ICLR 2025 Papers

Applying reinforcement learning (RL) to real-world problems is often made challenging by the inability to interact with the environment and the difficulty of designing reward functions. Offline RL addresses the first challenge by considering access to an offline dataset of environment interactions labeled by the reward function. In contrast, Preference-based RL does not assume access to the reward function and learns it from preferences, but typically requires an online interaction with the environment. We bridge the gap between these frameworks by exploring efficient methods for acquiring preference feedback in a fully offline setup. We propose Sim-OPRL, an offline preference-based reinforcement learning algorithm, which leverages a learned environment model to elicit preference feedback on simulated rollouts. Drawing on insights from both the offline RL and the preference-based RL literature, our algorithm employs a pessimistic approach for out-of-distribution data, and an optimistic approach for acquiring informative preferences about the optimal policy. We provide theoretical guarantees regarding the sample complexity of our approach, dependent on how well the offline data covers the optimal policy. Finally, we demonstrate the empirical performance of Sim-OPRL in various environments.

**InvestESG** is a novel multi-agent reinforcement learning (MARL) benchmark designed to study the impact of Environmental, Social, and Governance (ESG) disclosure mandates on corporate climate investments. The benchmark models an intertemporal social dilemma where companies balance short-term profit losses from climate mitigation efforts and long-term benefits from reducing climate risk, while ESG-conscious investors attempt to influence corporate behavior through their investment decisions. Companies allocate capital across mitigation, greenwashing, and resilience, with varying strategies influencing climate outcomes and investor preferences. We are releasing open-source versions of InvestESG in both PyTorch and JAX, which enable scalable and hardware-accelerated simulations for investigating competing incentives in mitigate climate change. Our experiments show that without ESG-conscious investors with sufficient capital, corporate mitigation efforts remain limited under the disclosure mandate. However, when a critical mass of investors prioritizes ESG, corporate cooperation increases, which in turn reduces climate risks and enhances long-term financial stability. Additionally, providing more information about global climate risks encourages companies to invest more in mitigation, even without investor involvement. Our findings align with empirical research using real-world data, highlighting MARL's potential to inform policy by providing insights into large-scale socio-economic challenges through efficient testing of alternative policy and market designs.

As LLMs become capable of complex tasks, there is growing potential for personalized interactions tailored to the subtle and idiosyncratic preferences of the user. We present a public benchmark, PersonalLLM, focusing on adapting LLMs to provide maximal benefits for a particular user. Departing from existing alignment benchmarks that implicitly assume uniform preferences, we curate open-ended prompts paired with many high-quality answers over which users would be expected to display heterogeneous latent preferences. Instead of persona prompting LLMs based on high-level attributes (e.g., user race or response length), which yields homogeneous preferences relative to humans, we develop a method that can simulate a large user base with diverse preferences from a set of pre-trained reward models. Our dataset and generated personalities offer an innovative testbed for developing personalization algorithms that grapple with continual data sparsity---few relevant feedback from the particular user---by leveraging historical data from other (similar) users. We explore basic in-context learning and meta-learning baselines to illustrate the utility of PersonalLLM and highlight the need for future methodological development.

This work studies the alignment of large language models with preference data from an imitation learning perspective. We establish a close theoretical connection between reinforcement learning from human feedback RLHF and imitation learning (IL), revealing that RLHF implicitly performs imitation learning on the preference data distribution. Building on this connection, we propose DIL, a principled framework that directly optimizes the imitation learning objective. DIL provides a unified imitation learning perspective on alignment, encompassing existing alignment algorithms as special cases while naturally introducing new variants. By bridging IL and RLHF, DIL offers new insights into alignment with RLHF. Extensive experiments demonstrate that DIL outperforms existing methods on various challenging benchmarks.

We present the first comprehensive finite-time global convergence analysis of policy gradient for infinite horizon average reward Markov decision processes (MDPs). Specifically, we focus on ergodic tabular MDPs with finite state and action spaces. Our analysis shows that the policy gradient iterates converge to the optimal policy at a sublinear rate of $O(\frac{1}{T})$, where $T$ represents the number of iterations. Performance bounds for discounted reward MDPs cannot be easily extended to average reward MDPs as the bounds grow proportional to the fifth power of the effective horizon. Recent work on such extensions makes a smoothness assumption that has not been verified. Thus, our primary contribution is in providing the first complete proof that the policy gradient algorithm converges globally for average-reward MDPs, without such an assumption. We also obtain the corresponding finite-time performance guarantees. In contrast to the existing discounted reward performance bounds, our performance bounds have an explicit dependence on constants that capture the complexity of the underlying MDP. Motivated by this observation, we reexamine and improve the existing performance bounds for discounted reward MDPs. We also present simulations that empirically validate the result.

