ICML 2025 Papers

Model merging seeks to integrate task-specific expert models into a unified architecture while preserving multi-task generalization capabilities, yet parameter interference between constituent models frequently induces performance degradation. Although prior work has explored many merging strategies, resolving interference without additional data for retraining or test-time computation remains challenging. In this paper, we theoretically demonstrate that the task vectors of the linear layer constitute an approximate linear subspace for its corresponding input. Therefore, we can minimize interference under the guidance of task vectors. Based on this insight, we propose **WUDI-Merging** (**W**hoever started the interference sho**U**ld en**D** **I**t), a simple yet effective model merging method that eliminates interference without any additional data or rescaling coefficients. Comprehensive empirical evaluations across vision and language benchmarks demonstrate our method's superiority, achieving state-of-the-art performance in data-free model merging scenarios (average 10.9% improvement versus baseline methods) while even outperforming mainstream test-time adaptation approaches by 3.3%, and only very few computing resources are required. The source code and implementation details are available at https://github.com/nathanielyvo/WUDI-Merging.

Pretrained transformers have demonstrated the ability to implement various algorithms at inference time without parameter updates. While theoretical works have established this capability through constructions and approximation guarantees, the optimization and statistical efficiency aspects remain understudied. In this work, we investigate how transformers learn features in-context -- a key mechanism underlying their inference-time adaptivity. We focus on the in-context learning of single-index models $y=\sigma_*(\langle \\boldsymbol{x},\\boldsymbol{\beta}\rangle)$, which are low-dimensional nonlinear functions parameterized by feature vector $\\boldsymbol\beta$. We prove that transformers pretrained by gradient-based optimization can perform *inference-time feature learning*, i.e., extract information of the target features $\\boldsymbol{\beta}$ solely from test prompts (despite $\\boldsymbol {\beta}$ varying across different prompts), hence achieving an in-context statistical efficiency that surpasses any non-adaptive (fixed-basis) algorithms such as kernel methods. Moreover, we show that the inference-time sample complexity surpasses the Correlational Statistical Query (CSQ) lower bound, owing to nonlinear label transformations naturally induced by the Softmax self-attention mechanism.

This paper addresses online learning with ''corrupted'' feedback. Our learner is provided with potentially corrupted gradients $\tilde g_t$ instead of the ''true'' gradients $g_t$. We make no assumptions about how the corruptions arise: they could be the result of outliers, mislabeled data, or even malicious interference. We focus on the difficult ``unconstrained'' setting in which our algorithm must maintain low regret with respect to any comparison point $u \in \mathbb{R}^d$. The unconstrained setting is significantly more challenging as existing algorithms suffer extremely high regret even with very tiny amounts of corruption (which is not true in the case of a bounded domain). Our algorithms guarantee regret $ \|u\|G (\sqrt{T} + k) $ when $G \ge \max_t \|g_t\|$ is known, where $k$ is a measure of the total amount of corruption. When $G$ is unknown we incur an extra additive penalty of $(\|u\|^2+G^2) k$.

This paper addresses the challenge of discontinuity in goal-achievement capabilities observed in Goal-conditioned Reinforcement Learning (GCRL) algorithms. Through a theoretical analysis, we identify that the reuse of successful trajectories or policies during training can aid in achieving adjacent goals of achievable goals. However, the policy discrepancy between achievable and adjacent goals must be carefully managed to avoid both overly trivial and excessively large differences, which can respectively hinder policy performance. To tackle this issue, we propose a margin-based policy self-regularization approach that optimizes the policy discrepancies between adjacent desired goals to a minimal acceptable threshold. This method can be integrated into popular GCRL algorithms, such as GC-SAC, HER, and GC-PPO. Systematic evaluations across two robotic arm control tasks and a complex fixed-wing aircraft control task demonstrate that our approach significantly improves the continuity of goal-achievement abilities of GCRL algorithms, thereby enhancing their overall performance.

The weighted Euclidean norm $||x||_w$ of a vector $x\in \mathbb{R}^d$ with weights $w\in \mathbb{R}^d$ is the Euclidean norm where the contribution of each dimension is scaled by a given weight. Approaches to dimensionality reduction that satisfy the Johnson–Lindenstrauss (JL) lemma can be easily adapted to the weighted Euclidean distance if weights are known and fixed: it suffices to scale each dimension of the input vectors according to the weights, and then apply any standard approach. However, this is not the case when weights are unknown during the dimensionality reduction or might dynamically change. In this paper, we address this issue by providing a linear function that maps vectors into a smaller complex vector space and allows to retrieve a JL-like estimate for the weighted Euclidean distance once weights are revealed. Our results are based on the decomposition of the complex dimensionality reduction into several Rademacher chaos random variables, which are studied using novel concentration inequalities for sums of independent Rademacher chaoses.

