ICML 2025 Papers

Most learning algorithms with formal regret guarantees assume that all mistakes are recoverable and essentially rely on trying all possible behaviors. This approach is problematic when some mistakes are *catastrophic*, i.e., irreparable. We propose an online learning problem where the goal is to minimize the chance of catastrophe. Specifically, we assume that the payoff in each round represents the chance of avoiding catastrophe in that round and try to maximize the product of payoffs (the overall chance of avoiding catastrophe) while allowing a limited number of queries to a mentor. We also assume that the agent can transfer knowledge between similar inputs. We first show that in general, any algorithm either queries the mentor at a linear rate or is nearly guaranteed to cause catastrophe. However, in settings where the mentor policy class is learnable in the standard online model, we provide an algorithm whose regret and rate of querying the mentor both approach 0 as the time horizon grows. Although our focus is the product of payoffs, we provide matching bounds for the typical additive regret. Conceptually, if a policy class is learnable in the absence of catastrophic risk, it is learnable in the presence of catastrophic risk if the agent can ask for help.

Maximizing a single submodular set function subject to a cardinality constraint is a well-studied and central topic in combinatorial optimization. However, finding a set that maximizes multiple functions at the same time is much less understood, even though it is a formulation which naturally occurs in robust maximization or problems with fairness considerations such as fair influence maximization or fair allocation. In this work, we consider the problem of maximizing the minimum over many submodular functions subject to a cardinality constraint, which is known as multiobjective submodular maximization. All known polynomial-time approximation algorithms either obtain a weak approximation guarantee or rely on the evaluation of the multilinear extension. The latter is expensive to evaluate and renders such algorithms impractical. We bridge this gap and introduce the first scalable and practical algorithm that obtains the best-known approximation guarantee. We furthermore introduce a novel application fair centrality maximization and show how it can be addressed via multiobjective submodular maximization. In our experimental evaluation, we show that our algorithm outperforms known algorithms in terms of objective value and running time.

A substantial body of work in machine learning (ML) and randomized numerical linear algebra (RandNLA) has exploited various sorts of random sketching methodologies, including random sampling and random projection, with much of the analysis using Johnson--Lindenstrauss and subspace embedding techniques. Recent studies have identified the issue of *inversion bias* -- the phenomenon that inverses of random sketches are *not* unbiased, despite the unbiasedness of the sketches themselves. This bias presents challenges for the use of random sketches in various ML pipelines, such as fast stochastic optimization, scalable statistical estimators, and distributed optimization. In the context of random projection, the inversion bias can be easily corrected for dense Gaussian projections (which are, however, too expensive for many applications). Recent work has shown how the inversion bias can be corrected for sparse sub-gaussian projections. In this paper, we show how the inversion bias can be corrected for random sampling methods, both uniform and non-uniform leverage-based, as well as for structured random projections, including those based on the Hadamard transform. Using these results, we establish problem-independent local convergence rates for sub-sampled Newton methods.

Recent theories suggest that Neural Scaling Laws arise whenever the task is linearly decomposed into units that are power-law distributed. Alternatively, scaling laws also emerge when data exhibit a hierarchically compositional structure, as is thought to occur in language and images. To unify these views, we consider classification and next-token prediction tasks based on probabilistic context-free grammars—probabilistic models that generate data via a hierarchy of production rules. For classification, we show that having power-law distributed production rules results in a power-law learning curve with an exponent depending on the rules’ distribution and a large multiplicative constant that depends on the hierarchical structure. By contrast, for next-token prediction, the distribution of production rules controls the fine details of the learning curve, but not the exponent describing the large-scale behaviour.

