Abstract:Alignment of Large Language models (LLMs) is crucial for safe and trustworthy deployment in applications. Reinforcement learning from human feedback (RLHF) has emerged as an effective technique to align LLMs to human preferences and broader utilities, but it requires updating billions of model parameters, which is computationally expensive. Controlled Decoding, by contrast, provides a mechanism for aligning a model at inference time without retraining. However, single-agent decoding approaches often struggle to adapt to diverse tasks due to the complexity and variability inherent in these tasks. To strengthen the test-time performance w.r.t the target task, we propose a mixture of agent-based decoding strategies leveraging the existing off-the-shelf aligned LLM policies. Treating each prior policy as an agent in the spirit of mixture of agent collaboration, we develop a decoding method that allows for inference-time alignment through a token-level selection strategy among multiple agents. For each token, the most suitable LLM is dynamically chosen from a pool of models based on a long-term utility metric. This policy-switching mechanism ensures optimal model selection at each step, enabling efficient collaboration and alignment among LLMs during decoding. Theoretical analysis of our proposed algorithm establishes optimal performance with respect to the target task represented via a target reward for the given off-the-shelf models. We conduct comprehensive empirical evaluations with open-source aligned models on diverse tasks and preferences, which demonstrates the merits of this approach over single-agent decoding baselines. Notably, Collab surpasses the current SoTA decoding strategy, achieving an improvement of up to 1.56x in average reward and 71.89% in GPT-4 based win-tie rate.
Abstract:In this paper, we study inverse game theory (resp. inverse multiagent learning) in which the goal is to find parameters of a game's payoff functions for which the expected (resp. sampled) behavior is an equilibrium. We formulate these problems as generative-adversarial (i.e., min-max) optimization problems, for which we develop polynomial-time algorithms to solve, the former of which relies on an exact first-order oracle, and the latter, a stochastic one. We extend our approach to solve inverse multiagent simulacral learning in polynomial time and number of samples. In these problems, we seek a simulacrum, meaning parameters and an associated equilibrium that replicate the given observations in expectation. We find that our approach outperforms the widely-used ARIMA method in predicting prices in Spanish electricity markets based on time-series data.
Abstract:The Koopman operator provides a powerful framework for representing the dynamics of general nonlinear dynamical systems. Data-driven techniques to learn the Koopman operator typically assume that the chosen function space is closed under system dynamics. In this paper, we study the Koopman operator via its action on the reproducing kernel Hilbert space (RKHS), and explore the mis-specified scenario where the dynamics may escape the chosen function space. We relate the Koopman operator to the conditional mean embeddings (CME) operator and then present an operator stochastic approximation algorithm to learn the Koopman operator iteratively with control over the complexity of the representation. We provide both asymptotic and finite-time last-iterate guarantees of the online sparse learning algorithm with trajectory-based sampling with an analysis that is substantially more involved than that for finite-dimensional stochastic approximation. Numerical examples confirm the effectiveness of the proposed algorithm.
Abstract:Mechanism design in resource allocation studies dividing limited resources among self-interested agents whose satisfaction with the allocation depends on privately held utilities. We consider the problem in a payment-free setting, with the aim of maximizing social welfare while enforcing incentive compatibility (IC), i.e., agents cannot inflate allocations by misreporting their utilities. The well-known proportional fairness (PF) mechanism achieves the maximum possible social welfare but incurs an undesirably high exploitability (the maximum unilateral inflation in utility from misreport and a measure of deviation from IC). In fact, it is known that no mechanism can achieve the maximum social welfare and exact incentive compatibility (IC) simultaneously without the use of monetary incentives (Cole et al., 2013). Motivated by this fact, we propose learning an approximate mechanism that desirably trades off the competing objectives. Our main contribution is to design an innovative neural network architecture tailored to the resource allocation problem, which we name Regularized Proportional Fairness Network (RPF-Net). RPF-Net regularizes the output of the PF mechanism by a learned function approximator of the most exploitable allocation, with the aim of reducing the incentive for any agent to misreport. We derive generalization bounds that guarantee the mechanism performance when trained under finite and out-of-distribution samples and experimentally demonstrate the merits of the proposed mechanism compared to the state-of-the-art.
