ICML 2025 Papers

Gradient-based bilevel optimisation is a powerful technique with applications in hyperparameter optimisation, task adaptation, algorithm discovery, meta-learning more broadly, and beyond. It often requires differentiating through the gradient-based optimisation process itself, leading to "gradient-of-a-gradient" calculations with computationally expensive second-order and mixed derivatives. While modern automatic differentiation libraries provide a convenient way to write programs for calculating these derivatives, they oftentimes cannot fully exploit the specific structure of these problems out-of-the-box, leading to suboptimal performance. In this paper, we analyse such cases and propose Mixed-Flow Meta-Gradients, or MixFlow-MG -- a practical algorithm that uses mixed-mode differentiation to construct more efficient and scalable computational graphs yielding over 10x memory and up to 25\% wall-clock time improvements over standard implementations in modern meta-learning setups.

Despite the potential of language model-based agents to solve real-world tasks such as web navigation, current methods still struggle with long-horizon tasks with complex action trajectories. In contrast, humans can flexibly solve complex tasks by learning reusable task workflows from past experiences and using them to guide future actions. To build agents that can similarly benefit from this process, we introduce Agent Workflow Memory (AWM), a method for inducing commonly reused routines, i.e., workflows, and selectively providing workflows to the agent to guide subsequent generations. AWM flexibly applies to both offline and online scenarios, where agents induce workflows from training examples beforehand or from test queries on the fly. We experiment on two major web navigation benchmarks — Mind2Web and WebArena — that collectively cover 1000+ tasks from 200+ domains across travel, shopping, and social media, among others. AWM substantially improves the baseline results by 24.6% and 51.1% relative success rate on Mind2Web and WebArena while reducing the number of steps taken to solve WebArena tasks successfully. Furthermore, online AWM robustly generalizes in cross-task, website, and domain evaluations, surpassing baselines from 8.9 to 14.0 absolute points as train-test task distribution gaps widen.

Large-batch training has become a cornerstone in accelerating the training of deep neural networks, yet it poses challenges in optimization and generalization. Existing optimizers like AdamW present performance degradation during language models' large-batch training, due to the information bottleneck in attention layers caused by the sharp increase of max attention logit. While the LAMB optimizer partially addresses this issue, some attention layers still face this issue. The reason is that $l_2$-norm-based trust ratios in LAMB are less effective in directly influencing the max value of query/key weights. Furthermore, the weight-wise trust ratio in LAMB is error-prone as it overlooks relationships of weight values within rows or columns. Building on these observations, we propose a novel optimizer, MERIT, which leverages the max-norm to calculate the trust ratio to constrain the max attention logit more effectively. Moreover, we further construct element-wise trust ratios to provide more robust update scaling by focusing on local weight structures. Extensive experiments of large-batch training across various sizes of GPT-2 models demonstrate the superior performance of MERIT. Notably, during the training of GPT-2 Medium, MERIT enables a 6k batch size without any performance degradation compared to the standard batch size (480) with 48B training tokens. This work highlights the importance of considering the max attention logit and finer-granularity trust ratio in large-batch training. It successfully improves the training stability and paves the way for larger batch usage, enabling faster development and iteration of large language models. Code is available at https://github.com/NUS-HPC-AI-Lab/MERIT.

We introduce a new tool, Transductive Local Complexity (TLC), to analyze the generalization performance of transductive learning methods and motivate new transductive learning algorithms. Our work extends the idea of the popular Local Rademacher Complexity (LRC) to the transductive setting with considerable and novel changes compared to the analysis of typical LRC methods in the inductive setting. While LRC has been widely used as a powerful tool in the analysis of inductive models with sharp generalization bounds for classification and minimax rates for nonparametric regression, it remains an open problem whether a localized version of Rademacher complexity based tool can be designed and applied to transductive learning and gain sharp bound for transductive learning which is consistent with the inductive excess risk bound by (LRC). We give a confirmative answer to this open problem by TLC. Similar to the development of LRC, we build TLC by first establishing a novel and sharp concentration inequality for supremum of empirical processes for the gap between test and training loss in the setting of sampling uniformly without replacement. Then a peeling strategy and a new surrogate variance operator are used to derive the following excess risk bound in the transductive setting, which is consistent with that of the classical LRC based excess risk bound in the inductive setting. As an application of TLC, we use the new TLC tool to analyze the Transductive Kernel Learning (TKL) model, and derive sharper excess risk bound than that by the current state-of-the-art. As a result of independent interest, the concentration inequality for the test-train process is used to derive a sharp concentration inequality for the general supremum of empirical process involving random variables in the setting of sampling uniformly without replacement, with comparison to current concentration inequalities.

