ICLR 2025 Papers

Reinforcement learning with human feedback (RLHF) has become the dominant method to align large models to user preferences. Unlike fine-tuning, for which there are many studies regarding training data memorization, it is not clear how memorization is affected by or introduced in the RLHF alignment process. Understanding this relationship is important as real user data may be collected and used to align large models; if user data is memorized during RLHF and later regurgitated, this could raise privacy concerns. In addition to RLHF, other methods such as Direct Preference Optimization (DPO) and $\Psi$PO have gained popularity for learning directly from human preferences, removing the need for optimizing intermediary reward models with reinforcement learning. In this work, we analyze how training data memorization can surface and propagate through each phase of RLHF and direct preference learning. We focus our study on code completion models, as code completion is one of the most popular use cases for large language models. We find that RLHF significantly decreases the chance that data used for reward modeling and reinforcement learning is memorized in comparison to directly fine-tuning on this data, but that examples already memorized during the fine-tuning stage of RLHF, will, in the majority of cases, remain memorized after RLHF. In contrast, we find that aligning by learning directly from human preference data via a special case of $\Psi$PO, Identity Preference Optimization (IPO), increases the likelihood that training data is regurgitated compared to RLHF. Our work suggests that RLHF, as opposed to direct preference learning, is a safer way to mitigate the risk of regurgitating sensitive preference data when aligning large language models. We find our conclusions are robust across multiple code completion datasets, tasks, and model scales.

Reinforcement learning-based methods for constructing solutions to combinatorial optimization problems are rapidly approaching the performance of human-designed algorithms. To further narrow the gap, learning-based approaches must efficiently explore the solution space during the search process. Recent approaches artificially increase exploration by enforcing diverse solution generation through handcrafted rules, however, these rules can impair solution quality and are difficult to design for more complex problems. In this paper, we introduce PolyNet, an approach for improving exploration of the solution space by learning complementary solution strategies. In contrast to other works, PolyNet uses only a single-decoder and a training schema that does not enforce diverse solution generation through handcrafted rules. We evaluate PolyNet on four combinatorial optimization problems and observe that the implicit diversity mechanism allows PolyNet to find better solutions than approaches that explicitly enforce diverse solution generation.

A significant aspiration of offline reinforcement learning (RL) is to develop a generalist agent with high capabilities from large and heterogeneous datasets. However, prior approaches that scale offline RL either rely heavily on expert trajectories or struggle to generalize to diverse unseen tasks. Inspired by the excellent generalization of world model in conditional video generation, we explore the potential of image observation-based world model for scaling offline RL and enhancing generalization on novel tasks. In this paper, we introduce JOWA: Jointly-Optimized World-Action model, an offline model-based RL agent pretrained on multiple Atari games with 6 billion tokens data to learn general-purpose representation and decision-making ability. Our method jointly optimizes a world-action model through a shared transformer backbone, which stabilize temporal difference learning with large models during pretraining. Moreover, we propose a provably efficient and parallelizable planning algorithm to compensate for the Q-value estimation error and thus search out better policies. Experimental results indicate that our largest agent, with 150 million parameters, achieves 78.9% human-level performance on pretrained games using only 10% subsampled offline data, outperforming existing state-of-the-art large-scale offline RL baselines by 31.6% on averange. Furthermore, JOWA scales favorably with model capacity and can sample-efficiently transfer to novel games using only 5k offline fine-tuning data (approximately 4 trajectories) per game, demonstrating superior generalization.

Sparse continuous policies are distributions that can choose some actions at random yet keep strictly zero probability for the other actions, which are radically different from the Gaussian. They have important real-world implications, e.g. in modeling safety-critical tasks like medicine. The combination of offline reinforcement learning and sparse policies provides a novel paradigm that enables learning completely from logged datasets a safety-aware sparse policy. However, sparse policies can cause difficulty with the existing offline algorithms which require evaluating actions that fall outside of the current support. In this paper, we propose the first offline policy optimization algorithm that tackles this challenge: Fat-to-Thin Policy Optimization (FtTPO). Specifically, we maintain a fat (heavy-tailed) proposal policy that effectively learns from the dataset and injects knowledge to a thin (sparse) policy, which is responsible for interacting with the environment. We instantiate FtTPO with the general $q$-Gaussian family that encompasses both heavy-tailed and sparse policies and verify that it performs favorably in a safety-critical treatment simulation and the standard MuJoCo suite. Our code is available at https://github.com/lingweizhu/fat2thin.

