ICML 2025 Papers

Despite the success of distillation in large language models (LLMs), most prior work applies identical loss functions to both teacher- and student-generated data. These strategies overlook the synergy between loss formulations and data types, leading to a suboptimal performance boost in student models. To address this, we propose DistiLLM-2, a contrastive approach that simultaneously increases the likelihood of teacher responses and decreases that of student responses by harnessing this synergy. Our extensive experiments show that DistiLLM-2 not only builds high-performing student models across a wide range of tasks, including instruction-following and code generation, but also supports diverse applications, such as preference alignment and vision-language extensions. These findings highlight the potential of a contrastive approach to enhance the efficacy of LLM distillation by effectively aligning teacher and student models across varied data types.

We consider the basic statistical challenge of designing an "all-purpose" mean estimation algorithm that is recommendable across a variety of settings and models. Recent work by [Lee and Valiant 2022] introduced the first 1-d mean estimator whose error in the standard finite-variance+i.i.d. setting is optimal even in its constant factors; experimental demonstration of its good performance was shown by [Gobet et al. 2022]. Yet, unlike for classic (but not necessarily practical) estimators such as median-of-means and trimmed mean, this new algorithm lacked proven robustness guarantees in other settings, including the settings of adversarial data corruption and heavy-tailed distributions with infinite variance. Such robustness is important for practical use cases. This raises a research question: is it possible to have a mean estimator that is robust, *without* sacrificing provably optimal performance in the standard i.i.d. setting? In this work, we show that Lee and Valiant's estimator is in fact an "all-purpose" mean estimator by proving: (A) It is robust to an $\eta$-fraction of data corruption, even in the strong contamination model; it has optimal estimation error $O(\sigma\sqrt{\eta})$ for distributions with variance $\sigma^2$. (B) For distributions with finite $z^\text{th}$ moment, for $z \in (1,2)$, it has optimal estimation error, matching the lower bounds of [Devroye et al. 2016] up to constants. We further show (C) that outlier robustness for 1-d mean estimators in fact implies neighborhood optimality, a notion of beyond worst-case and distribution-dependent optimality recently introduced by [Dang et al. 2023]. Previously, such an optimality guarantee was only known for median-of-means, but now it holds also for all estimators that are simultaneously *robust* and *sub-Gaussian*, including Lee and Valiant's, resolving a question raised by Dang et al. Lastly, we show (D) the asymptotic normality and efficiency of Lee and Valiant's estimator, as further evidence for its performance across many settings.

Neural Processes (NPs) are deep probabilistic models that represent stochastic processes by conditioning their prior distributions on a set of context points. Despite their advantages in uncertainty estimation for complex distributions, NPs enforce parameterization coupling between the conditional prior model and the posterior model. We show that this coupling amounts to prior misspecification and revisit the NP objective to address this issue. More specifically, we propose Rényi Neural Processes (RNP), a method that replaces the standard KL divergence with the Rényi divergence, dampening the effects of the misspecified prior during posterior updates. We validate our approach across multiple benchmarks including regression and image inpainting tasks, and show significant performance improvements of RNPs in real-world problems. Our extensive experiments show consistently better log-likelihoods over state-of-the-art NP models.

Diffusion models and Flow Matching generate high-quality samples but are slow at inference, and distilling them into few-step models often leads to instability and extensive tuning. To resolve these trade-offs, we propose Moment Matching Self-Distillation (MMSD), a new class of generative models for one- or few-step sampling with a single-stage training procedure. Unlike distillation, MMSD does not require pre-training initialization and optimization of two networks; and unlike Consistency Models, MMSD guarantees distribution-level convergence and remains stable under various hyperparameters and standard model architectures. MMSD surpasses diffusion models on ImageNet-256x256 with 2.13 FID using only 8 inference steps and achieves state-of-the-art 2-step FID of 2.05 on CIFAR-10 for a model trained from scratch.