Large-language models (LLMs) and chatbot agents are known to provide wrong outputs at times, and it was recently found that this can never be fully prevented. Hence, uncertainty quantification plays a crucial role, aiming to quantify the level of ambiguity in either one overall number or two numbers for aleatoric and epistemic uncertainty. This position paper argues that this traditional dichotomy of uncertainties is too limited for the open and interactive setup that LLM agents operate in when communicating with a user, and that we need to research avenues that enrich uncertainties in this novel scenario. We review the literature and find that popular definitions of aleatoric and epistemic uncertainties directly contradict each other and lose their meaning in interactive LLM agent settings. Hence, we propose three novel research directions that focus on uncertainties in such human-computer interactions: Underspecification uncertainties, for when users do not provide all information or define the exact task at the first go, interactive learning, to ask follow-up questions and reduce the uncertainty about the current context, and output uncertainties, to utilize the rich language and speech space to express uncertainties as more than mere numbers. We expect that these new ways of dealing with and communicating uncertainties will lead to LLM agent interactions that are more transparent, trustworthy, and intuitive.

Recent advancements in large language model (LLM)-based agents have demonstrated that collective intelligence can significantly surpass the capabilities of individual agents, primarily due to well-crafted inter-agent communication topologies. Despite the diverse and high-performing designs available, practitioners often face confusion when selecting the most effective pipeline for their specific task: \textit{Which topology is the best choice for my task, avoiding unnecessary communication token overhead while ensuring high-quality solution?} In response to this dilemma, we introduce G-Designer, an adaptive, efficient, and robust solution for multi-agent deployment, which dynamically designs task-aware, customized communication topologies. Specifically, G-Designer models the multi-agent system as a multi-agent network, leveraging a variational graph auto-encoder to encode both the nodes (agents) and a task-specific virtual node, and decodes a task-adaptive and high-performing communication topology. Extensive experiments on six benchmarks showcase that G-Designer is: \textbf{(1) high-performing}, achieving superior results on MMLU with accuracy at $84.50\\%$ and on HumanEval with pass@1 at $89.90\\%$; \textbf{(2) task-adaptive}, architecting communication protocols tailored to task difficulty, reducing token consumption by up to $95.33\\%$ on HumanEval; and \textbf{(3) adversarially robust}, defending against agent adversarial attacks with merely $0.3\\%$ accuracy drop.

We propose a data-driven method for reducing the dimensionality of linear programming problems (LPs) by generating instance-specific projection matrices using a neural network-based model. Once the model is trained using multiple LPs by maximizing the expected objective value, we can efficiently find high-quality feasible solutions of newly given LPs. Our method can shorten the computational time of any LP solvers due to its solver-agnostic nature, it can provide feasible solutions by relying on projection that reduces the number of variables, and it can handle LPs of different sizes using neural networks with permutation equivariance and invariance. We also provide a theoretical analysis of the generalization bound for learning a neural network to generate projection matrices that reduce the size of LPs. Our experimental results demonstrate that our method can obtain solutions with higher quality than the existing methods, while its computational time is significantly shorter than solving the original LPs.

Reinforcement Learning with Human Feedback (RLHF) and its variants have made huge strides toward the effective alignment of large language models (LLMs) to follow instructions and reflect human values. More recently, Direct Alignment Algorithms (DAAs) have emerged in which the reward modeling stage of RLHF is skipped by characterizing the reward directly as a function of the policy being learned. Some popular examples of DAAs include Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO). These methods often suffer from likelihood displacement, a phenomenon by which the probabilities of preferred responses are often reduced undesirably. In this paper, we argue that, for DAAs the reward (function) shape matters. We introduce \textbf{AlphaPO}, a new DAA method that leverages an $\alpha$-parameter to help change the shape of the reward function beyond the standard log reward. AlphaPO helps maintain fine-grained control over likelihood displacement and over-optimization. Compared to SimPO, one of the best performing DAAs, AlphaPO leads to about 7\% to 10\% relative improvement in alignment performance for the instruct versions of Mistral-7B and Llama3-8B while achieving 15\% to 50\% relative improvement over DPO on the same models. The analysis and results presented highlight the importance of the reward shape and how one can systematically change it to affect training dynamics, as well as improve alignment performance.