Reinforcement learning from human feedback (RLHF) has demonstrated great promise in aligning large language models (LLMs) with human preference. Depending on the availability of preference data, both online and offline RLHF are active areas of investigation. A key bottleneck is understanding how to incorporate uncertainty estimation in the reward function learned from the preference data for RLHF, regardless of how the preference data is collected. While the principles of optimism or pessimism under uncertainty are well-established in standard reinforcement learning (RL), a practically-implementable and theoretically-grounded form amenable to large language models is not yet available, as standard techniques for constructing confidence intervals become intractable under arbitrary policy parameterizations. In this paper, we introduce a unified approach to online and offline RLHF --- value-incentivized preference optimization (VPO) --- which regularizes the maximum-likelihood estimate of the reward function with the corresponding value function, modulated by a sign to indicate whether the optimism or pessimism is chosen. VPO also directly optimizes the policy with implicit reward modeling, and therefore shares a simpler RLHF pipeline similar to direct preference optimization. Theoretical guarantees of VPO are provided for both online and offline settings, matching the rates of their standard RL counterparts. Moreover, experiments on text summarization, dialogue, and standard benchmarks verify the practicality and effectiveness of VPO.

Traditional language model-based theorem proving assumes that by training on a sufficient amount of formal proof data, a model will learn to prove theorems. Our key observation is that a wealth of informal information that is not present in formal proofs can be useful for learning to prove theorems. For instance, humans think through steps of a proof, but this thought process is not visible in the resulting code. We present Lean-STaR, a framework for training language models to produce informal thoughts prior to each step of a proof, thereby boosting the model's theorem-proving capabilities. Lean-STaR uses retrospective ground-truth tactics to generate synthetic thoughts for training the language model. At inference time, the trained model directly generates the thoughts prior to the prediction of the tactics in each proof step. Building on the self-taught reasoner framework, we then apply expert iteration to further fine-tune the model on the correct proofs it samples and verifies using the Lean solver. Lean-STaR significantly outperform base models (43.4% → 46.3%, Pass@64). We also analyze the impact of the augmented thoughts on various aspects of the theorem proving process, providing insights into their effectiveness.

Large language models (LLMs), optimized through human feedback, have rapidly emerged as a leading paradigm for developing intelligent conversational assistants. However, despite their strong performance across many benchmarks, LLM-based agents might still lack conversational skills such as disambiguation -- when they are faced with ambiguity, they often overhedge or implicitly guess users' true intents rather than asking clarification questions. Under task-specific settings, high-quality conversation samples are often limited, constituting a bottleneck for LLMs' ability to learn optimal dialogue action policies. We propose Action-Based Contrastive Self-Training (ACT), a quasi-online preference optimization algorithm based on Direct Preference Optimization (DPO), that enables data-efficient dialogue policy learning in multi-turn conversation modeling. We demonstrate ACT's efficacy under data-efficient tuning scenarios, even when there is no action label available, using multiple real-world conversational tasks: tabular-grounded question-answering, machine reading comprehension, and AmbigSQL, a novel task for disambiguating information-seeking requests for complex SQL generation towards data analysis agents. Additionally, we propose evaluating LLMs' ability to function as conversational agents by examining whether they can implicitly recognize and reason about ambiguity in conversation. ACT demonstrates substantial conversation modeling improvements over standard tuning approaches like supervised fine-tuning and DPO.

Developing agents that can follow multimodal instructions remains a fundamental challenge in robotics and AI. Although large-scale pre-training on unlabeled datasets has enabled agents to learn diverse behaviors, these agents often struggle with following instructions. While augmenting the dataset with instruction labels can mitigate this issue, acquiring such high-quality annotations at scale is impractical. To address this issue, we frame the problem as a semi-supervised learning task and introduce \agent, a multimodal instructable agent trained using a novel approach that combines weak supervision with latent variable models. Our method consists of two key components: constrained self-imitating, which utilizes large amounts of unlabeled demonstrations to enable the policy to learn diverse behaviors, and human intention alignment, which uses a smaller set of labeled demonstrations to ensure the latent space reflects human intentions. \agent’s effectiveness is validated across four diverse environments, ranging from video games to robotic manipulation, demonstrating its robust multimodal instruction-following capabilities.