Recent advancements in reinforcement learning (RL) have achieved great success in fine-tuning diffusion-based generative models. However, fine-tuning continuous flow-based generative models to align with arbitrary user-defined reward functions remains challenging, particularly due to issues such as policy collapse from overoptimization and the prohibitively high computational cost of likelihoods in continuous-time flows. In this paper, we propose an easy-to-use and theoretically sound RL fine-tuning method, which we term Online Reward-Weighted Conditional Flow Matching with Wasserstein-2 Regularization (ORW-CFM-W2). Our method integrates RL into the flow matching framework to fine-tune generative models with arbitrary reward functions, without relying on gradients of rewards or filtered datasets. By introducing an online reward-weighting mechanism, our approach guides the model to prioritize high-reward regions in the data manifold. To prevent policy collapse and maintain diversity, we incorporate Wasserstein-2 (W2) distance regularization into our method and derive a tractable upper bound for it in flow matching, effectively balancing exploration and exploitation of policy optimization. We provide theoretical analyses to demonstrate the convergence properties and induced data distributions of our method, establishing connections with traditional RL algorithms featuring Kullback-Leibler (KL) regularization and offering a more comprehensive understanding of the underlying mechanisms and learning behavior of our approach. Extensive experiments on tasks including target image generation, image compression, and text-image alignment demonstrate the effectiveness of our method, where our method achieves optimal policy convergence while allowing controllable trade-offs between reward maximization and diversity preservation.

The loss of plasticity in learning agents, analogous to the solidification of neural pathways in biological brains, significantly impedes learning and adaptation in reinforcement learning due to its non-stationary nature. To address this fundamental challenge, we propose a novel approach, *Neuroplastic Expansion* (NE), inspired by cortical expansion in cognitive science. NE maintains learnability and adaptability throughout the entire training process by dynamically growing the network from a smaller initial size to its full dimension. Our method is designed with three key components: (1) elastic neuron generation based on potential gradients, (2) dormant neuron pruning to optimize network expressivity, and (3) neuron consolidation via experience review to strike a balance in the plasticity-stability dilemma. Extensive experiments demonstrate that NE effectively mitigates plasticity loss and outperforms state-of-the-art methods across various tasks in MuJoCo and DeepMind Control Suite environments. NE enables more adaptive learning in complex, dynamic environments, which represents a crucial step towards transitioning deep reinforcement learning from static, one-time training paradigms to more flexible, continually adapting models.

Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate *diverse* and *effective* attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.

Control policies that can achieve high task performance and satisfy safety constraints are desirable for any system, including multi-agent systems (MAS). One promising technique for ensuring the safety of MAS is distributed control barrier functions (CBF). However, it is difficult to design distributed CBF-based policies for MAS that can tackle unknown discrete-time dynamics, partial observability, changing neighborhoods, and input constraints, especially when a distributed high-performance nominal policy that can achieve the task is unavailable. To tackle these challenges, we propose **DGPPO**, a new framework that *simultaneously* learns both a *discrete* graph CBF which handles neighborhood changes and input constraints, and a distributed high-performance safe policy for MAS with unknown discrete-time dynamics. We empirically validate our claims on a suite of multi-agent tasks spanning three different simulation engines. The results suggest that, compared with existing methods, our DGPPO framework obtains policies that achieve high task performance (matching baselines that ignore the safety constraints), and high safety rates (matching the most conservative baselines), with a *constant* set of hyperparameters across all environments.

Reward shaping is critical in reinforcement learning (RL), particularly for complex tasks where sparse rewards can hinder learning. However, choosing effective shaping rewards from a set of reward functions in a computationally efficient manner remains an open challenge. We propose Online Reward Selection and Policy Optimization (ORSO), a novel approach that frames the selection of shaping reward function as an online model selection problem. ORSO automatically identifies performant shaping reward functions without human intervention with provable regret guarantees. We demonstrate ORSO's effectiveness across various continuous control tasks. Compared to prior approaches, ORSO significantly reduces the amount of data required to evaluate a shaping reward function, resulting in superior data efficiency and a significant reduction in computational time (up to 8×). ORSO consistently identifies high-quality reward functions outperforming prior methods by more than 50% and on average identifies policies as performant as the ones learned using manually engineered reward functions by domain experts.

