ICML 2025 Papers

Existing reinforcement learning (RL) methods for large language models (LLMs) rely on static prompt sets, where prompts are curated a priori, and sampled in a fixed schedule for training, regardless of their usefulness to the RL process. We design `eva`, the first method that allows LLMs to prioritize and adaptively create useful prompts during RL training by reward signals. In principle, `eva` (Evolving via A symmetric Self-Play) casts language model training as a game between: (1) a creator, who samples and generates training prompts, and (2) a solver, who generates responses to the prompts. `eva` is simple, suits both offline and online RL for LLMs, and sets a new state-of-the-art on challenging benchmarks without extra human prompts: it improves gemma-2-9b-it’s win-rate on Arena-Hard from 51.6% to 60.1% by DPO and 52.6% to 62.4% by RLOO, surpassing claude-3-opus and nearing gemini-1.5-pro, both are orders of magnitude larger. Further ablation studies show `eva` can induce meaningful learning curriculum, and effectively scale RL for LLMs beyond static human prompts.

Distributional reinforcement learning improves performance by capturing environmental stochasticity, but a comprehensive theoretical understanding of its effectiveness remains elusive. In addition, the intractable element of the infinite dimensionality of distributions has been overlooked. In this paper, we present a regret analysis of distributional reinforcement learning with general value function approximation in a finite episodic Markov decision process setting. We first introduce a key notion of Bellman unbiasedness which is essential for exactly learnable and provably efficient distributional updates in an online manner. Among all types of statistical functionals for representing infinite-dimensional return distributions, our theoretical results demonstrate that only moment functionals can exactly capture the statistical information. Secondly, we propose a provably efficient algorithm, SF-LSVI, that achieves a tight regret bound of $\tilde{O}(d_E H^{\frac{3}{2}}\sqrt{K})$ where $H$ is the horizon, $K$ is the number of episodes, and $d_E$ is the eluder dimension of a function class.

As multi-agent reinforcement learning (MARL) progresses towards solving larger and more complex problems, it becomes increasingly important that algorithms exhibit the key properties of (1) strong performance, (2) memory efficiency, and (3) scalability. In this work, we introduce Sable, a performant, memory-efficient, and scalable sequence modelling approach to MARL. Sable works by adapting the retention mechanism in Retentive Networks (Sun et al., 2023) to achieve computationally efficient processing of multi-agent observations with long context memory for temporal reasoning. Through extensive evaluations across six diverse environments, we demonstrate how **Sable is able to significantly outperform existing state-of-the-art methods in a large number of diverse tasks (34 out of 45 tested)**. Furthermore, Sable maintains performance as we scale the number of agents, handling environments with more than a thousand agents while exhibiting a linear increase in memory usage. Finally, we conduct ablation studies to isolate the source of Sable's performance gains and confirm its efficient computational memory usage. **All experimental data, hyperparameters, and code for a frozen version of Sable used in this paper are available on our website:** https://sites.google.com/view/sable-marl. **An improved and maintained version of Sable is available in Mava:** https://github.com/instadeepai/Mava.

Long-horizon tasks in robotic manipulation present significant challenges in reinforcement learning (RL) due to the difficulty of designing dense reward functions and effectively exploring the expansive state-action space. However, despite a lack of dense rewards, these tasks often have a multi-stage structure, which can be leveraged to decompose the overall objective into manageable sub-goals. In this work, we propose DEMO³, a framework that exploits this structure for efficient learning from visual inputs. Specifically, our approach incorporates multi-stage dense reward learning, a bi-phasic training scheme, and world model learning into a carefully designed demonstration-augmented RL framework that strongly mitigates the challenge of exploration in long-horizon tasks. Our evaluations demonstrate that our method improves data-efficiency by an average of 40% and by 70% on particularly difficult tasks compared to state-of-the-art approaches. We validate this across 16 sparse-reward tasks spanning four domains, including challenging humanoid visual control tasks using as few as five demonstrations.