The smoothness of control actions is a significant challenge faced by deep reinforcement learning (RL) techniques in solving optimal control problems. Existing RL-trained policies tend to produce non-smooth actions due to high-frequency input noise and unconstrained Lipschitz constants in neural networks. This article presents a Smooth ODE (SmODE) network capable of simultaneously addressing both causes of unsmooth control actions, thereby enhancing policy performance and robustness under noise condition. We first design a smooth ODE neuron with first-order low-pass filtering expression, which can dynamically filter out high frequency noises of hidden state by a learnable state-based system time constant. Additionally, we construct a state-based mapping function, $g$, and theoretically demonstrate its capacity to control the ODE neuron's Lipschitz constant. Then, based on the above neuronal structure design, we further advanced the SmODE network serving as RL policy approximators. This network is compatible with most existing RL algorithms, offering improved adaptability compared to prior approaches. Various experiments show that our SmODE network demonstrates superior anti-interference capabilities and smoother action outputs than the multi-layer perception and smooth network architectures like LipsNet.

Multi-Objective Markov Decision Processes (MO-MDPs) are receiving increasing attention, as real-world decision-making problems often involve conflicting objectives that cannot be addressed by a single-objective MDP. The Pareto front identifies the set of policies that cannot be dominated, providing a foundation for finding Pareto optimal solutions that can efficiently adapt to various preferences. However, finding the Pareto front is a highly challenging problem. Most existing methods either (i) rely on traversing the *continuous preference space*, which is impractical and results in approximations that are difficult to evaluate against the true Pareto front, or (ii) focus solely on deterministic Pareto optimal policies, from which there are no known techniques to characterize the full Pareto front. Moreover, finding the structure of the Pareto front itself remains unclear even in the context of dynamic programming, where the MDP is fully known in advance. In this work, we address the challenge of efficiently discovering the Pareto front, involving both deterministic and stochastic Pareto optimal policies. By investigating the geometric structure of the Pareto front in MO-MDPs, we uncover a key property: the Pareto front is on the boundary of a convex polytope whose vertices all correspond to deterministic policies, and neighboring vertices of the Pareto front differ by only one state-action pair of the deterministic policy, almost surely. This insight transforms the global comparison across all policies into a localized search among deterministic policies that differ by only one state-action pair, drastically reducing the complexity of searching for the exact Pareto front. We develop an efficient algorithm that identifies the vertices of the Pareto front by solving a single-objective MDP only once and then traversing the edges of the Pareto front, making it more efficient than existing methods. Furthermore, the entire Pareto front can be found in $V$ iterations, where $V$ represents the number of vertices on the Pareto front. Our empirical studies demonstrate the effectiveness of our theoretical strategy in discovering the Pareto front efficiently.

Structure-based drug discovery, encompassing the tasks of protein-ligand docking and pocket-aware 3D drug design, represents a core challenge in drug discovery. However, no existing work can deal with both tasks to effectively leverage the duality between them, and current methods for each task are hindered by challenges in modeling 3D information and the limitations of available data. To address these issues, we propose 3DMolFormer, a unified dual-channel transformer-based framework applicable to both docking and 3D drug design tasks, which exploits their duality by utilizing docking functionalities within the drug design process. Specifically, we represent 3D pocket-ligand complexes using parallel sequences of discrete tokens and continuous numbers, and we design a corresponding dual-channel transformer model to handle this format, thereby overcoming the challenges of 3D information modeling. Additionally, we alleviate data limitations through large-scale pre-training on a mixed dataset, followed by supervised and reinforcement learning fine-tuning techniques respectively tailored for the two tasks. Experimental results demonstrate that 3DMolFormer outperforms previous approaches in both protein-ligand docking and pocket-aware 3D drug design, highlighting its promising application in structure-based drug discovery. The code is available at: https://github.com/HXYfighter/3DMolFormer.