Diffusion Transformers (DiTs) excel in generative tasks but face practical deployment challenges due to high inference costs. Feature caching, which stores and retrieves redundant computations, offers the potential for acceleration. Existing learning-based caching, though adaptive, overlooks the impact of the prior timestep. It also suffers from misaligned objectives-*aligned predicted noise vs. high-quality images*-between training and inference. These two discrepancies compromise both performance and efficiency. To this end, we *harmonize* training and inference with a novel learning-based *caching* framework dubbed **HarmoniCa**. It first incorporates *Step-Wise Denoising Training* (SDT) to ensure the continuity of the denoising process, where prior steps can be leveraged. In addition, an *Image Error Proxy-Guided Objective* (IEPO) is applied to balance image quality against cache utilization through an efficient proxy to approximate the image error. Extensive experiments across $8$ models, $4$ samplers, and resolutions from $256\times256$ to $2K$ demonstrate superior performance and speedup of our framework. For instance, it achieves over $40\\%$ latency reduction (*i.e.*, $2.07\times$ theoretical speedup) and improved performance on PixArt-$\alpha$. Remarkably, our *image-free* approach reduces training time by $25\\%$ compared with the previous method. Our code is available at https://github.com/ModelTC/HarmoniCa.

The pursuit of universal black-box optimization (BBO) algorithms is a longstanding goal. However, unlike domains such as language or vision, where scaling structured data has driven generalization, progress in offline BBO remains hindered by the lack of unified representations for heterogeneous numerical spaces. Thus, existing offline BBO approaches are constrained to single-task and fixed-dimensional settings, failing to achieve cross-domain universal optimization. Recent advances in language models (LMs) offer a promising path forward: their embeddings capture latent relationships in a unifying way, enabling universal optimization across different data types possible. In this paper, we discuss multiple potential approaches, including an end-to-end learning framework in the form of next-token prediction, as well as prioritizing the learning of latent spaces with strong representational capabilities. To validate the effectiveness of these methods, we collect offline BBO tasks and data from open-source academic works for training. Experiments demonstrate the universality and effectiveness of our proposed methods. Our findings suggest that unifying language model priors and learning string embedding space can overcome traditional barriers in universal BBO, paving the way for general-purpose BBO algorithms. The code is provided at https://github.com/lamda-bbo/universal-offline-bbo.

Fine-Grained Openset Detection (FGOD) poses a fundamental challenge due to the similarity between the openset classes and those closed-set ones. Since real-world objects/entities tend to form a hierarchical structure, the fine-grained relationship among the closed-set classes as captured by the hierarchy could potentially improve the FGOD performance. Intuitively, the hierarchical dependency among different classes allows the model to recognize their subtle differences, which in turn makes it better at differentiating similar open-set classes even they may share the same parent. However, simply performing openset detection in a top-down fashion by building a local detector for each node may result in a poor detection performance. Our theoretical analysis also reveals that maximizing the probability of the path leading to the ground-truth leaf node also results in a sub-optimal training process. To systematically address this issue, we propose to formulate a novel state-based node representation, which constructs a state space based upon the entire hierarchical structure. We prove that the state-based representation guarantees to maximize the probability on the path leading to the ground-truth leaf node. Extensive experiments on multiple real-world hierarchical datasets clearly demonstrate the superior performance of the proposed method.

This paper introduces the SOAR framework for imitation learning. SOAR is an algorithmic template which learns a policy from expert demonstrations with a primal-dual style algorithm which alternates cost and policy updates. Within the policy updates the SOAR framework prescribe to use an actor critic method with multiple critics to estimate the critic uncertainty and therefore build an optimistic critic fundamental to drive exploration. When instantiated to the tabular setting, we get a provable algorithms dubbed FRA with guarantees matching the best known results in $\epsilon$. Practically, the SOAR template is shown to boost consistently the performance of primal dual IL algorithms building on actor critic routines for the policy updates. Approximately, thanks to SOAR, the required number of episodes to achieve the same performance is reduced by a half.

Despite their remarkable capabilities, large language models often struggle with tasks requiring complex reasoning and planning. While existing approaches like Chain-of-Thought prompting and tree search techniques show promise, they are limited by their reliance on predefined heuristics and computationally expensive exploration strategies. We propose Policy-Guided Tree Search (PGTS), a framework that combines reinforcement learning with structured tree exploration to efficiently navigate reasoning paths. Our key innovation is a learned policy that dynamically decides between expanding, branching, backtracking, or terminating exploration, eliminating the need for manual heuristics or exhaustive search. Experiments across mathematical reasoning, logical deduction, and planning benchmarks demonstrate that PGTS achieves superior reasoning performance while significantly reducing computational costs compared to existing methods. These results establish PGTS as a scalable and effective solution for tackling complex reasoning tasks with LLMs.

