ICML 2025 Papers

While foundation models have been exploited for various expert tasks with their fine-tuned parameters, any foundation model will be eventually outdated due to its old knowledge or limited capability, and thus should be replaced by a new foundation model. Subsequently, to benefit from its latest knowledge or improved capability, the new foundation model should be fine-tuned on each task again, which incurs not only the additional training cost but also the maintenance cost of the task-specific data. Existing work address this problem by inference-time tuning, i.e., modifying the output probability from the new foundation model by the outputs from the old foundation model and its fine-tuned model, which involves an additional inference cost by the latter two models. In this paper, we explore a new fine-tuning principle (which we call portable reward tuning; PRT) that reduces the inference cost by its nature, based on the reformulation of fine-tuning as the reward maximization with Kullback-Leibler regularization. Specifically, instead of fine-tuning parameters of the foundation models, PRT trains the reward model explicitly through the same loss as in fine-tuning. During inference, the reward model can be used with any foundation model (with the same set of vocabularies or labels) through the formulation of reward maximization. Experimental results, including both vision and language models, show that the PRT-trained model can achieve comparable accuracy with less inference cost, in comparison to the existing work of inference-time tuning.

We are born with the ability to learn concepts by comparing diverse observations. This helps us to understand the new world in a compositional manner and facilitates extrapolation, as objects naturally consist of multiple concepts. In this work, we argue that the cognitive mechanism of comparison, fundamental to human learning, is also vital for machines to recover true concepts underlying the data. This offers correctness guarantees for the field of concept learning, which, despite its impressive empirical successes, still lacks general theoretical support. Specifically, we aim to develop a theoretical framework for the identifiability of concepts with multiple classes of observations. We show that with sufficient diversity across classes, hidden concepts can be identified without assuming specific concept types, functional relations, or parametric generative models. Interestingly, even when conditions are not globally satisfied, we can still provide alternative guarantees for as many concepts as possible based on local comparisons, thereby extending the applicability of our theory to more flexible scenarios. Moreover, the hidden structure between classes and concepts can also be identified nonparametrically. We validate our theoretical results in both synthetic and real-world settings.

We study learning human preferences from limited comparison feedback, a core machine learning problem that is at the center of reinforcement learning from human feedback (RLHF). We formulate the problem as learning a Plackett-Luce (PL) model from a limited number of $K$-subset comparisons over a universe of $N$ items, where typically $K \ll N$. Our objective is to select the $K$-subsets such that all items can be ranked with minimal mistakes within the budget. We solve the problem using the D-optimal design, which minimizes the worst-case ranking loss under the estimated PL model. All known algorithms for this problem are computationally infeasible in our setting because we consider exponentially many subsets in $K$. To address this challenge, we propose a randomized Frank-Wolfe algorithm with memoization and sparse updates that has a low $O(N^2 + K^2)$ per-iteration complexity. We analyze it and demonstrate its empirical superiority on synthetic and open-source NLP datasets.

Marginalization -- summing a function over all assignments to a subset of its inputs -- is a fundamental computational problem with applications from probabilistic inference to formal verification. Despite its computational hardness in general, there exist many classes of functions (e.g., probabilistic models) for which marginalization remains tractable, and they can all be commonly expressed by arithmetic circuits computing multilinear polynomials. This raises the question, can *all* functions with polynomial time marginalization algorithms be succinctly expressed by such circuits? We give a negative answer, exhibiting simple functions with tractable marginalization yet no efficient representation by known models, assuming $\\mathsf{FP} \\neq \\#\\mathsf{P}$ (an assumption implied by $\\mathsf{P} \\neq \\mathsf{NP}$). To this end, we identify a hierarchy of complexity classes corresponding to stronger forms of marginalization, all of which are efficiently computable on the known circuit models. We conclude with a completeness result, showing that whenever there is an efficient real RAM performing virtual evidence marginalization for a function, then there are small arithmetic circuits for that function's multilinear representation.

In this study, we tackle an optimization problem with a known function and an unknown decision-dependent distribution, which arises in a variety of applications and is often referred to as a performative prediction problem. To solve the problem, several zeroth-order methods have been developed because the gradient of the objective function cannot be obtained explicitly due to the unknown distribution. Although these methods have theoretical convergence, they cannot utilize the information on the known function, which limits their efficiency in reducing the objective value. To overcome this issue, we propose new zeroth-order methods that generate effective update directions by utilizing information on the known function. As theoretical results, we show the convergence of our methods to stationary points and provide the worst-case sample complexity analysis, which improves the state of the arts when the maximum objective value dominates the cube root of the decision variable's dimensionality in order. Our simulation experiments on multiple applications show that our methods output solutions with lower objective values than the existing zeroth-order methods do.