How to balance between exploration and exploitation in an uncertain environment is a central challenge in reinforcement learning. In contrast, humans and animals have demonstrated superior exploration efficiency in novel environments. To understand how the brain’s neural network controls exploration under uncertainty, we analyzed the dynamical systems model of a biological neural network that controls explore-exploit decisions during foraging. Mathematically, this model (named the Brain Bandit Net, or BBN) is a special type of stochastic continuous Hopfield network. We show through theory and simulation that BBN can perform posterior sampling of action values with a tunable bias towards or against uncertain options. We then demonstrate that, in multi-armed bandit (MAB) tasks, BBN can generate probabilistic choice behavior with a flexible uncertainty bias resembling human and animal choice patterns. In addition to its high efficiency in MAB tasks, BBN can also be embedded with reinforcement learning algorithms to accelerate learning in MDP tasks. Altogether, our findings reveal the theoretical foundation for efficient exploration in biological neural networks and propose a general, brain-inspired algorithm for enhancing exploration in RL.

Designing reward functions in Reinforcement Learning (RL) often demands significant task-specific expertise. Offline Preference-based Reinforcement Learning (PbRL) provides an effective alternative to address the complexity of reward design by learning policies from offline datasets that contain human preferences between trajectory pairs. Existing offline PbRL studies typically model a reward function by maximizing its likelihood of generating the observed human preferences. However, due to the varying number of samples within the limited dataset, less frequently compared trajectories exhibit greater uncertainty, which potentially leads to unreliable behaviors during reward and policy updates. To solve this issue, in this work, we introduce Uncertainty-Aware PbRL (UA-PbRL) to learn a distributional reward model and a risk-sensitive policy from an offline preference dataset. Our approach employs a Maximum A Posteriori (MAP) objective to update trajectory rewards and incorporates an informative prior to account for the uncertainties. Building upon this reward update, we propose a generative reward model to capture the reward distribution, utilizing the offline distributional Bellman operator and the Conditional Value-at-Risk (CVaR) metric to train a risk-sensitive policy. Experimental results demonstrate that UA-PbRL effectively identifies and avoids states with high uncertainty, facilitating risk-averse behaviors across various tasks, including robot control and language model alignment. The code is available at https://github.com/Jasonxu1225/UA-PbRL.

Enabling robots to learn novel tasks in a data-efficient manner is a long-standing challenge. Common strategies involve carefully leveraging prior experiences, especially transition data collected on related tasks. Although much progress has been made for general pick-and-place manipulation, far fewer studies have investigated contact-rich assembly tasks, where precise control is essential. We introduce SRSA} (Skill Retrieval and Skill Adaptation), a novel framework designed to address this problem by utilizing a pre-existing skill library containing policies for diverse assembly tasks. The challenge lies in identifying which skill from the library is most relevant for fine-tuning on a new task. Our key hypothesis is that skills showing higher zero-shot success rates on a new task are better suited for rapid and effective fine-tuning on that task. To this end, we propose to predict the transfer success for all skills in the skill library on a novel task, and then use this prediction to guide the skill retrieval process. We establish a framework that jointly captures features of object geometry, physical dynamics, and expert actions to represent the tasks, allowing us to efficiently learn the transfer success predictor. Extensive experiments demonstrate that SRSA significantly outperforms the leading baseline. When retrieving and fine-tuning skills on unseen tasks, SRSA achieves a 19% relative improvement in success rate, exhibits 2.6x lower standard deviation across random seeds, and requires 2.4x fewer transition samples to reach a satisfactory success rate, compared to the baseline. In a continual learning setup, SRSA efficiently learns policies for new tasks and incorporates them into the skill library, enhancing future policy learning. Furthermore, policies trained with SRSA in simulation achieve a 90% mean success rate when deployed in the real world. Please visit our project webpage https://srsa2024.github.io/.