Large language models (LLMs) have recently achieved remarkable results in complex multi-step tasks, such as mathematical reasoning and agentic software engineering. However, they often struggle to maintain consistent performance across multiple solution attempts. One effective approach to narrow the gap between average-case and best-case performance is guided test-time search, which explores multiple solution paths to identify the most promising one. Unfortunately, effective search techniques (e.g. MCTS) are often unsuitable for *non-serializable* RL environments, such as Docker containers, where intermediate environment states cannot be easily saved and restored. We investigate two complementary search strategies applicable to such environments: 1-step lookahead and trajectory selection, both guided by a learned action-value function estimator. On the SWE-bench Verified benchmark, a key testbed for agentic software engineering, we find these methods to double the average success rate of a fine-tuned Qwen-72B model, achieving $40.8$\%, the new state-of-the-art for open-weights models. Additionally, we show that these techniques are transferable to more advanced closed models, yielding similar improvements with GPT-4o.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Our framework addresses two critical questions: (1) how to generate high-quality reasoning processes during inference automatically, and (2) how to integrate these processes into post-training. We propose the \emph{Bootstrapping Reinforced Thinking Process} (BRiTE) algorithm and demonstrate its theoretical convergence at a rate of $1/T$, where $T$ is the number of iterations. The algorithm operates in two steps. First, it generates high-quality rationales by approximating the desired posterior distribution using a reinforcement learning approach with a novel reward shaping mechanism. Second, it fine-tunes the base LLM by maximizing the joint probability of rationale generation with respect to LLM parameters. Empirical evaluation on GSM8K and MATH benchmarks demonstrates that our approach consistently improves performance across different model sizes without requiring human-annotated thinking processes, outperforming standard chain-of-thought prompting while enhancing existing post-training methods.

Previous influential work showed that infinite width limits of neural networks in the lazy training regime are described by kernel machines. Here, we show that neural networks trained in the rich infinite-width regime in two different settings are also described by kernel machines, but with data-dependent kernels. For both cases, we provide explicit expressions for the kernel predictors and prescriptions to numerically calculate them. To derive the first predictor, we study the large-width limit of feature-learning Bayesian networks, showing how feature learning leads to task-relevant adaptation of layer kernels and preactivation densities. The saddle point equations governing this limit result in a min-max optimization problem that defines the kernel predictor. To derive the second predictor, we study gradient flow training of randomly initialized networks trained with weight decay in the infinite-width limit using dynamical mean field theory (DMFT). The fixed point equations of the arising DMFT defines the task-adapted internal representations and the kernel predictor. We compare our kernel predictors to kernels derived from lazy regime and demonstrate that our adaptive kernels achieve lower test loss on benchmark datasets.

Score-based generative models are essential in various machine learning applications, with strong capabilities in generation quality. In particular, high-order derivatives (scores) of data density offer deep insights into data distributions, building on the proven effectiveness of first-order scores for modeling and generating synthetic data, unlocking new possibilities for applications. However, learning them typically requires complete data, which is often unavailable in domains such as healthcare and finance due to data corruption, acquisition constraints, or incomplete records. To tackle this challenge, we introduce MissScore, a novel framework for estimating high-order scores in the presence of missing data. We derive objective functions for estimating high-order scores under different missing data mechanisms and propose a new algorithm specifically designed to handle missing data effectively. Our empirical results demonstrate that MissScore accurately and efficiently learns the high-order scores from incomplete data and generates high-quality samples, resulting in strong performance across a range of downstream tasks.

We present Self-Organizing Visual Prototypes (SOP), a new training technique for unsupervised visual feature learning. Unlike existing prototypical self-supervised learning (SSL) methods that rely on a single prototype to encode all relevant features of a hidden cluster in the data, we propose the SOP strategy. In this strategy, a prototype is represented by many semantically similar representations, or support embeddings (SEs), each containing a complementary set of features that together better characterize their region in space and maximize training performance. We reaffirm the feasibility of non-parametric SSL by introducing novel non-parametric adaptations of two loss functions that implement the SOP strategy. Notably, we introduce the SOP Masked Image Modeling (SOP-MIM) task, where masked representations are reconstructed from the perspective of multiple non-parametric local SEs. We comprehensively evaluate the representations learned using the SOP strategy on a range of benchmarks, including retrieval, linear evaluation, fine-tuning, and object detection. Our pre-trained encoders achieve state-of-the-art performance on many retrieval benchmarks and demonstrate increasing performance gains with more complex encoders.