Accurate and verifiable large language model (LLM) simulations of human research subjects promise an accessible data source for understanding human behavior and training new AI systems. However, results to date have been limited, and few social scientists have adopted this method. In this position paper, we argue that the promise of LLM social simulations can be achieved by addressing five tractable challenges. We ground our argument in a review of empirical comparisons between LLMs and human research subjects, commentaries on the topic, and related work. We identify promising directions, including context-rich prompting and fine-tuning with social science datasets. We believe that LLM social simulations can already be used for pilot and exploratory studies, and more widespread use may soon be possible with rapidly advancing LLM capabilities. Researchers should prioritize developing conceptual models and iterative evaluations to make the best use of new AI systems.

We present a novel generative approach based on Denoising Diffusion Models (DDMs), which produces high-quality image samples *along* with their losslessly compressed bit-stream representations. This is obtained by replacing the standard Gaussian noise sampling in the reverse diffusion with a selection of noise samples from pre-defined codebooks of fixed iid Gaussian vectors. Surprisingly, we find that our method, termed *Denoising Diffusion Codebook Model* (DDCM), retains sample quality and diversity of standard DDMs, even for extremely small codebooks. We leverage DDCM and pick the noises from the codebooks that best match a given image, converting our generative model into a highly effective lossy image codec achieving state-of-the-art perceptual image compression results. More generally, by setting other noise selections rules, we extend our compression method to any conditional image generation task (e.g., image restoration), where the generated images are produced jointly with their condensed bit-stream representations. Our work is accompanied by a mathematical interpretation of the proposed compressed conditional generation schemes, establishing a connection with score-based approximations of posterior samplers for the tasks considered. Code and demo are available on our project's [website](https://ddcm-2025.github.io/).

Divergence-free symmetric tensors (DFSTs) are fundamental in continuum mechanics, encoding conservation laws such as mass and momentum conservation. We introduce Riemann Tensor Neural Networks (RTNNs), a novel neural architecture that inherently satisfies the DFST condition to machine precision, providing a strong inductive bias for enforcing these conservation laws. We prove that RTNNs can approximate any sufficiently smooth DFST with arbitrary precision and demonstrate their effectiveness as surrogates for conservative PDEs, achieving improved accuracy across benchmarks. This work is the first to use DFSTs as an inductive bias in neural PDE surrogates and to explicitly enforce the conservation of both mass and momentum within a physics-constrained neural architecture.

Self-supervised contrastive learning has emerged as a powerful tool in machine learning and computer vision to learn meaningful representations from unlabeled data. Meanwhile, its empirical success has encouraged many theoretical studies to reveal the learning mechanisms. However, in the existing theoretical research, the role of data augmentation is still under-exploited, especially the effects of specific augmentation types. To fill in the blank, we for the first time propose an augmentation-aware error bound for self-supervised contrastive learning, showing that the supervised risk is bounded not only by the unsupervised risk, but also explicitly by a trade-off induced by data augmentation. Then, under a novel semantic label assumption, we discuss how certain augmentation methods affect the error bound. Lastly, we conduct both pixel- and representation-level experiments to verify our proposed theoretical results.

Large Language Models (LLMs) have gained significant popularity due to their remarkable capabilities in text understanding and generation. However, despite their widespread deployment in inference services such as ChatGPT, concerns about the potential leakage of sensitive user data have arisen. Existing solutions primarily rely on privacy-enhancing technologies to mitigate such risks, facing the trade-off among efficiency, privacy, and utility. To narrow this gap, we propose Cape, a context-aware prompt perturbation mechanism based on differential privacy, to enable efficient inference with an improved privacy-utility trade-off. Concretely, we introduce a hybrid utility function that better captures the token similarity. Additionally, we propose a bucketized sampling mechanism to handle large sampling space, which might lead to long-tail phenomenons. Extensive experiments across multiple datasets, along with ablation studies, demonstrate that Cape achieves a better privacy-utility trade-off compared to prior state-of-the-art works.

