ICML 2025 论文列表

Effective visual representation learning is crucial for reinforcement learning (RL) agents to extract task-relevant information from raw sensory inputs and generalize across diverse environments. However, existing RL benchmarks lack the ability to systematically evaluate representation learning capabilities in isolation from other learning challenges. To address this gap, we introduce the Sliding Puzzles Gym (SPGym), a novel benchmark that transforms the classic 8-tile puzzle into a visual RL task with images drawn from arbitrarily large datasets. SPGym's key innovation lies in its ability to precisely control representation learning complexity through adjustable grid sizes and image pools, while maintaining fixed environment dynamics, observation, and action spaces. This design enables researchers to isolate and scale the visual representation challenge independently of other learning components. Through extensive experiments with model-free and model-based RL algorithms, we uncover fundamental limitations in current methods' ability to handle visual diversity. As we increase the pool of possible images, all algorithms exhibit in- and out-of-distribution performance degradation, with sophisticated representation learning techniques often underperforming simpler approaches like data augmentation. These findings highlight critical gaps in visual representation learning for RL and establish SPGym as a valuable tool for driving progress in robust, generalizable decision-making systems.

A key challenge in training Large Language Models (LLMs) is properly aligning them with human preferences. Reinforcement Learning with Human Feedback (RLHF) uses pairwise comparisons from human annotators to train reward functions and has emerged as a popular alignment method. However, input datasets in RLHF can be unbalanced due to adversarial manipulation or inadvertent repetition. Therefore, we want RLHF algorithms to perform well even when the set of alternatives is not uniformly distributed. Drawing on insights from social choice theory, we introduce robustness to approximate clones, a desirable property of RLHF algorithms which requires that adding near-duplicate alternatives does not significantly change the learned reward function. We first demonstrate that the standard RLHF algorithm based on regularized maximum likelihood estimation (MLE) fails to satisfy this property. We then propose the weighted MLE, a new RLHF algorithm that modifies the standard regularized MLE by weighting alternatives based on their similarity to other alternatives. This new algorithm guarantees robustness to approximate clones while preserving desirable theoretical properties.

Long-context models (LCMs) have shown great potential in processing long input sequences (even more than 100M tokens) conveniently and effectively. With significant progress, recent research has pointed out that LCMs can accurately locate token-level salient information within the context. Yet, the generation performance of these LCMs is far from satisfactory and might result in misaligned responses, such as hallucinations. To enhance the generation capability of LCMs, existing works have investigated the effects of data size and quality for both pre-training and instruction tuning. Though achieving meaningful improvement, previous methods fall short in either effectiveness or efficiency. In this paper, we introduce LOGO (Long cOntext aliGnment via efficient preference Optimization), a training strategy that first introduces preference optimization for long-context alignment. To overcome the GPU memory-bound issue caused by the long sequence, LOGO employs a reference-free preference optimization strategy and adopts a position synthesis method to construct the training data. By training with only 0.3B data on a single 8 x A800 GPU machine for 16 hours, LOGO allows the Llama-3-8B-Instruct-80K model to achieve comparable performance with GPT-4 in real-world long-context tasks while preserving the model's original capabilities on other tasks, e.g., language modeling and MMLU. Moreover, LOGO can extend the model's context window size while enhancing its generation performance.

Generalizing multi-agent reinforcement learning (MARL) to accommodate variations in problem configurations remains a critical challenge in real-world applications, where even subtle differences in task setups can cause pre-trained policies to fail. To address this, we propose Context-Aware Identity Generation (CAID), a novel framework to enhance MARL performance under the Contextual MARL (CMARL) setting. CAID dynamically generates unique agent identities through the agent identity decoder built on a causal Transformer architecture. These identities provide contextualized representations that align corresponding agents across similar problem variants, facilitating policy reuse and improving sample efficiency. Furthermore, the action regulator in CAID incorporates these agent identities into the action-value space, enabling seamless adaptation to varying contexts. Extensive experiments on CMARL benchmarks demonstrate that CAID significantly outperforms existing approaches by enhancing both sample efficiency and generalization across diverse context variants.