Reinforcement learning from human feedback (RLHF) is a key driver of quality and safety in state-of-the-art large language models. Yet, a surprisingly simple and strong inference-time strategy is Best-of-N sampling that selects the best generation among N candidates. In this paper, we propose Best-of-N Distillation (BOND), a novel RLHF algorithm that seeks to emulate Best-of-N but without its significant computational overhead at inference time. Specifically, BOND is a distribution matching algorithm that forces the distribution of generations from the policy to get closer to the Best-of-N distribution. We use the Jeffreys divergence (a linear combination of forward and backward KL) to balance between mode-covering and mode-seeking behavior, and derive an iterative formulation that utilizes a moving anchor for efficiency. We demonstrate the effectiveness of our approach and several design choices through experiments on abstractive summarization and Gemma models.

Hierarchical reinforcement learning (HRL) has the potential to solve complex long horizon tasks using temporal abstraction and increased exploration. However, hierarchical agents are difficult to train due to inherent non-stationarity. We present primitive enabled adaptive relabeling (PEAR), a two-phase approach where we first perform adaptive relabeling on a few expert demonstrations to generate efficient subgoal supervision, and then jointly optimize HRL agents by employing reinforcement learning (RL) and imitation learning (IL). We perform theoretical analysis to bound the sub-optimality of our approach and derive a joint optimization framework using RL and IL. Since PEAR utilizes only a few expert demonstrations and considers minimal limiting assumptions on the task structure, it can be easily integrated with typical off-policy \RL algorithms to produce a practical HRL approach. We perform extensive experiments on challenging environments and show that PEAR is able to outperform various hierarchical and non-hierarchical baselines and achieve upto 80% success rates in complex sparse robotic control tasks where other baselines typically fail to show significant progress. We also perform ablations to thoroughly analyze the importance of our various design choices. Finally, we perform real world robotic experiments on complex tasks and demonstrate that PEAR consistently outperforms the baselines.

Imitation learning aims at constructing an optimal policy by emulating expert demonstrations. However, the prevailing approaches in this domain typically presume that the demonstrations are optimal, an assumption that seldom holds true in the complexities of real-world applications. The data collected in practical scenarios often contains imperfections, encompassing both optimal and non-optimal examples. In this study, we propose Positive-Negative Generative Adversarial Imitation Learning (PN-GAIL), a novel approach that falls within the framework of Generative Adversarial Imitation Learning (GAIL). PN-GAIL innovatively leverages non-optimal information from imperfect demonstrations, allowing the discriminator to comprehensively assess the positive and negative risks associated with these demonstrations. Furthermore, it requires only a small subset of labeled confidence scores. Theoretical analysis indicates that PN-GAIL deviates from the non-optimal data while mimicking imperfect demonstrations. Experimental results demonstrate that PN-GAIL surpasses conventional baseline methods in dealing with imperfect demonstrations, thereby significantly augmenting the practical utility of imitation learning in real-world contexts. Our codes are available at https://github.com/QiangLiuT/PN-GAIL.

An important prerequisite for safe control is aligning the policy with the underlying constraints in the environment. In many real-world applications, due to the difficulty of manually specifying these constraints, existing works have proposed recovering constraints from expert demonstrations by solving the Inverse Constraint Learning (ICL) problem. However, ICL is inherently ill-posed, as multiple constraints can equivalently explain the experts' preferences, making the optimal solutions not uniquely identifiable. In this work, instead of focusing solely on a single constraint, we propose the novel approach of Exploratory ICL (ExICL). The goal of ExICL is to recover a diverse set of feasible constraints, thereby providing practitioners the flexibility to select the most appropriate constraint based on the practical needs of deployment. To achieve this goal, we design a generative diffusion verifier that guides the trajectory generation process using the probabilistic representation of an optimal constrained policy. By comparing these decisions with those made by expert agents, we can efficiently verify a candidate constraint. Driven by the verification feedback, ExICL implements an exploratory constraint update mechanism that strategically facilitates diversity within the collection of feasible constraints. Our empirical results demonstrate that ExICL can seamlessly and reliably generalize across different tasks and environments. The code is available at https://github.com/ZhaoRunyi/ExICL.