Reinforcement Learning with Human Feedback (RLHF) is a widely used fine-tuning approach that aligns machine learning models, particularly Language Models (LMs) with human preferences. There are typically multiple objectives driving the preference, hence humans find it easier to express per-objective comparisons rather than a global preference between two choices, e.g. compare two papers on their novelty, clarity, correctness, etc. Multi-Objective RLHF aims to use per-objective preference feedback and achieve a Pareto optimal tradeoff among these objectives by aggregating them into a single unified objective for optimization. However, nearly all prior works rely on linear aggregation, which rules out policies that favor specific objectives such as the worst one. The only existing approach using non-linear aggregation is computationally expensive due to its reward-based nature and the need for retraining whenever the aggregation parameters change. In this work, we address this limitation by transforming the non-linear aggregation maximization problem into a series of sub-problems. Each sub-problem involves only linear aggregation, making it computationally efficient to solve. We further extend our framework to handle multi-group scenarios, where each group has distinct weights for the objectives. Our method enables achieving consensus or maximizing the aggregated objective across all groups. Theoretically, we demonstrate that our algorithmic framework achieves sublinear regret and can be easily adapted to a reward-free algorithm. Empirically, leveraging our theoretical insights, we propose a nearly training-free algorithm once the optimal policies for individual objectives are obtained.

We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers from scratch, including understanding paper contributions, developing a codebase, and successfully executing experiments. For objective evaluation, we develop rubrics that hierarchically decompose each replication task into smaller sub-tasks with clear grading criteria. In total, PaperBench contains 8,316 individually gradable tasks. Rubrics are co-developed with the author(s) of each ICML paper for accuracy and realism. To enable scalable evaluation, we also develop an LLM-based judge to automatically grade replication attempts against rubrics, and assess our judge's performance by creating a separate benchmark for judges. We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline. We open-source our code (https://github.com/openai/preparedness) to facilitate future research in understanding the AI engineering capabilities of AI agents.

We propose a novel class of language models, Latent Thought Models (LTMs), which incorporate explicit latent thought vectors that follow an explicit prior model in latent space. These latent thought vectors guide the autoregressive generation of ground tokens through a Transformer decoder. Training employs a dual-rate optimization process within the classical variational Bayes framework: fast learning of local variational parameters for the posterior distribution of latent vectors (inference-time computation), and slow learning of global decoder parameters. Empirical studies reveal that LTMs possess additional scaling dimensions beyond traditional Large Language Models (LLMs), such as the number of iterations in inference-time computation and number of latent thought vectors. Higher sample efficiency can be achieved by increasing training compute per token, with further gains possible by trading model size for more inference steps. Designed based on these scaling properties, LTMs demonstrate superior sample and parameter efficiency compared to autoregressive models and discrete diffusion models. They significantly outperform these counterparts in validation perplexity and zero-shot language modeling tasks. Additionally, LTMs exhibit emergent few-shot in-context reasoning capabilities that scale with model size, and achieve competitive performance in conditional and unconditional text generation. The project page is available at [https://deqiankong.github.io/blogs/ltm](https://deqiankong.github.io/blogs/ltm).

Protein *sequence* design has seen significant advances through discrete diffusion and autoregressive approaches, yet the potential of continuous diffusion remains underexplored. Here, we present *DiMA*, a latent diffusion framework that operates on protein language model representations. Through systematic exploration of architectural choices and diffusion components, we develop a robust methodology that generalizes across multiple protein encoders ranging from 8M to 3B parameters. We demonstrate that our framework achieves consistently high performance across sequence-only (ESM-2, ESMc), dual-decodable (CHEAP), and multimodal (SaProt) representations using the same architecture and training approach. We conduct extensive evaluation of existing methods alongside *DiMA* using multiple metrics across two protein modalities, covering quality, diversity, novelty, and distribution matching of generated proteins. *DiMA* consistently produces novel, high-quality and diverse protein sequences and achieves strong results compared to baselines such as autoregressive, discrete diffusion and flow matching language models. The model demonstrates versatile functionality, supporting conditional generation tasks including protein family-generation, motif scaffolding and infilling, and fold-specific sequence design, despite being trained solely on sequence data. This work provides a universal continuous diffusion framework for protein sequence generation, offering both architectural insights and practical applicability across various protein design scenarios. Code is released at [GitHub](https://github.com/MeshchaninovViacheslav/DiMA).