Understanding and constructing brain communications that capture dynamic communications across multiple regions is fundamental to modern system neuroscience, yet current methods struggle to find time-varying region-level communications or scale to large neural datasets with long recording durations. We present a novel framework using Markovian Gaussian Processes to learn brain communications with time-varying temporal delays from multi-region neural recordings, named Adaptive Delay Model (ADM). Our method combines Gaussian Processes with State Space Models and employs parallel scan inference algorithms, enabling efficient scaling to large datasets while identifying concurrent communication patterns that evolve over time. This time-varying approach captures how brain region interactions shift dynamically during cognitive processes. Validated on synthetic and multi-region neural recordings datasets, our approach discovers both the directionality and temporal dynamics of neural communication. This work advances our understanding of distributed neural computation and provides a scalable tool for analyzing dynamic brain networks. Code is available at https://github.com/BRAINML-GT/Adaptive-Delay-Model.

Low-rank adaptation (LoRA) has become a standard approach for fine-tuning large foundation models. However, our theoretical understanding of LoRA remains limited as prior analyses of LoRA's training dynamics either rely on linearization arguments or consider highly simplified setups. In this work, we analyze the LoRA loss landscape without such restrictive assumptions. We define two regimes: a "special regime", which includes idealized setups where linearization arguments hold, and a "generic regime" representing more realistic setups where linearization arguments do not hold. In the generic regime, we show that LoRA training converges to a global minimizer with low rank and small magnitude, or a qualitatively distinct solution with high rank and large magnitude. Finally, we argue that the zero-initialization and weight decay in LoRA training induce an implicit bias toward the low-rank, small-magnitude region of the parameter space—where global minima lie—thus shedding light on why LoRA training usually succeeds in finding global minima.

Constrained Reinforcement Learning (CRL) aims to maximize cumulative rewards while satisfying constraints. However, existing CRL algorithms often encounter significant constraint violations during training, limiting their applicability in safety-critical scenarios. In this paper, we identify the underestimation of the cost value function as a key factor contributing to these violations. To address this issue, we propose the Memory-driven Intrinsic Cost Estimation (MICE) method, which introduces intrinsic costs to mitigate underestimation and control bias to promote safer exploration. Inspired by flashbulb memory, where humans vividly recall dangerous experiences to avoid risks, MICE constructs a memory module that stores previously explored unsafe states to identify high-cost regions. The intrinsic cost is formulated as the pseudo-count of the current state visiting these risk regions. Furthermore, we propose an extrinsic-intrinsic cost value function that incorporates intrinsic costs and adopts a bias correction strategy. Using this function, we formulate an optimization objective within the trust region, along with corresponding optimization methods. Theoretically, we provide convergence guarantees for the proposed cost value function and establish the worst-case constraint violation for the MICE update. Extensive experiments demonstrate that MICE significantly reduces constraint violations while preserving policy performance comparable to baselines.

Efficiently counting Markov equivalent directed acyclic graphs (DAGs) is crucial in graphical causal analysis. Wienöbst et al. (2023) introduced a polynomial-time algorithm, known as the Clique-Picking algorithm, to count the number of Markov equivalent DAGs for a given completed partially directed acyclic graph (CPDAG). This algorithm iteratively selects a root clique, determines fixed orientations with outgoing edges from the clique, and generates the unresolved undirected connected components (UCCGs). In this work, we propose a more efficient approach to UCCG generation by utilizing previously computed results for different root cliques. Our method introduces the concept of super cliques within rooted clique trees, enabling their efficient transfer between trees with different root cliques. The proposed algorithm effectively reduces the computational complexity of the Clique-Picking method, particularly when the number of cliques is substantially smaller than the number of vertices and edges.

Effectively scaling up deep reinforcement learning models has proven notoriously difficult due to network pathologies during training, motivating various targeted interventions such as periodic reset and architectural advances such as layer normalization. Instead of pursuing more complex modifications, we show that introducing static network sparsity alone can unlock further scaling potential beyond their dense counterparts with state-of-the-art architectures. This is achieved through simple one-shot random pruning, where a predetermined percentage of network weights are randomly removed once before training. Our analysis reveals that, in contrast to naively scaling up dense DRL networks, such sparse networks achieve both higher parameter efficiency for network expressivity and stronger resistance to optimization challenges like plasticity loss and gradient interference. We further extend our evaluation to visual and streaming RL scenarios, demonstrating the consistent benefits of network sparsity.

Score matching is a vital tool for learning the distribution of data with applications across many areas including diffusion processes, energy based modelling, and graphical model estimation. Despite all these applications, little work explores its use when data is incomplete. We address this by adapting score matching (and its major extensions) to work with missing data in a flexible setting where data can be partially missing over any subset of the coordinates. We provide two separate score matching variations for general use, an importance weighting (IW) approach, and a variational approach. We provide finite sample bounds for our IW approach in finite domain settings and show it to have especially strong performance in small sample lower dimensional cases. Complementing this, we show our variational approach to be strongest in more complex high-dimensional settings which we demonstrate on graphical model estimation tasks on both real and simulated data.