This paper investigates *list replicability* [Dixon et al., 2023] in the context of multi-armed (also linear) bandits (MAB). We define an algorithm $A$ for MAB to be $(\ell,\delta)$-list replicable if with probability at least $1-\delta$, $A$ has at most $\ell$ traces in independent executions even with different random bits, where a trace means sequence of arms played during an execution. For $k$-armed bandits, although the total number of traces can be $\Omega(k^T)$ for a time horizon $T$, we present several surprising upper bounds that either independent of or logarithmic of $T$: (1) a $(2^{k},\delta)$-list replicable algorithm with near-optimal regret, $\widetilde{O}({\sqrt{kT}})$, (2) a $(O(k/\delta),\delta)$-list replicable algorithm with regret $\widetilde{O}\left(\frac{k}{\delta}\sqrt{kT}\right)$, (3) a $((k+1)^{B-1}, \delta)$-list replicable algorithm with regret $\widetilde{O}(k^{\frac{3}{2}}T^{{\frac{1}{2}}+2^{-(B+1)}})$ for any integer $B>1$. On the other hand, for the sublinear regret regime, we establish a matching lower bound on the list complexity (parameter $\ell$). We prove that there is no $(k-1,\delta)$-list replicable algorithm with $o(T)$-regret. This is optimal in list complexity in the sub-linear regret regime as there is a $(k, 0)$-list replicable algorithm with $O(T^{2/3})$-regret. We further show that for linear bandits with $d$-dimensional features, there is a $\widetilde{O}(d^2T^{1/2+2^{-(B+1)}})$-regret algorithm with $((2d+1)^{B-1},\delta)$-list replicability, for $B>1$, even when the number of possible arms can be infinite.

Online unsupervised reinforcement learning (URL) can discover diverse skills via reward-free pre-training and exhibits impressive downstream task adaptation abilities through further fine-tuning. However, online URL methods face challenges in achieving zero-shot generalization, i.e., directly applying pre-trained policies to downstream tasks without additional planning or learning. In this paper, we propose a novel Dual-Value Forward-Backward representation (DVFB) framework with a contrastive entropy intrinsic reward to achieve both zero-shot generalization and fine-tuning adaptation in online URL. On the one hand, we demonstrate that poor exploration in forward-backward representations can lead to limited data diversity in online URL, impairing successor measures, and ultimately constraining generalization ability. To address this issue, the DVFB framework learns successor measures through a skill value function while promoting data diversity through an exploration value function, thus enabling zero-shot generalization. On the other hand, and somewhat surprisingly, by employing a straightforward dual-value fine-tuning scheme combined with a reward mapping technique, the pre-trained policy further enhances its performance through fine-tuning on downstream tasks, building on its zero-shot performance. Through extensive multi-task generalization experiments, DVFB demonstrates both superior zero-shot generalization (outperforming on all 12 tasks) and fine-tuning adaptation (leading on 10 out of 12 tasks) abilities, surpassing state-of-the-art URL methods.

In offline reinforcement learning (RL), managing the distribution shift between the learned policy and the static offline dataset is a persistent challenge that can result in overestimated values and suboptimal policies. Traditional offline RL methods address this by introducing conservative biases that limit exploration to well-understood regions, but they often overly restrict the agent's generalization capabilities. Recent work has sought to generate trajectories using generative models to augment the offline dataset, yet these methods still struggle with overestimating synthesized data, especially when out-of-distribution samples are produced. To overcome this issue, we propose RTDiff, a novel diffusion-based data augmentation technique that synthesizes trajectories *in reverse*, moving from unknown to known states. Such reverse generation naturally mitigates the risk of overestimation by ensuring that the agent avoids planning through unknown states. Additionally, reverse trajectory synthesis allows us to generate longer, more informative trajectories that take full advantage of diffusion models' generative strengths while ensuring reliability. We further enhance RTDiff by introducing flexible trajectory length control and improving the efficiency of the generation process through noise management. Our empirical results show that RTDiff significantly improves the performance of several state-of-the-art offline RL algorithms across diverse environments, achieving consistent and superior results by effectively overcoming distribution shift.

Robust federated learning aims to maintain reliable performance despite the presence of adversarial or misbehaving workers. While state-of-the-art (SOTA) robust distributed gradient descent (Robust-DGD) methods were proven theoretically optimal, their empirical success has often relied on pre-aggregation gradient clipping. However, existing static clipping strategies yield inconsistent results: enhancing robustness against some attacks while being ineffective or even detrimental against others. To address this limitation, we propose a principled adaptive clipping strategy, Adaptive Robust Clipping (ARC), which dynamically adjusts clipping thresholds based on the input gradients. We prove that ARC not only preserves the theoretical robustness guarantees of SOTA Robust-DGD methods but also provably improves asymptotic convergence when the model is well-initialized. Extensive experiments on benchmark image classification tasks confirm these theoretical insights, demonstrating that ARC significantly enhances robustness, particularly in highly heterogeneous and adversarial settings.