Sampling from binary quadratic distributions (BQDs) is a fundamental but challenging problem in discrete optimization and probabilistic inference. Previous work established theoretical guarantees for stochastic localization (SL) in continuous domains, where MCMC methods efficiently estimate the required posterior expectations during SL iterations. However, achieving similar convergence guarantees for discrete MCMC samplers in posterior estimation presents unique theoretical challenges. In this work, we present the first application of SL to general BQDs, proving that after a certain number of iterations, the external field of posterior distributions constructed by SL tends to infinity almost everywhere, hence satisfy Poincaré inequalities with probability near to 1, leading to polynomial-time mixing. This theoretical breakthrough enables efficient sampling from general BQDs, even those that may not originally possess fast mixing properties. Furthermore, our analysis, covering enormous discrete MCMC samplers based on Glauber dynamics and Metropolis-Hastings algorithms, demonstrates the broad applicability of our theoretical framework. Experiments on instances with quadratic unconstrained binary objectives, including maximum independent set, maximum cut, and maximum clique problems, demonstrate consistent improvements in sampling efficiency across different discrete MCMC samplers.

Preference-based reinforcement learning (PbRL) bypasses explicit reward engineering by inferring reward functions from human preference comparisons, enabling better alignment with human intentions. However, humans often struggle to label a clear preference between similar segments, reducing label efficiency and limiting PbRL’s real-world applicability. To address this, we propose an offline PbRL method: Contrastive LeArning for ResolvIng Ambiguous Feedback (CLARIFY), which learns a trajectory embedding space that incorporates preference information, ensuring clearly distinguished segments are spaced apart, thus facilitating the selection of more unambiguous queries. Extensive experiments demonstrate that CLARIFY outperforms baselines in both non-ideal teachers and real human feedback settings. Our approach not only selects more distinguished queries but also learns meaningful trajectory embeddings.

Unlike its vanilla counterpart with i.i.d. samples, stochastic optimization with Markovian sampling allows the sampling scheme following a Markov chain. This problem encompasses various applications that range from asynchronous distributed optimization to reinforcement learning. In this work, we lower bound the sample complexity of finding $\epsilon$-approximate critical solutions for any first-order methods when sampling is Markovian. We show that for samples drawn from stationary Markov processes with countable state space, any algorithm that accesses smooth, non-convex functions through queries to a stochastic gradient oracle, requires at least $\Omega(\epsilon^{-4})$ samples. Moreover, for finite Markov chains, we show a $\Omega(\epsilon^{-2})$ lower bound and propose a new algorithm, called MaC-SAGE, that is proven to (nearly) match our lower bound.

This paper proposes Constrained Exploitability Descent (CED), a model-free offline reinforcement learning (RL) algorithm for solving adversarial Markov games (MGs). CED combines the game-theoretical approach of Exploitability Descent (ED) with policy constraint methods from offline RL. While policy constraints can perturb the optimal pure-strategy solutions in single-agent scenarios, we find the side effect less detrimental in adversarial games, where the optimal policy can be a mixed-strategy Nash equilibrium. We theoretically prove that, under the uniform coverage assumption on the dataset, CED converges to a stationary point in deterministic two-player zero-sum Markov games. We further prove that the min-player policy at the stationary point follows the property of mixed-strategy Nash equilibrium in MGs. Compared to the model-based ED method that optimizes the max-player policy, our CED method no longer relies on a generalized gradient. Experiments in matrix games, a tree-form game, and an infinite-horizon soccer game verify that CED can find an equilibrium policy for the min-player as long as the offline dataset guarantees uniform coverage. Besides, CED achieves a significantly lower NashConv compared to an existing pessimism-based method and can gradually improve the behavior policy even under non-uniform data coverages. When combined with neural networks, CED also outperforms behavior cloning and offline self-play in a large-scale two-team robotic combat game.

Model immunization aims to pre-train models that are difficult to fine-tune on harmful tasks while retaining their utility on other non-harmful tasks. Though prior work has shown empirical evidence for immunizing text-to-image models, the key understanding of when immunization is possible and a precise definition of an immunized model remain unclear. In this work, we propose a framework, based on the condition number of a Hessian matrix, to analyze model immunization for linear models. Building on this framework, we design an algorithm with regularization terms to control the resulting condition numbers after pre-training. Empirical results on linear models and non-linear deep-nets demonstrate the effectiveness of the proposed algorithm on model immunization. The code is available at https://github.com/amberyzheng/model-immunization-cond-num.