This paper studies off-policy evaluation (OPE) in reinforcement learning with a focus on behavior policy estimation for importance sampling. Prior work has shown empirically that estimating a history-dependent behavior policy can lead to lower mean squared error (MSE) even when the true behavior policy is Markovian. However, the question of *why* the use of history should lower MSE remains open. In this paper, we theoretically demystify this paradox by deriving a bias-variance decomposition of the MSE of ordinary importance sampling (IS) estimators, demonstrating that history-dependent behavior policy estimation decreases their asymptotic variances while increasing their finite-sample biases. Additionally, as the estimated behavior policy conditions on a longer history, we show a consistent decrease in variance. We extend these findings to a range of other OPE estimators, including the sequential IS estimator, the doubly robust estimator and the marginalized IS estimator, with the behavior policy estimated either parametrically or non-parametrically.

Bias problems in the estimation of Q-values are a well-known obstacle that slows down convergence of Q-learning and actor-critic methods. One of the reasons of the success of modern RL algorithms is partially a direct or indirect overestimation reduction mechanism. We introduce an easy to implement method built on top of distributional reinforcement learning (DRL) algorithms to deal with the overestimation in a locally adaptive way. Our framework ADDQ is simple to implement, existing DRL implementations can be improved with a few lines of code. We provide theoretical backup and experimental results in tabular, Atari, and MuJoCo environments, comparisons with state-of-the-art methods, and a proof of convergence in the tabular case.

The Lottery Ticket Hypothesis (LTH) suggests there exists a sparse LTH mask and weights that achieve the same generalization performance as the dense model while using significantly fewer parameters. However, finding a LTH solution is computationally expensive, and a LTH sparsity mask does not generalize to other random weight initializations. Recent work has suggested that neural networks trained from random initialization find solutions within the same basin modulo permutation, and proposes a method to align trained models within the same loss basin. We hypothesize that misalignment of basins is the reason why LTH masks do not generalize to new random initializations and propose permuting the LTH mask to align with the new optimization basin when performing sparse training from a different random init. We empirically show a significant increase in generalization when sparse training from random initialization with the permuted mask as compared to using the non-permuted LTH mask, on multiple datasets (CIFAR-10/100 & ImageNet) and models (VGG11 & ResNet20/50).

When rodents learn to navigate in a novel environment, a high density of place fields emerges at reward locations, fields elongate against the trajectory, and individual fields change spatial selectivity while demonstrating stable behavior. Why place fields demonstrate these characteristic phenomena during learning remains elusive. We develop a normative framework using a reward maximization objective, whereby the temporal difference (TD) error drives place field reorganization to improve policy learning. Place fields are modeled using Gaussian radial basis functions to represent states in an environment, and directly synapse to an actor-critic for policy learning. Each field's amplitude, center, and width, as well as downstream weights, are updated online at each time step to maximize rewards. We demonstrate that this framework unifies three disparate phenomena observed in navigation experiments. Furthermore, we show that these place field phenomena improve policy convergence when learning to navigate to a single target and relearning multiple new targets. To conclude, we develop a simple normative model that recapitulates several aspects of hippocampal place field learning dynamics and unifies mechanisms to offer testable predictions for future experiments.

Online linear programming (OLP) has found broad applications in revenue management and resource allocation. State-of-the-art OLP algorithms achieve low regret by repeatedly solving linear programming (LP) subproblems that incorporate updated resource information. However, LP-based methods are computationally expensive and often inefficient for large-scale applications. By contrast, recent first-order OLP algorithms are more computationally efficient but typically suffer from weaker regret guarantees. To address these shortcomings, we propose a new algorithm that combines the strengths of LP-based and first-order OLP algorithms. Our algorithm re-solves the LP subproblems periodically at a predefined frequency $f$ and uses the latest dual prices to guide online decision-making. In parallel, a first-order method runs during each interval between LP re-solves and smooths resource consumption. Our algorithm achieves $\mathcal{O}(\log (T/f) + \sqrt{f})$ regret and delivers a "wait-less" online decision-making process that balances computational efficiency and regret guarantees. Extensive experiments demonstrate at least $10$-fold improvements in regret over first-order methods and $100$-fold improvements in runtime over LP-based methods.

This paper investigates the potential of quantum acceleration in addressing infinite horizon Markov Decision Processes (MDPs) to enhance average reward outcomes. We introduce an innovative quantum framework for the agent's engagement with an unknown MDP, extending the conventional interaction paradigm. Our approach involves the design of an optimism-driven tabular Reinforcement Learning algorithm that harnesses quantum signals acquired by the agent through efficient quantum mean estimation techniques. Through thorough theoretical analysis, we demonstrate that the quantum advantage in mean estimation leads to exponential advancements in regret guarantees for infinite horizon Reinforcement Learning. Specifically, the proposed Quantum algorithm achieves a regret bound of $\tilde{\mathcal{O}}(1)$\footnote{$\tilde{\mathcal{O}}(\cdot)$ conceals logarithmic terms of $T$.}, a significant improvement over the $\tilde{\mathcal{O}}(\sqrt{T})$ bound exhibited by classical counterparts, where $T$ is the length of the time horizon.