We study the algorithmic problem of robust mean estimation of an identity covariance Gaussian in the presence of mean-shift contamination. In this contamination model, we are given a set of points in $\mathbb{R}^d$ generated i.i.d. via the following process. For a parameter $\alpha<1/2$, the $i$-th sample $x_i$ is obtained as follows: with probability $1-\alpha$, $x_i$ is drawn from $\mathcal{N}(\mu, I)$, where $\mu \in \mathbb{R}^d$ is the target mean; and with probability $\alpha$, $x_i$ is drawn from $\mathcal{N}(z_i, I)$, where $z_i$ is unknown and potentially arbitrary. Prior work characterized the information-theoretic limits of this task. Specifically, it was shown that— in contrast to Huber contamination— in the presence of mean-shift contamination consistent estimation is possible. On the other hand, all known robust estimators in the mean-shift model have running times exponential in the dimension. Here we give the first computationally efficient algorithm for high-dimensional robust mean estimation with mean-shift contamination that can tolerate a constant fraction of outliers. In particular, our algorithm has near-optimal sample complexity, runs in sample-polynomial time, and approximates the target mean to any desired accuracy. Conceptually, our result contributes to a growing body of work that studies inference with respect to natural noise models lying in between fully adversarial and random settings.

Optimal prompt selection is crucial for maximizing large language model (LLM) performance on downstream tasks, especially in black-box settings where models are only accessible via APIs. Black-box prompt selection is challenging due to potentially large, combinatorial search spaces, absence of gradient information, and high evaluation cost of prompts on a validation set. We propose HbBoPs, a novel method that combines a structural-aware deep kernel Gaussian Process with Hyperband as a multi-fidelity scheduler to efficiently select prompts. HbBoPs uses embeddings of instructions and few-shot exemplars, treating them as modular components within prompts. This enhances the surrogate model's ability to predict which prompt to evaluate next in a sample-efficient manner. Hyperband improves query-efficiency by adaptively allocating resources across different fidelity levels, reducing the number of validation instances required for evaluating prompts. Extensive experiments across ten diverse benchmarks and three LLMs demonstrate that HbBoPs outperforms state-of-the-art methods in both performance and efficiency.

Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, are commonly used to adapt LLMs. However, the effectiveness of standard PEFT methods is limited in low-resource scenarios with only a few hundred examples. Recent advances in interpretability research have inspired the emergence of activation editing (or steering) techniques, which modify the activations of specific model components. Due to their extremely small parameter counts, these methods show promise for small datasets. However, their performance is highly dependent on identifying the correct modules to edit and often lacks stability across different datasets. In this paper, we propose Joint Localization and Activation Editing (JoLA), a method that jointly learns (1) which heads in the Transformer to edit (2) whether the intervention should be additive, multiplicative, or both and (3) the intervention parameters themselves - the vectors applied as additive offsets or multiplicative scalings to the head output. Through evaluations on three benchmarks spanning commonsense reasoning, natural language understanding, and natural language generation, we demonstrate that JoLA consistently outperforms existing methods.

The remarkable generalization performance of contrastive vision-language models like CLIP is often attributed to the diversity of their training distributions. However, key questions remain unanswered: Can CLIP generalize to an entirely unseen domain when trained on a diverse mixture of domains (domain generalization)? Can it generalize to unseen classes within partially seen domains (compositional generalization)? What factors affect such generalization? To answer these questions, we trained CLIP models on systematically constructed training distributions with controlled domain diversity and object class exposure. Our experiments show that domain diversity is essential for both domain and compositional generalization, yet compositional generalization can be surprisingly weaker than domain generalization when the training distribution contains a suboptimal subset of the test domain. Through data-centric *and* mechanistic analyses, we find that successful generalization requires the learning of sufficiently shared representations in intermediate layers and circuits.