Reinforcement learning (RL) algorithms aim to balance exploiting the current best strategy with exploring new options that could lead to higher rewards. Most common RL algorithms use undirected exploration, i.e., select random sequences of actions. Exploration can also be directed using intrinsic rewards, such as curiosity or model epistemic uncertainty. However, effectively balancing task and intrinsic rewards is challenging and often task-dependent. In this work, we introduce a framework, MaxInfoRL, for balancing intrinsic and extrinsic exploration. MaxInfoRL steers exploration towards informative transitions, by maximizing intrinsic rewards such as the information gain about the underlying task. When combined with Boltzmann exploration, this approach naturally trades off maximization of the value function with that of the entropy over states, rewards, and actions. We show that our approach achieves sublinear regret in the simplified setting of multi-armed bandits. We then apply this general formulation to a variety of off-policy model-free RL methods for continuous state-action spaces, yielding novel algorithms that achieve superior performance across hard exploration problems and complex scenarios such as visual control tasks.

We consider the problem of how a trusted, but computationally bounded agent (a 'verifier') can learn to interact with one or more powerful but untrusted agents ('provers') in order to solve a given task. More specifically, we study the case in which agents are represented using neural networks and refer to solutions of this problem as neural interactive proofs. First we introduce a unifying framework based on prover-verifier games (Anil et al., 2021), which generalises previously proposed interaction protocols. We then describe several new protocols for generating neural interactive proofs, and provide a theoretical comparison of both new and existing approaches. Finally, we support this theory with experiments in two domains: a toy graph isomorphism problem that illustrates the key ideas, and a code validation task using large language models. In so doing, we aim to create a foundation for future work on neural interactive proofs and their application in building safer AI systems.

Reinforcement learning (RL) promises a framework for near-universal problem-solving. In practice however, RL algorithms are often tailored to specific benchmarks, relying on carefully tuned hyperparameters and algorithmic choices. Recently, powerful model-based RL methods have shown impressive general results across benchmarks but come at the cost of increased complexity and slow run times, limiting their broader applicability. In this paper, we attempt to find a unifying model-free deep RL algorithm that can address a diverse class of domains and problem settings. To achieve this, we leverage model-based representations that approximately linearize the value function, taking advantage of the denser task objectives used by model-based RL while avoiding the costs associated with planning or simulated trajectories. We evaluate our algorithm, MR.Q, on a variety of common RL benchmarks with a single set of hyperparameters and show a competitive performance against domain-specific and general baselines, providing a concrete step towards building general-purpose model-free deep RL algorithms.

Offline reinforcement learning (RL) has gained considerable attention for its ability to learn policies from pre-collected data without real-time interaction, which makes it particularly useful for high-risk applications. However, due to its reliance on offline datasets, existing works inevitably introduce assumptions to ensure effective learning, which, however, often lead to a trade-off between robustness to model mismatch and scalability to large environments. In this paper, we enhance both aspects with a novel double-pessimism principle, which conservatively estimates performance and accounts for both limited data and potential model mismatches, two major reasons for the previous trade-off. We then propose a universal, model-free algorithm to learn a policy that is robust to potential environment mismatches, which enhances robustness in a scalable manner. Furthermore, we provide a sample complexity analysis of our algorithm when the mismatch is modeled by the $l_\alpha$-norm, which also theoretically demonstrates the efficiency of our method. Extensive experiments further demonstrate that our approach significantly improves robustness in a more scalable manner than existing methods.

We investigate off-policy evaluation (OPE), a central and fundamental problem in reinforcement learning (RL), in the challenging setting of Partially Observable Markov Decision Processes (POMDPs) with large observation spaces. Recent works of Uehara et al. (2023a); Zhang & Jiang (2024) developed a model-free framework and identified important coverage assumptions (called belief and outcome coverage) that enable accurate OPE of memoryless policies with polynomial sample complexities, but handling more general target policies that depend on the entire observable history remained an open problem. In this work, we prove information-theoretic hardness for model-free OPE of history-dependent policies in several settings, characterized by additional assumptions imposed on the behavior policy (memoryless vs. history-dependent) and/or the state-revealing property of the POMDP (single-step vs. multi-step revealing). We further show that some hardness can be circumvented by a natural model-based algorithm—whose analysis has surprisingly eluded the literature despite the algorithm’s simplicity—demonstrating provable separation between model-free and model-based OPE in POMDPs.