ICML 2025 Papers

Even when massively overparameterized, deep neural networks show a remarkable ability to generalize. Research on this phenomenon has focused on generalization within distribution, via smooth interpolation. Yet in some settings neural networks also learn to extrapolate to data far beyond the bounds of the original training set, sometimes even allowing for infinite generalization, implying that an algorithm capable of solving the task has been learned. Here we undertake a case study of the learning dynamics of recurrent neural networks trained on the streaming parity task in order to develop an effective theory of algorithm development. The streaming parity task is a simple but nonlinear task defined on sequences up to arbitrary length. We show that, with sufficient finite training experience, RNNs exhibit a phase transition to perfect infinite generalization. Using an effective theory for the representational dynamics, we find an implicit representational merger effect which can be interpreted as the construction of a finite automaton that reproduces the task. Overall, our results disclose one mechanism by which neural networks can generalize infinitely from finite training experience.

Ensuring safety is a critical challenge in applying Reinforcement Learning (RL) to real-world scenarios. Constrained Reinforcement Learning (CRL) addresses this by maximizing returns under predefined constraints, typically formulated as the expected cumulative cost. However, expectation-based constraints overlook rare but high-impact extreme value events in the tail distribution, such as black swan incidents, which can lead to severe constraint violations. To address this issue, we propose the Extreme Value policy Optimization (EVO) algorithm, leveraging Extreme Value Theory (EVT) to model and exploit extreme reward and cost samples, reducing constraint violations. EVO introduces an extreme quantile optimization objective to explicitly capture extreme samples in the cost tail distribution. Additionally, we propose an extreme prioritization mechanism during replay, amplifying the learning signal from rare but high-impact extreme samples. Theoretically, we establish upper bounds on expected constraint violations during policy updates, guaranteeing strict constraint satisfaction at a zero-violation quantile level. Further, we demonstrate that EVO achieves a lower probability of constraint violations than expectation-based methods and exhibits lower variance than quantile regression methods. Extensive experiments show that EVO significantly reduces constraint violations during training while maintaining competitive policy performance compared to baselines.

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as *AI Oversight*. We study how model similarity affects both aspects of AI oversight by proposing *Chance Adjusted Probabilistic Agreement (CAPA)*--a metric for LM similarity based on overlap in model mistakes. Using CAPA, we first show that *LLM-as-a-judge* scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from *weak-to-strong generalization*. As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend--model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

Experimentation with interference poses a significant challenge in contemporary online platforms. Prior research on experimentation with interference has concentrated on the final output of a policy. Cumulative performance, while equally important, is less well understood. To address this gap, we introduce the problem of Multi-armed Bandits with Interference (MABI), where the learner assigns an arm to each of $N$ experimental units over $T$ rounds. The reward of each unit in each round depends on the treatments of all units, where the interference between two units decays in their distance. The reward functions are chosen by an adversary and may vary arbitrarily over time and across different units. We first show that the optimal expected regret (against the best fixed-arm policy) is $\tilde O(\sqrt T)$, and can be achieved by a switchback policy. However, the regret (as a random variable) for any switchback policy suffers a high variance, since it does not account for $N$. We propose a policy based on a novel clustered randomization scheme, whose regret (i) is optimal in expectation and (ii) admits a high probability bound that vanishes in $N$.

Fine-tuning large language models (LLMs) with low-rank adaptations (LoRAs) has become common practice, often yielding numerous copies of the same LLM differing only in their LoRA updates. This paradigm presents challenges for systems that serve real-time responses to queries that each involve a different LoRA. Prior works optimize the design of such systems but still require continuous loading and offloading of LoRAs, as it is infeasible to store thousands of LoRAs in GPU memory. To mitigate this issue, we investigate the efficacy of compression when serving LoRAs. We propose a method for the joint compression of LoRAs into a shared basis paired with LoRA-specific scaling matrices. We extend our algorithm to learn clusters of LoRAs that are amenable to joint compression, allowing it to scale gracefully to large LoRA collections. Our experiments with up to 1000 LoRAs demonstrate that compressed LoRAs preserve performance while offering major throughput gains in realistic serving scenarios with over a thousand LoRAs, maintaining 80\% of the throughput of serving a single LoRA.

