, and aims to learn a shared representation to be used in finding a good policy for a target task. Unlike in online RL where the agent interacts with the environment while learning a policy, in the offline setting there cannot be such interactions in either the source tasks or the target task; thus multi-task offline RL can suffer from incomplete coverage. We propose an algorithm to compute pointwise uncertainty measures for the learnt representation, and establish a data-dependent upper bound for the suboptimality of the learnt policy for the target task. Our algorithm leverages the collective exploration done by source tasks to mitigate poor coverage at some points by a few tasks, thus overcoming the limitation of needing uniformly good coverage for a meaningful transfer by existing offline algorithms. We complement our theoretical results with empirical evaluation on a rich-observation MDP which requires many samples for complete coverage. Our findings illustrate the benefits of penalizing and quantifying the uncertainty in the learnt representation.

Learning Optimal Tax Design in Nonatomic Congestion Games

We study how to learn the optimal tax design to maximize the efficiency in nonatomic congestion games. It is known that self-interested behavior among the players can damage the system's efficiency. Tax mechanisms is a common method to alleviate this issue and induce socially optimal behavior. In this work, we take the initial step for learning the optimal tax that can minimize the social cost with \emph{equilibrium feedback}, i.e., the tax designer can only observe the equilibrium state under the enforced tax. Existing algorithms are not applicable due to the exponentially large tax function space, nonexistence of the gradient, and nonconvexity of the objective. To tackle these challenges, our algorithm leverages several novel components: (1) piece-wise linear tax to approximate the optimal tax; (2) an extra linear term to guarantee a strongly convex potential function; (3) efficient subroutine to find the ``boundary'' tax. The algorithm can find an $\epsilon$-optimal tax with $O(\beta F^2/\epsilon)$ sample complexity, where $\beta$ is the smoothness of the cost function and $F$ is the number of facilities.

Initializing Services in Interactive ML Systems for Diverse Users

This paper studies ML systems that interactively learn from users across multiple subpopulations with heterogeneous data distributions. The primary objective is to provide specialized services for different user groups while also predicting user preferences. Once the users select a service based on how well the service anticipated their preference, the services subsequently adapt and refine themselves based on the user data they accumulate, resulting in an iterative, alternating minimization process between users and services (learning dynamics). Employing such tailored approaches has two main challenges: (i) Unknown user preferences: Typically, data on user preferences are unavailable without interaction, and uniform data collection across a large and diverse user base can be prohibitively expensive. (ii) Suboptimal Local Solutions: The total loss (sum of loss functions across all users and all services) landscape is not convex even if the individual losses on a single service are convex, making it likely for the learning dynamics to get stuck in local minima. The final outcome of the aforementioned learning dynamics is thus strongly influenced by the initial set of services offered to users, and is not guaranteed to be close to the globally optimal outcome. In this work, we propose a randomized algorithm to adaptively select very few users to collect preference data from, while simultaneously initializing a set of services. We prove that under mild assumptions on the loss functions, the expected total loss achieved by the algorithm right after initialization is within a factor of the globally optimal total loss with complete user preference data, and this factor scales only logarithmically in the number of services. Our theory is complemented by experiments on real as well as semi-synthetic datasets.

A/B Testing and Best-arm Identification for Linear Bandits with Robustness to Non-stationarity

We investigate the fixed-budget best-arm identification (BAI) problem for linear bandits in a potentially non-stationary environment. Given a finite arm set $\mathcal{X}\subset\mathbb{R}^d$, a fixed budget $T$, and an unpredictable sequence of parameters $\left\lbrace\theta_t\right\rbrace_{t=1}^{T}$, an algorithm will aim to correctly identify the best arm $x^* := \arg\max_{x\in\mathcal{X}}x^\top\sum_{t=1}^{T}\theta_t$ with probability as high as possible. Prior work has addressed the stationary setting where $\theta_t = \theta_1$ for all $t$ and demonstrated that the error probability decreases as $\exp(-T /\rho^*)$ for a problem-dependent constant $\rho^*$. But in many real-world $A/B/n$ multivariate testing scenarios that motivate our work, the environment is non-stationary and an algorithm expecting a stationary setting can easily fail. For robust identification, it is well-known that if arms are chosen randomly and non-adaptively from a G-optimal design over $\mathcal{X}$ at each time then the error probability decreases as $\exp(-T\Delta^2_{(1)}/d)$, where $\Delta_{(1)} = \min_{x \neq x^*} (x^* - x)^\top \frac{1}{T}\sum_{t=1}^T \theta_t$. As there exist environments where $\Delta_{(1)}^2/ d \ll 1/ \rho^*$, we are motivated to propose a novel algorithm $\mathsf{P1}$-$\mathsf{RAGE}$ that aims to obtain the best of both worlds: robustness to non-stationarity and fast rates of identification in benign settings. We characterize the error probability of $\mathsf{P1}$-$\mathsf{RAGE}$ and demonstrate empirically that the algorithm indeed never performs worse than G-optimal design but compares favorably to the best algorithms in the stationary setting.