With advancements in visual language models (VLMs) and large language models (LLMs), video anomaly detection (VAD) has progressed beyond binary classification to fine-grained categorization and multidimensional analysis. However, existing methods focus mainly on coarse-grained detection, lacking anomaly explanations. To address these challenges, we propose Ex-VAD, an Explainable Fine-grained Video Anomaly Detection approach that combines fine-grained classification with detailed explanations of anomalies. First, we use a VLM to extract frame-level captions, and an LLM converts them to video-level explanations, enhancing the model's explainability. Second, integrating textual explanations of anomalies with visual information greatly enhances the model's anomaly detection capability. Finally, we apply label-enhanced alignment to optimize feature fusion, enabling precise fine-grained detection. Extensive experimental results on the UCF-Crime and XD-Violence datasets demonstrate that Ex-VAD significantly outperforms existing State-of-The-Art methods.

Multi-kernel clustering (MKC) has emerged as a powerful method for capturing diverse data patterns, offering robust and generalized representations of data structures. However, the increasing deployment of MKC in real-world applications raises concerns about its vulnerability to adversarial perturbations. While adversarial robustness has been extensively studied in other domains, its impact on MKC remains largely unexplored. In this paper, we address the challenge of assessing the adversarial robustness of MKC methods in a black-box setting. Specifically, we propose *AdvMKC*, a novel reinforcement-learning-based adversarial attack framework designed to inject imperceptible perturbations into data and mislead MKC methods. AdvMKC leverages proximal policy optimization with an advantage function to overcome the instability of clustering results during optimization. Additionally, it introduces a generator-clusterer framework, where a generator produces adversarial perturbations, and a clusterer approximates MKC behavior, significantly reducing computational overhead. We provide theoretical insights into the impact of adversarial perturbations on MKC and validate these findings through experiments. Evaluations across seven datasets and eleven MKC methods (seven traditional and four robust) demonstrate AdvMKC's effectiveness, robustness, and transferability.

Trajectory prediction for surrounding agents is a challenging task in autonomous driving due to its inherent uncertainty and underlying multimodality. Unlike prevailing data-driven methods that primarily rely on supervised learning, in this paper, we introduce a novel **G**raph-**o**riented **I**nverse **R**einforcement **L**earning (GoIRL) framework, which is an IRL-based predictor equipped with vectorized context representations. We develop a feature adaptor to effectively aggregate lane-graph features into grid space, enabling seamless integration with the maximum entropy IRL paradigm to infer the reward distribution and obtain the policy that can be sampled to induce multiple plausible plans. Furthermore, conditioned on the sampled plans, we implement a hierarchical parameterized trajectory generator with a refinement module to enhance prediction accuracy and a probability fusion strategy to boost prediction confidence. Extensive experimental results showcase our approach not only achieves state-of-the-art performance on the large-scale Argoverse & nuScenes motion forecasting benchmarks but also exhibits superior generalization abilities compared to existing supervised models.

This paper studies phase transitions for the existence of unregularized M-estimators under proportional asymptotics where the sample size $n$ and feature dimension $p$ grow proportionally with $n/p \to \delta \in (1, \infty)$. We study the existence of M-estimators in single-index models where the response $y_i$ depends on covariates $x_i \sim N(0, I_p)$ through an unknown index ${w} \in \mathbb{R}^p$ and an unknown link function. An explicit expression is derived for the critical threshold $\delta_\infty$ that determines the phase transition for the existence of the M-estimator, generalizing the results of Candés & Sur (2020) for binary logistic regression to other single-index models. Furthermore, we investigate the existence of a solution to the nonlinear system of equations governing the asymptotic behavior of the M-estimator when it exists. The existence of solution to this system for $\delta > \delta_\infty$ remains largely unproven outside the global null in binary logistic regression. We address this gap with a proof that the system admits a solution if and only if $\delta > \delta_\infty$, providing a comprehensive theoretical foundation for proportional asymptotic results that require as a prerequisite the existence of a solution to the system.

We empirically investigate how well popular approximate inference algorithms for Bayesian Neural Networks (BNNs) respect the theoretical properties of Bayesian belief updating. We find strong evidence on synthetic regression and real-world image classification tasks that common BNN algorithms such as variational inference, Laplace approximation, SWAG, and SGLD fail to update in a consistent manner, forget about old data under sequential updates, and violate the predictive coherence properties that would be expected of Bayesian methods. These observed behaviors imply that care should be taken when treating BNNs as true Bayesian models, particularly when using them beyond static prediction settings, such as for active, continual, or transfer learning.

Orientation is a fundamental attribute of objects, essential for understanding their spatial pose and arrangement. However, practical solutions for estimating the orientation of open-world objects in monocular images remain underexplored. In this work, we introduce Orient Anything, the first foundation model for zero-shot object orientation estimation. A key challenge in this task is the scarcity of orientation annotations for open-world objects. To address this, we propose leveraging the vast resources of 3D models. By developing a pipeline to annotate the front face of 3D objects and render them from random viewpoints, we curate 2 million images with precise orientation annotations across a wide variety of object categories. To fully leverage the dataset, we design a robust training objective that models the 3D orientation as probability distributions over three angles and predicts the object orientation by fitting these distributions. Besides, we propose several strategies to further enhance the synthetic-to-real transfer. Our model achieves state-of-the-art orientation estimation accuracy on both rendered and real images, demonstrating impressive zero-shot capabilities across various scenarios. Furthermore, it shows great potential in enhancing high-level applications, such as understanding complex spatial concepts in images and adjusting 3D object pose.