Multi-View Reinforcement Learning (MVRL) seeks to provide agents with multi-view observations, enabling them to perceive environment with greater effectiveness and precision. Recent advancements in MVRL focus on extracting latent representations from multiview observations and leveraging them in control tasks. However, it is not straightforward to learn compact and task-relevant representations, particularly in the presence of redundancy, distracting information, or missing views. In this paper, we propose **M**ulti-view **F**usion **S**tate for **C**ontrol (**MFSC**), firstly incorporating bisimulation metric learning into MVRL to learn task-relevant representations. Furthermore, we propose a multiview-based mask and latent reconstruction auxiliary task that exploits shared information across views and improves MFSC’s robustness in missing views by introducing a mask token. Extensive experimental results demonstrate that our method outperforms existing approaches in MVRL tasks. Even in more realistic scenarios with interference or missing views, MFSC consistently maintains high performance. The project code is available at [https://github.com/zpwdev/MFSC](https://github.com/zpwdev/MFSC).

Image steganography hides multiple images for multiple recipients into a single cover image. All secret images are usually revealed without authentication, which reduces security among multiple recipients. It is elegant to design an authentication mechanism for isolated reception. We explore such mechanism through sufficient experiments, and uncover that additional authentication information will affect the distribution of hidden information and occupy more hiding space of the cover image. This severely decreases effectiveness and efficiency in large-capacity hiding. To overcome such a challenge, we first prove the authentication feasibility within image steganography. Then, this paper proposes an image steganography network collaborating with separate authentication and efficient scheme. Specifically, multiple pairs of lock-key are generated during hiding and revealing. Unlike traditional methods, our method has two stages to make appropriate distribution adaptation between locks and secret images, simultaneously extracting more reasonable primary information from secret images, which can release hiding space of the cover image to some extent. Furthermore, due to separate authentication, fused information can be hidden in parallel with a single network rather than traditional serial hiding with multiple networks, which can largely decrease the model size. Extensive experiments demonstrate that the proposed method achieves more secure, effective, and efficient image steganography. Code is available at https://github.com/Revive624/Authentication-Image-Steganography.

We present MILS: Multimodal Iterative LLM Solver, a surprisingly simple, training-free approach, to imbue multimodal capabilities into your favorite LLM. Leveraging their innate ability to perform multi-step reasoning, MILS prompts the LLM to generate candidate outputs, each of which are scored and fed back iteratively, eventually generating a solution to the task. This enables various applications that typically require training specialized models on task-specific data. In particular, we establish a new state-of-the-art on emergent zero-shot image, video and audio captioning. MILS seamlessly applies to media generation as well, discovering prompt rewrites to improve text-to-image generation, and even edit prompts for style transfer! Finally, being a gradient-free optimization approach, MILS can invert multimodal embeddings into text, enabling applications like cross-modal arithmetic.

Event cameras are innovative sensors that capture brightness changes as asynchronous events rather than traditional intensity frames. These cameras offer substantial advantages over conventional cameras, including high temporal resolution, high dynamic range, and the elimination of motion blur. However, defocus blur, a common image quality degradation resulting from out-of-focus lenses, complicates the challenge of event-based imaging. Due to the unique imaging mechanism of event cameras, existing focusing algorithms struggle to operate efficiently on sparse event data. In this work, we propose EvFocus, a novel architecture designed to reconstruct sharp images from defocus event streams for the first time. Our work includes the development of an event-based out-of-focus camera model and a simulator to generate realistic defocus event streams for robust training and testing. EvDefous integrates a temporal information encoder, a blur-aware two-branch decoder, and a reconstruction and re-defocus module to effectively learn and correct defocus blur. Extensive experiments on both simulated and real-world datasets demonstrate that EvFocus outperforms existing methods across varying lighting conditions and blur sizes, proving its robustness and practical applicability in event-based defocus imaging.

We study reinforcement learning in infinite-horizon average-reward settings with linear MDPs. Previous work addresses this problem by approximating the average-reward setting by discounted setting and employing a value iteration-based algorithm that uses clipping to constrain the span of the value function for improved statistical efficiency. However, the clipping procedure requires computing the minimum of the value function over the entire state space, which is prohibitive since the state space in linear MDP setting can be large or even infinite. In this paper, we introduce a value iteration method with efficient clipping operation that only requires computing the minimum of value functions over the set of states visited by the algorithm. Our algorithm enjoys the same regret bound as the previous work while being computationally efficient, with computational complexity that is independent of the size of the state space.

