Correcting the Mythos of KL-Regularization: Direct Alignment without Overparameterization via Chi-squared Preference Optimization

Language model alignment methods, such as reinforcement learning from human feedback (RLHF), have led to impressive advances in language model capabilities, but existing techniques are limited by a widely observed phenomenon known as overoptimization, where the quality of the language model plateaus or degrades over the course of the alignment process. Overoptimization is often attributed to overfitting to an inaccurate reward model, and while it can be mitigated through online data collection, this is infeasible in many settings. This raises a fundamental question: Do existing offline alignment algorithms make the most of the data they have, or can their sample-efficiency be improved further? We address this question with a new algorithm for offline alignment, $\chi^2$-Preference Optimization ($\chi$PO). $\chi$PO is a one-line change to Direct Preference Optimization (DPO; Rafailov et al., 2023), which only involves modifying the logarithmic link function in the DPO objective. Despite this minimal change, $\chi$PO implicitly implements the principle of pessimism in the face of uncertainty via regularization with the $\chi^2$-divergence -- which quantifies uncertainty more effectively than KL-regularization -- and provably alleviates overoptimization, achieving sample-complexity guarantees based on single-policy concentrability -- the gold standard in offline reinforcement learning. $\chi$PO's simplicity and strong guarantees make it the first practical and general-purpose offline alignment algorithm that is provably robust to overoptimization.

On Speeding Up Language Model Evaluation

Large language models (LLMs) currently dominate the field of natural language processing (NLP), representing the state-of-the-art across a diverse array of tasks. Developing a model of this nature, from training to inference, requires making numerous decisions which define a combinatorial search problem. For example, selecting the optimal pre-trained LLM, prompt, or hyperparameters to attain the best performance for a task often requires evaluating multiple candidates on an entire test set. This exhaustive evaluation can be time-consuming and costly, as both inference and metric computation with LLMs are resource-intensive. In this paper, we address the challenge of identifying the best method within a limited budget for evaluating methods on test examples. By leveraging the well-studied multi-armed bandit framework, which sequentially selects the next method-example pair to evaluate, our approach, combining multi-armed bandit algorithms with low-rank factorization, significantly reduces the required resources. Experiments show that our algorithms can identify the top-performing method using only 5-15\% of the typically needed resources, resulting in an 85-95\% reduction in cost.

Orchestrating LLMs with Different Personalizations

This paper presents a novel approach to aligning large language models (LLMs) with individual human preferences, sometimes referred to as Reinforcement Learning from \textit{Personalized} Human Feedback (RLPHF). Given stated preferences along multiple dimensions, such as helpfulness, conciseness, or humor, the goal is to create an LLM without re-training that best adheres to this specification. Starting from specialized expert LLMs, each trained for one such particular preference dimension, we propose a black-box method that merges their outputs on a per-token level. We train a lightweight Preference Control Model (PCM) that dynamically translates the preference description and current context into next-token prediction weights. By combining the expert models' outputs at the token level, our approach dynamically generates text that optimizes the given preference. Empirical tests show that our method matches or surpasses existing preference merging techniques, providing a scalable, efficient alternative to fine-tuning LLMs for individual personalization.

Computationally Efficient RL under Linear Bellman Completeness for Deterministic Dynamics

We study computationally and statistically efficient Reinforcement Learning algorithms for the linear Bellman Complete setting, a setting that uses linear function approximation to capture value functions and unifies existing models like linear Markov Decision Processes (MDP) and Linear Quadratic Regulators (LQR). While it is known from the prior works that this setting is statistically tractable, it remained open whether a computationally efficient algorithm exists. Our work provides a computationally efficient algorithm for the linear Bellman complete setting that works for MDPs with large action spaces, random initial states, and random rewards but relies on the underlying dynamics to be deterministic. Our approach is based on randomization: we inject random noise into least square regression problems to perform optimistic value iteration. Our key technical contribution is to carefully design the noise to only act in the null space of the training data to ensure optimism while circumventing a subtle error amplification issue.

Understanding Preference Fine-Tuning Through the Lens of Coverage

Learning from human preference data has emerged as the dominant paradigm for fine-tuning large language models (LLMs). The two most common families of techniques -- online reinforcement learning (RL) such as Proximal Policy Optimization (PPO) and offline contrastive methods such as Direct Preference Optimization (DPO) -- were positioned as equivalent in prior work due to the fact that both have to start from the same offline preference dataset. To further expand our theoretical understanding of the similarities and differences between online and offline techniques for preference fine-tuning, we conduct a rigorous analysis through the lens of dataset coverage, a concept that captures how the training data covers the test distribution and is widely used in RL. We prove that a global coverage condition is both necessary and sufficient for offline contrastive methods to converge to the optimal policy, but a weaker partial coverage condition suffices for online RL methods. This separation provides one explanation of why online RL methods can perform better than offline methods, especially when the offline preference data is not diverse enough. Finally, motivated by our preceding theoretical observations, we derive a hybrid preference optimization (HyPO) algorithm that uses offline data for contrastive-based preference optimization and online data for KL regularization. Theoretically and empirically, we demonstrate that HyPO is more performant than its pure offline counterpart DPO, while still preserving its computation and memory efficiency.

REBEL: Reinforcement Learning via Regressing Relative Rewards

While originally developed for continuous control problems, Proximal Policy Optimization (PPO) has emerged as the work-horse of a variety of reinforcement learning (RL) applications including the fine-tuning of generative models. Unfortunately, PPO requires multiple heuristics to enable stable convergence (e.g. value networks, clipping) and is notorious for its sensitivity to the precise implementation of these components. In response, we take a step back and ask what a minimalist RL algorithm for the era of generative models would look like. We propose REBEL, an algorithm that cleanly reduces the problem of policy optimization to regressing the relative rewards via a direct policy parameterization between two completions to a prompt, enabling strikingly lightweight implementation. In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL, which allows us to match the strongest known theoretical guarantees in terms of convergence and sample complexity in the RL literature. REBEL can also cleanly incorporate offline data and handle the intransitive preferences we frequently see in practice. Empirically, we find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO, all while being simpler to implement and more computationally tractable than PPO.