In the differentially private partition selection problem (a.k.a. set union, key discovery), users hold subsets of items from an unbounded universe. The goal is to output as many items as possible from the union of the users' sets while maintaining user-level differential privacy. Solutions to this problem are a core building block for many privacy-preserving ML applications including vocabulary extraction in a private corpus, computing statistics over categorical data and learning embeddings over user-provided items. We propose an algorithm for this problem, MaxAdaptiveDegree (MAD), which adaptively reroutes weight from items with weight far above the threshold needed for privacy to items with smaller weight, thereby increasing the probability that less frequent items are output. Our algorithm can be efficiently implemented in massively parallel computation systems allowing scalability to very large datasets. We prove that our algorithm stochastically dominates the standard parallel algorithm for this problem. We also develop a two-round version of our algorithm, MAD2R, where results of the computation in the first round are used to bias the weighting in the second round to maximize the number of items output. In experiments, our algorithms provide the best results among parallel algorithms and scale to datasets with hundreds of billions of items, up to three orders of magnitude larger than those analyzed in prior works.

Vision-Language Models (VLMs) face two critical limitations in visual representation learning: degraded supervision due to information loss during gradient propagation, and the inherent semantic sparsity of textual supervision compared to visual data. We propose the Diffusion Supervision Vision-Language Model (DS-VLM), a plug-and-play framework that introduces diffusion-based direct supervision for vision-language alignment. By reconstructing input images through a diffusion model conditioned on outputs of the visual encoder and the connector, our method establishes a short-path gradient propagation channel from pixel space to visual features. This approach simultaneously preserves high-level semantic alignment through conventional text supervision while enhancing visual feature quality via pixel-level reconstruction constraints. Extensive experiments conducted across various visual encoders and LLMs of different scales demonstrate the effectiveness of our approach.

We introduce Mean Field Trust Region Policy Optimization (MF-TRPO), a novel algorithm designed to compute approximate Nash equilibria for ergodic Mean Field Games (MFGs) in finite state-action spaces. Building on the well-established performance of TRPO in the reinforcement learning (RL) setting, we extend its methodology to the MFG framework, leveraging its stability and robustness in policy optimization. Under standard assumptions in the MFG literature, we provide a rigorous analysis of MF-TRPO, establishing theoretical guarantees on its convergence. Our results cover both the exact formulation of the algorithm and its sample-based counterpart, where we derive high-probability guarantees and finite sample complexity. This work advances MFG optimization by bridging RL techniques with mean-field decision-making, offering a theoretically grounded approach to solving complex multi-agent problems.

Inductive knowledge graph completion aims to predict missing triplets in an incomplete knowledge graph that differs from the one observed during training. While subgraph reasoning models have demonstrated empirical success in this task, their theoretical properties, such as stability and generalization capability, remain unexplored. In this work, we present the first theoretical analysis of the relationship between the stability and the generalization capability for subgraph reasoning models. Specifically, we define stability as the degree of consistency in a subgraph reasoning model's outputs in response to differences in input subgraphs and introduce the Relational Tree Mover’s Distance as a metric to quantify the differences between the subgraphs. We then show that the generalization capability of subgraph reasoning models, defined as the discrepancy between the performance on training data and test data, is proportional to their stability. Furthermore, we empirically analyze the impact of stability on generalization capability using real-world datasets, validating our theoretical findings.

Diffusion-based generative models (diffusion models) often require a large amount of data to train a score-based model that learns the score function of the data distribution through denoising score matching. However, collecting and cleaning such data can be expensive, time-consuming, and even infeasible. In this paper, we present a novel theoretical insight for diffusion models that two factors, i.e., the denoiser function hypothesis space and the number of training samples, can affect the denoising score matching error of all training samples. Based on this theoretical insight, it is evident that minimizing the total denoising score matching error is challenging within the denoiser function hypothesis space in existing methods, when training diffusion models with limited data. To address this, we propose a new diffusion model called Limited Data Diffusion (LD-Diffusion), which consists of two main components: a compressing model and a novel mixed augmentation with fixed probability (MAFP) strategy. Specifically, the compressing model can constrain the complexity of the denoiser function hypothesis space and MAFP can effectively increase the training samples by providing more informative guidance than existing data augmentation methods in the compressed hypothesis space. Extensive experiments on several datasets demonstrate that LD-Diffusion can achieve better performance compared to other diffusion models. Codes are available at https://github.com/zzhang05/LD-Diffusion.