The alignment of large language models (LLMs) with human preferences remains a key challenge. While post-training techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have achieved notable success, they often experience computational inefficiencies and training instability. In this paper, we propose \textbf{F}eature-level constrained \textbf{P}reference \textbf{O}ptimization (FPO), a novel method designed to simplify the alignment process while ensuring stability. FPO leverages pre-trained Sparse Autoencoders (SAEs) and introduces feature-level constraints, allowing for efficient, sparsity-enforced alignment. Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence by using the feature-level offline reference. Experimental results on benchmark datasets demonstrate that FPO achieves an above 5\% absolute improvement in win rate with much lower computational cost compared to state-of-the-art baselines, making it a promising solution for efficient and controllable LLM alignments.

Robust reinforcement learning (RL) under the average reward criterion, which seeks to optimize long-term system performance in uncertain environments, remains a largely unexplored area. To address this challenge, we propose a reduction-based framework that transforms robust average reward optimization into the more extensively studied robust discounted reward optimization by employing a specific discount factor. Our framework provides two key advantages. **Data Efficiency**: We design a model-based reduction algorithm that achieves near-optimal sample complexity, enabling efficient identification of optimal robust policies; **Scalability**: By bypassing the inherent challenges of scaling up average reward optimization, our framework facilitates the design of scalable, convergent algorithms for robust average reward optimization leveraging function approximation. Our algorithmic design, supported by theoretical and empirical analyses, provides a concrete solution to robust average reward RL with the first data efficiency and scalability guarantees, highlighting the framework’s potential to optimize long-term performance under model uncertainty in practical problems.

Iterative extension of empirical game models through deep reinforcement learning (RL) has proved an effective approach for finding equilibria in complex games. When multiple equilibria exist, we may also be interested in finding solutions with particular characteristics. We address this issue of equilibrium selection in the context of Policy Space Response Oracles (PSRO), a flexible game-solving framework based on deep RL, by skewing the strategy exploration process towards higher-welfare solutions. At each iteration, we create an exploration policy that imitates high welfare-yielding behavior and train a response to the current solution, regularized to be similar to the exploration policy. With no additional simulation expense, our approach, named Ex$^2$PSRO, tends to find higher welfare equilibria than vanilla PSRO in two benchmarks: a sequential bargaining game and a social dilemma game. Further experiments demonstrate Ex$^2$PSRO's composability with other PSRO variants and illuminate the relationship between exploration policy choice and algorithmic performance.

Maximum entropy reinforcement learning (MaxEnt-RL) has become the standard approach to RL due to its beneficial exploration properties. Traditionally, policies are parameterized using Gaussian distributions, which significantly limits their representational capacity. Diffusion-based policies offer a more expressive alternative, yet integrating them into MaxEnt-RL poses challenges—primarily due to the intractability of computing their marginal entropy. To overcome this, we propose Diffusion-Based Maximum Entropy RL (DIME). DIME leverages recent advances in approximate inference with diffusion models to derive a lower bound on the maximum entropy objective. Additionally, we propose a policy iteration scheme that provably converges to the optimal diffusion policy. Our method enables the use of expressive diffusion-based policies while retaining the principled exploration benefits of MaxEnt-RL, significantly outperforming other diffusion-based methods on challenging high-dimensional control benchmarks. It is also competitive with state-of-the-art non-diffusion based RL methods while requiring fewer algorithmic design choices and smaller update-to-data ratios, reducing computational complexity.

Language models exhibit complex, diverse behaviors when prompted with free-form text, making it hard to characterize the space of possible outputs. We study the problem of behavioral elicitation, where the goal is to search for prompts that induce specific target behaviors (e.g., hallucinations, harmful responses) from a target language model. To navigate the exponentially large space of possible prompts, we train amortized investigator models to emulate the posterior distribution over the prompts, conditioned on the target behavior. Specifically, we first fit a reverse model and then use reinforcement learning to optimize likelihood of generating the target behavior. To improve the diversity of the prompt distribution, we further propose a novel iterative training objective based on the Frank-Wolfe algorithm that encourages each iteration to discover different sets of prompts not captured by previous iterations. Our investigator models produce prompts that exhibit a variety of effective and human-interpretable strategies for behavior elicitation, obtaining a 100% attack success rate on AdvBench (Harmful Behaviors) and an 85% hallucination rate.