Quantifying the elemental composition of a material is a general scientific challenge with broad relevance to environmental sustainability. Existing techniques for the measurement of atomic abundances generally require laboratory conditions and expensive equipment. As a result, they cannot be deployed *in situ* without significant capital investment, limiting their proliferation. Measurement techniques based on nuclear magnetic resonance (NMR) hold promise in this setting due to their applicability across the periodic table, their non-destructive manipulation of samples, and their amenability to *in silico* optimization. In this work, we learn policies to modulate NMR pulses for rapid atomic abundance quantification. Our approach involves three inter-operating agents which (1) rapidly align nuclear spins for measurement, (2) quickly force relaxation to equilibrium, and (3) toggle control between agents (1) and (2) to minimize overall measurement time. To demonstrate this technique, we consider a specific use case of low-magnetic-field carbon-13 quantification for low-cost, portable analysis of foodstuffs and soils. We find significant performance improvements relative to traditional NMR pulse sequencing, and discuss limitations on the applicability of this approach.

Large language models (LLMs) are making inroads into classical AI problems such as automated planning, yet key shortcomings continue to hamper their integration. Chain-of-Thought (CoT) struggles in complex multi-step reasoning, and Tree-of-Thoughts requires multiple queries that increase computational overhead. Recently, Algorithm-of-Thoughts (AoT) have shown promise using in-context examples, at the cost of significantly longer solutions compared to CoT. Aimed at bridging the solution length gap between CoT and AoT, this paper introduces AoT-O3, which combines supervised finetuning on AoT-style plans with a reinforcement learning (RL) framework designed to reduce solution length. The RL component uses a reward model that favors concise, valid solutions while maintaining planning accuracy. Empirical evaluations indicate that AoT-O3 shortens solution length by up to 80\% compared to baseline AoT while maintaining or surpassing prior performance. These findings suggest a promising pathway for more efficient, scalable LLM-based planning.

We give an efficient algorithm for producing multi-dimensional forecasts in an online adversarial environment that have low bias subject to any polynomial number of conditioning events, that can depend both on external context and on our predictions themselves. We demonstrate the use of this algorithm with several applications. We show how to make predictions that can be transparently consumed by any polynomial number of downstream decision makers with different utility functions, guaranteeing them diminishing swap regret at optimal rates. We also give the first efficient algorithms for guaranteeing diminishing conditional regret in online combinatorial optimization problems for an arbitrary polynomial number of conditioning events --- i.e. on an arbitrary number of intersecting subsequences determined both by context and our own predictions. Finally, we give the first efficient algorithm for online multicalibration with $O(T^{2/3})$ rates in the ECE metric.

Offline reinforcement learning (RL) enables policy optimization using static datasets, avoiding the risks and costs of extensive real-world exploration. However, it struggles with suboptimal offline behaviors and inaccurate value estimation due to the lack of environmental interaction. We present Video-Enhanced Offline RL (VeoRL), a model-based method that constructs an interactive world model from diverse, unlabeled video data readily available online. Leveraging model-based behavior guidance, our approach transfers commonsense knowledge of control policy and physical dynamics from natural videos to the RL agent within the target domain. VeoRL achieves substantial performance gains (over 100% in some cases) across visual control tasks in robotic manipulation, autonomous driving, and open-world video games. Project page: https://panmt.github.io/VeoRL.github.io.

In this work, we design quantum algorithms that are more efficient than classical algorithms to solve time-dependent and finite-horizon Markov Decision Processes (MDPs) in two distinct settings: (1) In the exact dynamics setting, where the agent has full knowledge of the environment's dynamics (i.e., transition probabilities), we prove that our **Quantum Value Iteration (QVI)** algorithm **QVI-1** achieves a quadratic speedup in the size of the action space $(A)$ compared with the classical value iteration algorithm for computing the optimal policy ($\pi^{\ast}$) and the optimal V-value function ($V_{0}^{\ast}$). Furthermore, our algorithm **QVI-2** provides an additional speedup in the size of the state space $(S)$ when obtaining near-optimal policies and V-value functions. Both **QVI-1** and **QVI-2** achieve quantum query complexities that provably improve upon classical lower bounds, particularly in their dependences on $S$ and $A$. (2) In the generative model setting, where samples from the environment are accessible in quantum superposition, we prove that our algorithms **QVI-3** and **QVI-4** achieve improvements in sample complexity over the state-of-the-art (SOTA) classical algorithm in terms of $A$, estimation error $(\epsilon)$, and time horizon $(H)$. More importantly, we prove quantum lower bounds to show that **QVI-3** and **QVI-4** are asymptotically optimal, up to logarithmic factors, assuming a constant time horizon.

Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks. However, existing approaches mainly rely on imitation learning and struggle to achieve effective test-time scaling. While reinforcement learning (RL) holds promise for enabling self-exploration, recent attempts yield modest improvements in complex reasoning. In this paper, we present T1 to scale RL by encouraging exploration and understand inference scaling. We first initialize the LLM using synthesized chain-of-thought data that integrates trial-and-error and self-verification. To scale RL training, we promote increased sampling diversity through over-sampling. We demonstrate that T1 with open LLMs as its base exhibits inference scaling behavior and achieves superior performance on challenging math reasoning benchmarks. More importantly, we present a simple strategy to examine inference scaling, where increased inference budgets directly lead to T1’s better performance without any additional verification. The model weights and training data are publicly available at https://github.com/THUDM/T1.

Developing autonomous agents that quickly explore an environment and adapt their behavior online is a canonical challenge in robotics and machine learning. While humans are able to achieve such fast online exploration and adaptation, often acquiring new information and skills in only a handful of interactions, existing algorithmic approaches tend to rely on random exploration and slow, gradient-based behavior updates. How can we endow autonomous agents with such capabilities on par with humans? Taking inspiration from recent progress on both in-context learning and large-scale behavioral cloning, in this work we propose behavioral exploration: training agents to internalize what it means to explore and adapt in-context over the space of ''expert'' behaviors. To achieve this, given access to a dataset of expert demonstrations, we train a long-context generative model to predict expert actions conditioned on a context of past observations and a measure of how ''exploratory'' the expert's behaviors are relative to this context. This enables the model to not only mimic the behavior of an expert, but also, by feeding its past history of interactions into its context, to select different expert behaviors than what have been previously selected, thereby allowing for fast online adaptation and targeted, ''expert-like'' exploration. We demonstrate the effectiveness of our method in both simulated locomotion and manipulation settings, as well as on real-world robotic manipulation tasks, illustrating its ability to learn adaptive, exploratory behavior.

The ability of agents to collaborate with previously unknown teammates on the fly, known as ad hoc teamwork (AHT), is crucial in many real-world applications. Existing approaches to AHT require online interactions with the environment and some carefully designed teammates. However, these prerequisites can be infeasible in practice. In this work, we extend the AHT problem to the offline setting, where the policy of the ego agent is directly learned from a multi-agent interaction dataset. We propose a hierarchical sequence modeling framework called TAGET that addresses critical challenges in the offline setting, including limited data, partial observability and online adaptation. The core idea of TAGET is to dynamically predict teammate-aware rewards-to-go and sub-goals, so that the ego agent can adapt to the changes of teammates’ behaviors in real time. Extensive experimental results show that TAGET significantly outperforms existing solutions to AHT in the offline setting.

Reinforcement Learning from Human Feedback (RLHF) is widely used to align models with human preferences, particularly to enhance the safety of responses generated by LLMs. This method traditionally relies on choosing preferred responses from response pairs. However, due to variations in human opinions and the difficulty of making an overall comparison of two responses, there is a growing shift towards a fine-grained annotation approach, assessing responses based on multiple specific metrics or rules. Selecting and applying these rules efficiently while accommodating the diversity of preference data remains a significant challenge. In this paper, we introduce a dynamic approach that adaptively selects the most critical rules for each pair of responses. We develop a mathematical framework that leverages the maximum discrepancy between each paired responses and theoretically show that this strategy optimizes the mutual information between the rule-based labeling and the hidden ground-truth preferences. We then train an 8B reward model using the adaptively labeled preference dataset and evaluate its performance on RewardBench. As of May 25, 2025, our model achieved the highest safety performance on the leaderboard, outperforming various larger models.

Training reinforcement learning (RL) agents requires extensive trials and errors, which becomes prohibitively time-consuming in systems with costly reward evaluations. To address this challenge, we propose adaptive reward modeling (AdaReMo) which accelerates RL training by decomposing the complicated reward function into multiple localized fast reward models approximating direct reward evaluation with neural networks. These models dynamically adapt to the agent’s evolving policy by fitting the currently explored subspace with the latest trajectories, ensuring accurate reward estimation throughout the entire training process while significantly reducing computational overhead. We empirically show that AdaReMo not only achieves over 1,000 times speedup but also improves the performance by 14.6% over state-of-the-art approaches across three expensive-to-evaluate systems---molecular generation, epidemic control, and spatial planning. Code and data for the project are provided at https://github.com/tsinghua-fib-lab/AdaReMo.