Dataset Reset Policy Optimization for RLHF

Reinforcement Learning (RL) from Human Preference-based feedback is a popular paradigm for fine-tuning generative models, which has produced impressive models such as GPT-4 and Claude3 Opus. This framework often consists of two steps: learning a reward model from an offline preference dataset followed by running online RL to optimize the learned reward model. In this work, leveraging the idea of reset, we propose a new RLHF algorithm with provable guarantees. Motivated by the fact that offline preference dataset provides informative states (i.e., data that is preferred by the labelers), our new algorithm, Dataset Reset Policy Optimization (DR-PO), integrates the existing offline preference dataset into the online policy training procedure via dataset reset: it directly resets the policy optimizer to the states in the offline dataset, instead of always starting from the initial state distribution. In theory, we show that DR-PO learns to perform at least as good as any policy that is covered by the offline dataset under general function approximation with finite sample complexity. In experiments, we demonstrate that on both the TL;DR summarization and the Anthropic Helpful Harmful (HH) dataset, the generation from DR-PO is better than that from Proximal Policy Optimization (PPO) and Direction Preference Optimization (DPO), under the metric of GPT4 win-rate. Code for this work can be found at https://github.com/Cornell-RL/drpo.

Adversarial Imitation Learning via Boosting

Adversarial imitation learning (AIL) has stood out as a dominant framework across various imitation learning (IL) applications, with Discriminator Actor Critic (DAC) (Kostrikov et al.,, 2019) demonstrating the effectiveness of off-policy learning algorithms in improving sample efficiency and scalability to higher-dimensional observations. Despite DAC's empirical success, the original AIL objective is on-policy and DAC's ad-hoc application of off-policy training does not guarantee successful imitation (Kostrikov et al., 2019; 2020). Follow-up work such as ValueDICE (Kostrikov et al., 2020) tackles this issue by deriving a fully off-policy AIL objective. Instead in this work, we develop a novel and principled AIL algorithm via the framework of boosting. Like boosting, our new algorithm, AILBoost, maintains an ensemble of properly weighted weak learners (i.e., policies) and trains a discriminator that witnesses the maximum discrepancy between the distributions of the ensemble and the expert policy. We maintain a weighted replay buffer to represent the state-action distribution induced by the ensemble, allowing us to train discriminators using the entire data collected so far. In the weighted replay buffer, the contribution of the data from older policies are properly discounted with the weight computed based on the boosting framework. Empirically, we evaluate our algorithm on both controller state-based and pixel-based environments from the DeepMind Control Suite. AILBoost outperforms DAC on both types of environments, demonstrating the benefit of properly weighting replay buffer data for off-policy training. On state-based environments, DAC outperforms ValueDICE and IQ-Learn (Gary et al., 2021), achieving competitive performance with as little as one expert trajectory.

Efficient and Sharp Off-Policy Evaluation in Robust Markov Decision Processes

We study evaluating a policy under best- and worst-case perturbations to a Markov decision process (MDP), given transition observations from the original MDP, whether under the same or different policy. This is an important problem when there is the possibility of a shift between historical and future environments, due to e.g. unmeasured confounding, distributional shift, or an adversarial environment. We propose a perturbation model that can modify transition kernel densities up to a given multiplicative factor or its reciprocal, which extends the classic marginal sensitivity model (MSM) for single time step decision making to infinite-horizon RL. We characterize the sharp bounds on policy value under this model, that is, the tightest possible bounds given by the transition observations from the original MDP, and we study the estimation of these bounds from such transition observations. We develop an estimator with several appealing guarantees: it is semiparametrically efficient, and remains so even when certain necessary nuisance functions such as worst-case Q-functions are estimated at slow nonparametric rates. It is also asymptotically normal, enabling easy statistical inference using Wald confidence intervals. In addition, when certain nuisances are estimated inconsistently we still estimate a valid, albeit possibly not sharp bounds on the policy value. We validate these properties in numeric simulations. The combination of accounting for environment shifts from train to test (robustness), being insensitive to nuisance-function estimation (orthogonality), and accounting for having only finite samples to learn from (inference) together leads to credible and reliable policy evaluation.

RL for Consistency Models: Faster Reward Guided Text-to-Image Generation

Reinforcement learning (RL) has improved guided image generation with diffusion models by directly optimizing rewards that capture image quality, aesthetics, and instruction following capabilities. However, the resulting generative policies inherit the same iterative sampling process of diffusion models that causes slow generation. To overcome this limitation, consistency models proposed learning a new class of generative models that directly map noise to data, resulting in a model that can generate an image in as few as one sampling iteration. In this work, to optimize text-to-image generative models for task specific rewards and enable fast training and inference, we propose a framework for fine-tuning consistency models via RL. Our framework, called Reinforcement Learning for Consistency Model (RLCM), frames the iterative inference process of a consistency model as an RL procedure. RLCM improves upon RL fine-tuned diffusion models on text-to-image generation capabilities and trades computation during inference time for sample quality. Experimentally, we show that RLCM can adapt text-to-image consistency models to objectives that are challenging to express with prompting, such as image compressibility, and those derived from human feedback, such as aesthetic quality. Comparing to RL finetuned diffusion models, RLCM trains significantly faster, improves the quality of the generation measured under the reward objectives, and speeds up the inference procedure by generating high quality images with as few as two inference steps. Our code is available at https://rlcm.owenoertell.com