We explore ensembles of axis-aligned decision stumps, which can be viewed as a generalized additive model (GAM). In this model, stumps utilizing the same feature are grouped to form a shape function for that feature. Instead of relying on boosting or bagging, we employ alternating optimization to learn a fixed-size stump forest. We optimize the parameters of each stump exactly through enumeration, given the other stumps are fixed. For fixed stump splits, the leaf values are optimized jointly by solving a convex problem. To address the overfitting issue inherent in naive optimization of stump forests, we propose effective regularization techniques. Our regularized stump forests achieve accuracy comparable to state-of-the-art GAM methods while using fewer parameters. This work is the first to successfully learn stump forests without employing traditional ensembling techniques like bagging or boosting.

Reflectance attributes in LiDAR point clouds provide essential information for downstream tasks but remain underexplored in neural compression methods. To address this, we introduce SerLiC, a serialization-based neural compression framework to fully exploit the intrinsic characteristics of LiDAR reflectance. SerLiC first transforms 3D LiDAR point clouds into 1D sequences via scan-order serialization, offering a device-centric perspective for reflectance analysis. Each point is then tokenized into a contextual representation comprising its sensor scanning index, radial distance, and prior reflectance, for effective dependencies exploration. For efficient sequential modeling, Mamba is incorporated with a dual parallelization scheme, enabling simultaneous autoregressive dependency capture and fast processing. Extensive experiments demonstrate that SerLiC attains over 2$\times$ volume reduction against the original reflectance data, outperforming the state-of-the-art method by up to 22\% reduction of compressed bits while using only 2\% of its parameters. Moreover, a lightweight version of SerLiC achieves $\geq 10$ fps (frames per second) with just 111K parameters, which is attractive for real applications.

We study how multi-head softmax attention models are trained to perform in-context learning on linear data. Through extensive empirical experiments and rigorous theoretical analysis, we demystify the emergence of elegant attention patterns: a diagonal and homogeneous pattern in the key-query weights, and a last-entry-only and zero-sum pattern in the output-value weights. Remarkably, these patterns consistently appear from gradient-based training starting from random initialization. Our analysis reveals that such emergent structures enable multi-head attention to approximately implement a debiased gradient descent predictor --- one that outperforms single-head attention and nearly achieves Bayesian optimality up to proportional factor. We also extend our study to scenarios with anisotropic covariates and multi-task linear regression. Our results reveal that in-context learning ability emerges from the trained transformer as an aggregated effect of its architecture and the underlying data distribution, paving the way for deeper understanding and broader applications of in-context learning.

Fine-tuning large-scale pre-trained models is prohibitively expensive in terms of computation and memory costs. Low-Rank Adaptation (LoRA), a popular Parameter-Efficient Fine-Tuning (PEFT) method, offers an efficient solution by optimizing only low-rank matrices. Despite recent progress in improving LoRA's performance, the relationship between the LoRA optimization space and the full parameter space is often overlooked. A solution that appears flat in the loss landscape of the LoRA space may still exhibit sharp directions in the full parameter space, potentially compromising generalization. We introduce Flat-LoRA, which aims to identify a low-rank adaptation situated in a flat region of the full parameter space. Instead of adopting the well-established sharpness-aware minimization approach, which incurs significant computation and memory overheads, we employ a Bayesian expectation loss objective to preserve training efficiency. Further, we design a refined strategy for generating random perturbations to enhance performance and carefully manage memory overhead using random seeds. Experiments across diverse tasks—including mathematical reasoning, coding abilities, dialogue generation, instruction following, and text-to-image generation—demonstrate that Flat-LoRA improves both in-domain and out-of-domain generalization. Code is available at https://github.com/nblt/Flat-LoRA.

Building a virtual cell capable of accurately simulating cellular behaviors in silico has long been a dream in computational biology. We introduce CellFlux, an image-generative model that simulates cellular morphology changes induced by chemical and genetic perturbations using flow matching. Unlike prior methods, CellFlux models distribution-wise transformations from unperturbed to perturbed cell states, effectively distinguishing actual perturbation effects from experimental artifacts such as batch effects—a major challenge in biological data. Evaluated on chemical (BBBC021), genetic (RxRx1), and combined perturbation (JUMP) datasets, CellFlux generates biologically meaningful cell images that faithfully capture perturbation-specific morphological changes, achieving a 35% improvement in FID scores and a 12% increase in mode-of-action prediction accuracy over existing methods. Additionally, CellFlux enables continuous interpolation between cellular states, providing a potential tool for studying perturbation dynamics. These capabilities mark a significant step toward realizing virtual cell modeling for biomedical research. Project page: https://yuhui-zh15.github.io/CellFlux/.