Abstract:Equivariant neural networks have shown great success in reinforcement learning, improving sample efficiency and generalization when there is symmetry in the task. However, in many problems, only approximate symmetry is present, which makes imposing exact symmetry inappropriate. Recently, approximately equivariant networks have been proposed for supervised classification and modeling physical systems. In this work, we develop approximately equivariant algorithms in reinforcement learning (RL). We define approximately equivariant MDPs and theoretically characterize the effect of approximate equivariance on the optimal Q function. We propose novel RL architectures using relaxed group convolutions and experiment on several continuous control domains and stock trading with real financial data. Our results demonstrate that approximate equivariance matches prior work when exact symmetries are present, and outperforms them when domains exhibit approximate symmetry. As an added byproduct of these techniques, we observe increased robustness to noise at test time.
Abstract:Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining.
Abstract:The standard contextual bandit framework assumes fully observable and actionable contexts. In this work, we consider a new bandit setting with partially observable, correlated contexts and linear payoffs, motivated by the applications in finance where decision making is based on market information that typically displays temporal correlation and is not fully observed. We make the following contributions marrying ideas from statistical signal processing with bandits: (i) We propose an algorithmic pipeline named EMKF-Bandit, which integrates system identification, filtering, and classic contextual bandit algorithms into an iterative method alternating between latent parameter estimation and decision making. (ii) We analyze EMKF-Bandit when we select Thompson sampling as the bandit algorithm and show that it incurs a sub-linear regret under conditions on filtering. (iii) We conduct numerical simulations that demonstrate the benefits and practical applicability of the proposed pipeline.
Abstract:Reinforcement Learning from Human Feedback (RLHF) is a key method for aligning large language models (LLMs) with human preferences. However, current offline alignment approaches like DPO, IPO, and SLiC rely heavily on fixed preference datasets, which can lead to sub-optimal performance. On the other hand, recent literature has focused on designing online RLHF methods but still lacks a unified conceptual formulation and suffers from distribution shift issues. To address this, we establish that online LLM alignment is underpinned by bilevel optimization. By reducing this formulation to an efficient single-level first-order method (using the reward-policy equivalence), our approach generates new samples and iteratively refines model alignment by exploring responses and regulating preference labels. In doing so, we permit alignment methods to operate in an online and self-improving manner, as well as generalize prior online RLHF methods as special cases. Compared to state-of-the-art iterative RLHF methods, our approach significantly improves alignment performance on open-sourced datasets with minimal computational overhead.
Abstract:The conditional mean embedding (CME) encodes Markovian stochastic kernels through their actions on probability distributions embedded within the reproducing kernel Hilbert spaces (RKHS). The CME plays a key role in several well-known machine learning tasks such as reinforcement learning, analysis of dynamical systems, etc. We present an algorithm to learn the CME incrementally from data via an operator-valued stochastic gradient descent. As is well-known, function learning in RKHS suffers from scalability challenges from large data. We utilize a compression mechanism to counter the scalability challenge. The core contribution of this paper is a finite-sample performance guarantee on the last iterate of the online compressed operator learning algorithm with fast-mixing Markovian samples, when the target CME may not be contained in the hypothesis space. We illustrate the efficacy of our algorithm by applying it to the analysis of an example dynamical system.
Abstract:In the context of average-reward reinforcement learning, the requirement for oracle knowledge of the mixing time, a measure of the duration a Markov chain under a fixed policy needs to achieve its stationary distribution-poses a significant challenge for the global convergence of policy gradient methods. This requirement is particularly problematic due to the difficulty and expense of estimating mixing time in environments with large state spaces, leading to the necessity of impractically long trajectories for effective gradient estimation in practical applications. To address this limitation, we consider the Multi-level Actor-Critic (MAC) framework, which incorporates a Multi-level Monte Carlo (MLMC) gradient estimator. With our approach, we effectively alleviate the dependency on mixing time knowledge, a first for average-reward MDPs global convergence. Furthermore, our approach exhibits the tightest-available dependence of $\mathcal{O}\left( \sqrt{\tau_{mix}} \right)$ relative to prior work. With a 2D gridworld goal-reaching navigation experiment, we demonstrate that MAC achieves higher reward than a previous PG-based method for average reward, Parameterized Policy Gradient with Advantage Estimation (PPGAE), especially in cases with relatively small training sample budget restricting trajectory length.