The number of agents can be an effective curriculum variable for controlling the difficulty of multi-agent reinforcement learning (MARL) tasks. Existing work typically uses manually defined curricula such as linear schemes. We identify two potential flaws while applying existing reward-based automatic curriculum learning methods in MARL: (1) The expected episode return used to measure task difficulty has high variance; (2) Credit assignment difficulty can be exacerbated in tasks where increasing the number of agents yields higher returns which is common in many MARL tasks. To address these issues, we propose to control the curriculum by using a TD-error based *learning progress* measure and by letting the curriculum proceed from an initial context distribution to the final task specific one. Since our approach maintains a distribution over the number of agents and measures learning progress rather than absolute performance, which often increases with the number of agents, we alleviate problem (2). Moreover, the learning progress measure naturally alleviates problem (1) by aggregating returns. In three challenging sparse-reward MARL benchmarks, our approach outperforms state-of-the-art baselines.

We propose Diff-Instruct*(DI*), a data-efficient post-training approach to one-step text-to-image generative models to improve its human preferences without requiring image data. Our method frames alignment as online reinforcement learning from human feedback (RLHF), which optimizes a human reward function while regularizing the generator to stay close to a reference diffusion process. Unlike traditional RLHF approaches, which rely on the KL divergence for regularization, we introduce a novel score-based divergence regularization that substantially improves performance. Although such a score-based RLHF objective seems intractable when optimizing, we derive a strictly equivalent tractable loss function in theory that can efficiently compute its gradient for optimizations. Building upon this framework, we train DI*-SDXL-1step, a 1-step text-to-image model based on Stable Diffusion-XL (2.6B parameters), capable of generating 1024x1024 resolution images in a single step. The 2.6B DI*-SDXL-1step model outperforms the 12B FLUX-dev model in ImageReward, PickScore, and CLIP score on the Parti prompts benchmark while using only 1.88% of the inference time. This result strongly supports the thought that with proper post-training, the small one-step model is capable of beating huge multi-step models. We will open-source our industry-ready model to the community.

Fine-tuning large language models (LLMs) to align with user preferences is challenging due to the high cost of quality human annotations in Reinforcement Learning from Human Feedback (RLHF) and the generalizability limitations of AI Feedback. To address these challenges, we propose RLTHF, a human-AI hybrid framework that combines LLM-based initial alignment with selective human annotations to achieve full-human annotation alignment with minimal effort. RLTHF identifies hard-to-annotate samples mislabeled by LLMs using a reward model's reward distribution and iteratively enhances alignment by integrating strategic human corrections while leveraging LLM's correctly labeled samples. Evaluations on HH-RLHF and TL;DR datasets show that RLTHF reaches full-human annotation-level alignment with only 6-7% of the human annotation effort. Furthermore, models trained on RLTHF's curated datasets for downstream tasks outperform those trained on fully human-annotated datasets, underscoring the effectiveness of RLTHF.

Many systems contain latent variables that make their dynamics partially unidentifiable or cause distribution shifts in the observed statistics between offline and online data. However, existing control techniques often assume access to complete dynamics or perfect simulators with fully observable states, which are necessary to verify whether the system remains within a safe set (forward invariance) or safe actions are consistently feasible at all times. To address this limitation, we propose a technique for designing probabilistic safety certificates for systems with latent variables. A key technical enabler is the formulation of invariance conditions in probability space, which can be constructed using observed statistics in the presence of distribution shifts due to latent variables. We use this invariance condition to construct a safety certificate that can be implemented efficiently in real-time control. The proposed safety certificate can continuously find feasible actions that control long-term risk to stay within tolerance. Stochastic safe control and (causal) reinforcement learning have been studied in isolation until now. To the best of our knowledge, the proposed work is the first to use causal reinforcement learning to quantify long-term risk for the design of safety certificates. This integration enables safety certificates to efficiently ensure long-term safety in the presence of latent variables. The effectiveness of the proposed safety certificate is demonstrated in numerical simulations.