In *fully-dynamic consistent clustering*, we are given a finite metric space $(M,d)$, and a set $F\subseteq M$ of possible locations for opening centers. Data points arrive and depart, and the goal is to maintain an approximately optimal clustering solution at all times while minimizing the *recourse*, the total number of additions/deletions of centers over time. Specifically, we study fully dynamic versions of the classical $k$-center, facility location, and $k$-median problems. We design algorithms that, given a parameter $\beta\geq 1$, maintain an $O(\beta)$-approximate solution at all times, and whose total recourse is bounded by $O(\log |F| \log \Delta) \cdot OPT_{rec}^{\beta}$. Here $OPT_{rec}^{\beta}$ is the minimal recourse of an offline algorithm that maintains a $\beta$-approximate solution at all times, and $\Delta$ is the metric aspect ratio. We obtain our results via a reduction to the recently proposed *Positive Body Chasing* framework of [Bhattacharya Buchbinder Levin Saranurak, FOCS 2023], which we show gives fractional solutions to our clustering problems online. Our contribution is to round these fractional solutions while preserving the approximation and recourse guarantees. We complement our positive results with logarithmic lower bounds which show that our bounds are nearly tight.

Machine unlearning presents a promising approach to mitigating privacy and safety concerns in large language models (LLMs) by enabling the selective removal of targeted data or knowledge while preserving model utility. However, existing unlearning methods remain over-sensitive to downstream fine-tuning, which can rapidly recover what is supposed to be unlearned information even when the fine-tuning task is entirely unrelated to the unlearning objective. To enhance robustness, we introduce the concept of `invariance' into unlearning for the first time from the perspective of invariant risk minimization (IRM), a principle for environment-agnostic training. By leveraging IRM, we develop a new invariance-regularized LLM unlearning framework, termed invariant LLM unlearning (ILU). We show that the proposed invariance regularization, even using only a single fine-tuning dataset during ILU training, can enable unlearning robustness to generalize effectively across diverse and new fine-tuning tasks at test time. A task vector analysis is also provided to further elucidate the rationale behind ILU's effectiveness. Extensive experiments on the WMDP benchmark, which focuses on removing an LLM's hazardous knowledge generation capabilities, reveal that ILU significantly outperforms state-of-the-art unlearning methods, including negative preference optimization (NPO) and representation misdirection for unlearning (RMU). Notably, ILU achieves superior unlearning robustness across diverse downstream fine-tuning scenarios (e.g., math, paraphrase detection, and sentiment analysis) while preserving the fine-tuning performance.

Large language models (LLMs) consist of numerous Transformer modules, and while the models can perform various functions, it remains an open question of how these modules are combined to elicit distinct inherent functionalities. In this paper, we investigate the modules inside LLMs and demonstrate that, by simply masking or retaining specific attention heads during inference, LLMs can exhibit specific task functionalities without requiring explicit instructions or modifications to the model parameters. Experiments across various models and tasks reveal that LLMs inherently encode ``functional pathways'', the structured groups of interdependent attention heads that are crucial for executing specific tasks. These pathways not only govern the model's functional behaviors but also enhance parameter efficiency, as suppressing attention heads outside the pathway can improve task performance. The code is available in this repository: [https://github.com/OpenDFM/HeadsUp](https://github.com/OpenDFM/HeadsUp).

Motivated by the first priority of safety in many real-world applications, we propose \textsc{MaxSafe}, a chance-constrained bi-level optimization framework for safe reinforcement learning. \textsc{MaxSafe} first minimizes the unsafe probability and then maximizes the return among the safest policies. We provide a tailored Q-learning algorithm for the \textsc{MaxSafe} objective, featuring a novel learning process for \emph{optimal action masks} with theoretical convergence guarantees. To enable the application of our algorithm to large-scale experiments, we introduce two key techniques: \emph{safety polarization} and \emph{safety prioritized experience replay}. Safety polarization generalizes the optimal action masking by polarizing the Q-function, which assigns low values to unsafe state-action pairs, effectively discouraging their selection. In parallel, safety prioritized experience replay enhances the learning of optimal action masks by prioritizing samples based on temporal-difference (TD) errors derived from our proposed state-action reachability estimation functions. This approach efficiently addresses the challenges posed by sparse cost signals. Experiments on diverse autonomous driving and safe control tasks show that our methods achieve near-maximal safety and an optimal reward-safety trade-off.