Aligning large language models with humans is challenging due to the inherently multifaceted nature of preference feedback. While existing approaches typically frame this as a multi-objective optimization problem, they often overlook how humans actually make decisions. Research on bounded rationality suggests that human decision making follows satisficing strategies- optimizing primary objectives while ensuring others meet acceptable thresholds. To bridge this gap and operationalize the notion of satisficing alignment, we propose SITAlign: an inference-time framework that addresses the multifaceted nature of alignment by maximizing a primary objective while satisfying threshold-based constraints on secondary criteria. We provide theoretical insights by deriving sub-optimality bounds of our satisficing-based inference alignment approach. We empirically validate SITAlign's performance through extensive experimentation on multiple benchmarks. For instance, on the PKU-SafeRLHF dataset with the primary objective of maximizing helpfulness while ensuring a threshold on harmlessness, SITAlign outperforms the state-of-the-art multi-objective decoding strategy by a margin of 22.3% in terms of GPT-4 win-tie rate for helpfulness reward while adhering to the threshold on harmlessness.

Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of specifying the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. Inspired by reinforcement learning from observations, we propose **Co**ntrolling Large Language Models with **L**atent **A**ctions **CoLA**, a framework that integrates a latent action space into pre-trained LLMs. **CoLA** employs an \emph{inverse dynamics model} to extract latent actions conditioned on future tokens, ensuring that the next token prediction is partially influenced by these actions. Simultaneously, **CoLA** fine-tunes the pre-trained LLM to function as a \emph{language world model}, capable of incorporating latent actions as inputs. Additionally, **CoLA** trains a \emph{policy model} to generate actions within this language world model. The policy model can be trained via behavior cloning to mimic a standard language model or through RL to maximize task-specific rewards. In this work, we apply **CoLA** to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, **CoLA**'s latent actions enable greater semantic diversity. For enhancing downstream tasks, we show that **CoLA** with RL achieves a score of 42.4 on the \emph{math500} benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, **CoLA** with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, **CoLA** reduces computation time by half in tasks involving enhanced thinking prompts for LLMs via RL. These results highlight **CoLA**'s potential to advance RL-based adaptation of LLMs for downstream applications. The CoLA model is available at \url{https://huggingface.co/LAMDA-RL/Llama-3.1-CoLA-10B}.

Many dynamic decision problems, such as robotic control, involve a series of tasks, many of which are unknown at training time. Typical approaches for these problems, such as multi-task and meta reinforcement learning, do not generalize well when the tasks are diverse. On the other hand, approaches that aim to tackle task diversity, such as using task embedding as policy context and task clustering, typically lack performance guarantees and require a large number of training tasks. To address these challenges, we propose a novel approach for learning a policy committee that includes at least one near-optimal policy with high probability for tasks encountered during execution. While we show that this problem is in general inapproximable, we present two practical algorithmic solutions. The first yields provable approximation and task sample complexity guarantees when tasks are low-dimensional (the best we can do due to inapproximability), whereas the second is a general and practical gradient-based approach. In addition, we provide a provable sample complexity bound for few-shot learning. Our experiments on MuJoCo and Meta-World show that the proposed approach outperforms state-of-the-art multi-task, meta-, and task clustering baselines in training, generalization, and few-shot learning, often by a large margin. Our code is available at https://github.com/CERL-WUSTL/PACMAN.

Self-awareness, i.e., the ability to assess and correct one's generation, is a fundamental aspect of human intelligence, making its replication in large language models (LLMs) an important yet challenging task. Previous works tackle this by employing extensive reinforcement learning or relying on large external verifiers. In this work, we propose Refine via Intrinsic Self-Verification (ReVISE), an efficient and effective framework that enables LLMs to self-correct their outputs through self-verification. The core idea of ReVISE is to enable LLMs to verify their reasoning processes and continually rethink reasoning trajectories based on its verification. To implement this efficiently, we introduce a structured curriculum based on preference learning. Specifically, as ReVISE involves two challenging tasks (i.e., self-verification and reasoning correction), we tackle each task sequentially using curriculum learning, collecting both failed and successful reasoning paths to construct preference pairs for efficient training. During inference, our approach enjoys natural test-time scaling by integrating self-verification and correction capabilities, further enhanced by our proposed confidence-aware decoding mechanism. Our experiments on various reasoning tasks demonstrate that ReVISE achieves efficient self-correction and significantly improves the reasoning performance of LLMs.