The peer review process in major artificial intelligence (AI) conferences faces unprecedented challenges with the surge of paper submissions (exceeding 10,000 submissions per venue), accompanied by growing concerns over review quality and reviewer responsibility. This position paper argues for **the need to transform the traditional one-way review system into a bi-directional feedback loop where authors evaluate review quality and reviewers earn formal accreditation, creating an accountability framework that promotes a sustainable, high-quality peer review system.** The current review system can be viewed as an interaction between three parties: the authors, reviewers, and system (i.e., conference), where we posit that all three parties share responsibility for the current problems. However, issues with authors can only be addressed through policy enforcement and detection tools, and ethical concerns can only be corrected through self-reflection. As such, this paper focuses on reforming reviewer accountability with systematic rewards through two key mechanisms: (1) a two-stage bi-directional review system that allows authors to evaluate reviews while minimizing retaliatory behavior, (2) a systematic reviewer reward system that incentivizes quality reviewing. We ask for the community's strong interest in these problems and the reforms that are needed to enhance the peer review process.

Process Reward Models (PRMs) have proven effective at enhancing mathematical reasoning for Large Language Models (LLMs) by leveraging increased inference-time computation. However, they are predominantly trained on mathematical data and their generalizability to non-mathematical domains has not been rigorously studied. In response, this work first shows that current PRMs have poor performance in other domains. To address this limitation, we introduce ***VersaPRM***, a multi-domain PRM trained on synthetic reasoning data generated using our novel data generation and annotation method. VersaPRM achieves consistent performance gains across diverse domains. For instance, in the MMLU-Pro category of Law, VersaPRM via weighted majority voting, achieves a 7.9% performance gain over the majority voting baseline–surpassing Qwen2.5-Math-PRM's gain of 1.3%. We further contribute to the community by open-sourcing all data, code and models for VersaPRM.

We analyze nonlinearly preconditioned gradient methods for solving smooth minimization problems. We introduce a generalized smoothness property, based on the notion of abstract convexity, that is broader than Lipschitz smoothness and provide sufficient first- and second-order conditions. Notably, our framework encapsulates algorithms associated with the gradient clipping method and brings out novel insights for the class of $(L_0,L_1)$-smooth functions that has received widespread interest recently, thus allowing us to extend beyond already established methods. We investigate the convergence of the proposed method in both the convex and nonconvex setting.

A modern paradigm for generalization in machine learning and AI consists of pre-training a task-agnostic foundation model, generally obtained using self-supervised and multimodal contrastive learning. The resulting representations can be used for prediction on a downstream task for which no labeled data is available. We present a theoretical framework to better understand this approach, called zero-shot prediction. We identify the target quantities that zero-shot prediction aims to learn, or learns in passing, and the key conditional independence relationships that enable its generalization ability.

Recently differential privacy has been used for a number of streaming, data structure, and dynamic graph problems as a means of hiding the internal randomness of the data structure, so that multiple possibly adaptive queries can be made without sacrificing the correctness of the responses. Although these works use differential privacy to show that for some problems it is possible to tolerate $T$ queries using $\widetilde{O}(\sqrt{T})$ copies of a data structure, such results only apply to {\it numerical estimation problems}, and only return the {\it cost} of an optimization problem rather than the solution itself. In this paper we investigate the use of differential privacy for adaptive queries to {\it search} problems, which are significantly more challenging since the responses to queries can reveal much more about the internal randomness than a single numerical query. We focus on two classical search problems: nearest neighbor queries and regression with arbitrary turnstile updates. We identify key parameters to these problems, such as the number of $c$-approximate near neighbors and the matrix condition number, and use different differential privacy techniques to design algorithms returning the solution point or solution vector with memory and time depending on these parameters. We give algorithms for each of these problems that achieve similar tradeoffs.