Deep networks are prone to catastrophic forgetting during sequential task learning, i.e., losing the knowledge about old tasks upon learning new tasks. To this end, continual learning (CL) has emerged, whose existing methods focus mostly on regulating or protecting the parameters associated with the previous tasks. However, parameter protection is often impractical, since the size of parameters for storing the old-task knowledge increases linearly with the number of tasks, otherwise it is hard to preserve the parameters related to the old-task knowledge. In this work, we bring a dual opinion from neuroscience and physics to CL: in the whole networks, the pathways matter more than the parameters when concerning the knowledge acquired from the old tasks. Following this opinion, we propose a novel CL framework, learning without isolation (LwI), where model fusion is formulated as graph matching and the pathways occupied by the old tasks are protected without being isolated. Thanks to the sparsity of activation channels in a deep network, LwI can adaptively allocate available pathways for a new task, realizing pathway protection and addressing catastrophic forgetting in a parameter-effcient manner. Experiments on popular benchmark datasets demonstrate the superiority of the proposed LwI.

We study how inherent randomness in the training process—where each sample (or client in federated learning) contributes only to a randomly selected portion of training—can be leveraged for privacy amplification. This includes (1) data partitioning, where a sample participates in only a subset of training iterations, and (2) model partitioning, where a sample updates only a subset of the model parameters. We apply our framework to model parallelism in federated learning, where each client updates a randomly selected subnetwork to reduce memory and computational overhead, and show that existing methods, e.g. model splitting or dropout, provide a significant privacy amplification gain not captured by previous privacy analysis techniques. Additionally, we introduce balanced iteration subsampling, a new data partitioning method where each sample (or client) participates in a fixed number of training iterations. We show that in certain regimes, this method yields stronger privacy amplification than Poisson (i.i.d.) sampling of data (or clients). Our results demonstrate that randomness in the training process, which is structured rather than i.i.d. and interacts with data in complex ways, can be systematically leveraged for nontrivial privacy amplification.

This work investigates the problem of best arm identification for multi-agent multi-armed bandits. We consider $N$ agents grouped into $M$ clusters, where each cluster solves a stochastic bandit problem. The mapping between agents and bandits is \textit{a priori} unknown. Each bandit is associated with $K$ arms, and the goal is to identify the best arm for each agent under a $\delta$-probably correct ($\delta$-PC) framework, while minimizing sample complexity and communication overhead. We propose two novel algorithms: \emph{Clustering then Best Arm Identification} (\texttt{Cl-BAI}) and \emph{Best Arm Identification then Clustering} (\texttt{BAI-Cl}). \texttt{Cl-BAI} employs a two-phase approach that first clusters agents based on the bandit problems they are learning, followed by identifying the best arm for each cluster. \texttt{BAI-Cl} reverses the sequence by identifying the best arms first and then clustering agents accordingly. Both algorithms exploit the successive elimination framework to ensure computational efficiency and high accuracy. Theoretical analysis establishes $\delta$-PC guarantees for both methods, derives bounds on their sample complexity, and provides a lower bound for the problem class. Moreover, when $M$ is small (a constant), we show that the sample complexity of (a variant of) \texttt{BAI-Cl} is (order-wise) minimax optimal. Experiments on synthetic and real-world (Movie Lens, Yelp) data demonstrates the superior performance of the proposed algorithms in terms of sample and communication efficiency, particularly in settings where $M \ll N$.

Offline algorithms for Reinforcement Learning from Human Preferences (RLHF), which use only a fixed dataset of sampled responses given an input, and preference feedback among these responses, have gained increasing prominence in the literature on aligning language models. In this paper, we study how the different design choices made in methods such as DPO, IPO, SLiC and many variants influence the quality of the learned policy, from a theoretical perspective. Our treatment yields insights into the choices of loss function, the policy which is used to normalize log-likelihoods, and also the role of the data sampling policy. Notably, our results do not rely on the standard reparameterization-style arguments used to motivate some of the algorithms in this family, which allows us to give a unified treatment to a broad class of methods. We also conduct a small empirical study to verify some of the theoretical findings on a standard summarization benchmark.

Human vision is dynamic and continuous. However, in video understanding with multimodal large language models (LLMs), existing methods primarily rely on static features extracted from images sampled at a fixed low frame rate of frame-per-second (FPS) $\leqslant$2, leading to critical visual information loss. In this paper, we introduce F-16, the first multimodal LLM designed for high-frame-rate video understanding. By increasing the frame rate to 16 FPS and compressing visual tokens within each 1-second clip, F-16 efficiently captures dynamic visual features while preserving key semantic information. Experimental results demonstrate that higher frame rates considerably enhance video understanding across multiple benchmarks, providing a new approach to improving video LLMs beyond scaling model size or training data. F-16 achieves state-of-the-art performance among 7-billion-parameter video LLMs on both general and fine-grained video understanding benchmarks, such as Video-MME and TemporalBench. Furthermore, F-16 excels in complex spatiotemporal tasks, including high-speed sports analysis (*e.g.*, basketball, football, gymnastics, and diving), outperforming SOTA proprietary visual models like GPT-4o and Gemini-1.5-pro. Additionally, we introduce a novel decoding method for F-16 that enables highly efficient low-frame-rate inference without requiring model retraining. We will release the source code, model checkpoints, and data at [https://github.com/bytedance/F-16](https://github.com/bytedance/F-16).

Despite the superior capability in complementary information exploration and consistent clustering structure learning, most current weight-based multi-modal clustering methods still contain three limitations: 1) lack of trustworthiness in learned weights; 2) isolated view weight learning; 3) extra weight parameters. Motivated by the peer-review mechanism in the academia, we in this paper give a new peer-review look on the multi-modal clustering problem and propose to iteratively treat one modality as "author" and the remaining modalities as "reviewers" so as to reach a peer-review score for each modality. It essentially explores the underlying relationships among modalities. To improve the trustworthiness, we further design a new trustworthy score with a self-supervision working mechanism. Following that, we propose a novel Peer-review Trustworthy Information Bottleneck (PTIB) method for weighted multi-modal clustering, where both the above scores are simultaneously taken into account for accurate and parameter-free modality weight learning. Extensive experiments on eight multi-modal datasets suggest that PTIB can outperform the state-of-the-art multi-modal clustering methods.

This work identifies the first privacy-aware stratified sampling scheme that minimizes the variance for general private mean estimation under the Laplace, Discrete Laplace (DLap) and Truncated-Uniform-Laplace (TuLap) mechanisms within the framework of differential privacy (DP). We view stratified sampling as a subsampling operation, which amplifies the privacy guarantee; however, to have the same final privacy guarantee for each group, different nominal privacy budgets need to be used depending on the subsampling rate. Ignoring the effect of DP, traditional stratified sampling strategies risk significant variance inflation. We phrase our optimal survey design as an optimization problem, where we determine the optimal subsampling sizes for each group with the goal of minimizing the variance of the resulting estimator. We establish strong convexity of the variance objective, propose an efficient algorithm to identify the integer-optimal design, and offer insights on the structure of the optimal design.

LLM-based multi-agent systems (MAS) have shown significant potential in tackling diverse tasks. However, to design effective MAS, existing approaches heavily rely on manual configurations or multiple calls of advanced LLMs, resulting in inadaptability and high inference costs. In this paper, we simplify the process of building an MAS by reframing it as a generative language task, where the input is a user query and the output is a corresponding MAS. To address this novel task, we unify the representation of MAS as executable code and propose a consistency-oriented data construction pipeline to create a high-quality dataset comprising coherent and consistent query-MAS pairs. Using this dataset, we train MAS-GPT, an open-source medium-sized LLM that is capable of generating query-adaptive MAS within a single LLM inference. The generated MAS can be seamlessly applied to process user queries and deliver high-quality responses. Extensive experiments on 9 benchmarks and 5 LLMs show that the proposed MAS-GPT consistently outperforms 10+ baseline MAS methods on diverse settings, indicating MAS-GPT's high effectiveness, efficiency and strong generalization ability. The codes are released at \url{https://github.com/rui-ye/MAS-GPT}.

While large-scale training data is fundamental for developing capable large language models (LLMs), strategically selecting high-quality data has emerged as a critical approach to enhance training efficiency and reduce computational costs. Current data selection methodologies predominantly rely on static, training-agnostic criteria, failing to account for the dynamic model training and data interactions. In this paper, we propose a new Data Weighting Model (DWM) to adjust the weight of selected data within each batch to achieve a dynamic data utilization during LLM training. Specially, to better capture the dynamic data preference of the trained model, a bi-level optimization framework is implemented to update the weighting model. Our experiments demonstrate that DWM enhances the performance of models trained with randomly-selected data, and the learned weighting model can be transferred to enhance other data selection methods and models of different sizes. Moreover, we further analyze how a model’s data preferences evolve throughout training, providing new insights into the data preference of the model during training.