Realizing the vision of using AI agents to automate critical IT tasks depends on the ability to measure and understand effectiveness of proposed solutions. We introduce ITBench, a framework that offers a systematic methodology for benchmarking AI agents to address real-world IT automation tasks. Our initial release targets three key areas: Site Reliability Engineering (SRE), Compliance and Security Operations (CISO), and Financial Operations (FinOps). The design enables AI researchers to understand the challenges and opportunities of AI agents for IT automation with push-button workflows and interpretable metrics. IT-Bench includes an initial set of 102 real-world scenarios, which can be easily extended by community contributions. Our results show that agents powered by state-of-the-art models resolve only 11.4% of SRE scenarios, 25.2% of CISO scenarios, and 25.8% of FinOps scenarios (excluding anomaly detection). For FinOps-specific anomaly detection (AD) scenarios, AI agents achieve an F1 score of 0.35. We expect ITBench to be a key enabler of AI-driven IT automation that is correct, safe, and fast. IT-Bench, along with a leaderboard and sample agent implementations, is available at https://github.com/ibm/itbench.

Predictive models of the future are fundamental for an agent's ability to reason and plan. A common strategy learns a world model and unrolls it step-by-step at inference, where small errors can rapidly compound. Geometric Horizon Models (GHMs) offer a compelling alternative by directly making predictions of future states, avoiding cumulative inference errors. While GHMs can be conveniently learned by a generative analog to temporal difference (TD) learning, existing methods are negatively affected by bootstrapping predictions at train time and struggle to generate high-quality predictions at long horizons. This paper introduces Temporal Difference Flows (TD-Flow), which leverages the structure of a novel Bellman equation on probability paths alongside flow-matching techniques to learn accurate GHMs at over 5x the horizon length of prior methods. Theoretically, we establish a new convergence result and primarily attribute TD-Flow's efficacy to reduced gradient variance during training. We further show that similar arguments can be extended to diffusion-based methods. Empirically, we validate TD-Flow across a diverse set of domains on both generative metrics and downstream tasks, including policy evaluation. Moreover, integrating TD-Flow with recent behavior foundation models for planning over policies demonstrates substantial performance gains, underscoring its promise for long-horizon decision-making.

Large Language Model (LLM)-empowered multi-agent systems extend the cognitive boundaries of individual agents through disciplined collaboration and interaction, while constructing these systems often requires labor-intensive manual designs. Despite the availability of methods to automate the design of agentic workflows, they typically seek to identify a static, complex, one-size-fits-all system, which, however, fails to dynamically allocate inference resources based on the difficulty and domain of each query. To address this challenge, we shift away from the pursuit of a monolithic agentic system, instead optimizing the \textbf{agentic supernet}, a probabilistic and continuous distribution of agentic architectures. We introduce \textbf{MaAS}, an automated framework that samples query-dependent agentic systems from the supernet, delivering high-quality solutions and tailored resource allocation (\textit{e.g.}, LLM calls, tool calls, token cost). Comprehensive evaluation across six benchmarks demonstrates that MaAS \textbf{(I)} requires only $6\\sim45\\%$ of the inference costs of existing handcrafted or automated multi-agent systems, \textbf{(II)} surpasses them by $0.54\\%\sim11.82\\%$, and \textbf{(III)} enjoys superior cross-dataset and cross-LLM-backbone transferability.

We obtain an analytic, interpretable and predictive theory of creativity in convolutional diffusion models. Indeed, score-matching diffusion models can generate highly original images that lie far from their training data. However, optimal score-matching theory suggests that these models should only be able to produce memorized training examples. To reconcile this theory-experiment gap, we identify two simple inductive biases, locality and equivariance, that: (1) induce a form of combinatorial creativity by preventing optimal score-matching; (2) result in fully analytic, completely mechanistically interpretable, local score (LS) and equivariant local score (ELS) machines that, (3) after calibrating a single time-dependent hyperparameter can quantitatively predict the outputs of trained convolution only diffusion models (like ResNets and UNets) with high accuracy (median $r^2$ of $0.95, 0.94, 0.94, 0.96$ for our top model on CIFAR10, FashionMNIST, MNIST, and CelebA). Our model reveals a {\it locally consistent patch mosaic} mechanism of creativity, in which diffusion models create exponentially many novel images by mixing and matching different local training set patches at different scales and image locations. Our theory also partially predicts the outputs of pre-trained self-attention enabled UNets (median $r^2 \sim 0.77$ on CIFAR10), revealing an intriguing role for attention in carving out semantic coherence from local patch mosaics.