Cross-Domain Few-Shot Segmentation (CD-FSS) aims to transfer knowledge from a large-scale source-domain dataset to unseen target-domain datasets with limited annotated samples. Current methods typically compare the distance between training and testing samples for mask prediction. However, a problem of feature entanglement exists in this well-adopted method, which binds multiple patterns together and harms the transferability. However, we find an entanglement problem exists in this widely adopted method, which tends to bind source-domain patterns together and make each of them hard to transfer. In this paper, we aim to address this problem for the CD-FSS task. We first find a natural decomposition of the ViT structure, based on which we delve into the entanglement problem for an interpretation. We find the decomposed ViT components are crossly compared between images in distance calculation, where the rational comparisons are entangled with those meaningless ones by their equal importance, leading to the entanglement problem. Based on this interpretation, we further propose to address the entanglement problem by learning to weigh for all comparisons of ViT components, which learn disentangled features and re-compose them for the CD-FSS task, benefiting both the generalization and finetuning. Experiments show that our model outperforms the state-of-the-art CD-FSS method by 1.92% and 1.88% in average accuracy under 1-shot and 5-shot settings, respectively.

Benchmarks and evaluations are central to machine learning methodology and direct research in the field. Current evaluations commonly test systems in the absence of humans. This position paper argues that the machine learning community should increasingly use _centaur evaluations_, in which humans and AI jointly solve tasks. Centaur Evaluations refocus machine learning development toward human augmentation instead of human replacement, they allow for direct evaluation of human-centered desiderata, such as interpretability and helpfulness, and they can be more challenging and realistic than existing evaluations. By shifting the focus from _automation_ toward _collaboration_ between humans and AI, centaur evaluations can drive progress toward more effective and human-augmenting machine learning systems.

Previous studies have extensively addressed the attribution problem for binary outcome variables. However, in many practical scenarios, the outcome variable is continuous, and simply binarizing it may result in information loss or biased conclusions. To address this issue, we propose a series of posterior causal estimands for retrospectively evaluating multiple correlated causes from a continuous outcome. These estimands include posterior intervention effects, posterior total causal effects, and posterior natural direct effects. Under assumptions of sequential ignorability, monotonicity, and perfect positive rank, we show that the posterior causal estimands of interest are identifiable and present the corresponding identification equations. We also provide a simple but effective estimation procedure and establish asymptotic properties of the proposed estimators. An artificial hypertension example and a real developmental toxicity dataset are employed to illustrate our method.

In self-supervised contrastive learning, negative pairs are typically constructed using an anchor image and a sample drawn from the entire dataset, excluding the anchor. However, this approach can result in the creation of negative pairs with similar semantics, referred to as "false negatives", leading to their embeddings being falsely pushed apart. To address this issue, we introduce *GloFND*, an optimization-based approach that automatically learns on the fly the threshold for each anchor data to *identify* its false negatives during training. In contrast to previous methods for false negative discovery, our approach *globally* detects false negatives across the entire dataset rather than locally within the mini-batch. Moreover, its per-iteration computation cost remains independent of the dataset size. Experimental results on image and image-text data demonstrate the effectiveness of the proposed method. Our implementation is available at https://github.com/vibalcam/GloFND.

Recent theoretical results demonstrate that the convergence rates of permutation-based SGD (e.g., random reshuffling SGD) are faster than uniform-sampling SGD; however, these studies focus mainly on the large epoch regime, where the number of epochs $K$ exceeds the condition number $\kappa$. In contrast, little is known when $K$ is smaller than $\kappa$, and it is still a challenging open question whether permutation-based SGD can converge faster in this small epoch regime (Safran and Shamir, 2021). As a step toward understanding this gap, we study the naive deterministic variant, Incremental Gradient Descent (IGD), on smooth and strongly convex functions. Our lower bounds reveal that for the small epoch regime, IGD can exhibit surprisingly slow convergence even when all component functions are strongly convex. Furthermore, when some component functions are allowed to be nonconvex, we prove that the optimality gap of IGD can be significantly worse throughout the small epoch regime. Our analyses reveal that the convergence properties of permutation-based SGD in the small epoch regime may vary drastically depending on the assumptions on component functions. Lastly, we supplement the paper with tight upper and lower bounds for IGD in the large epoch regime.

In this paper, we seek to develop a versatile test-time adaptation (TTA) objective for a variety of tasks — classification and regression across image-, object-, and pixel-level predictions. We achieve this through a self-bootstrapping scheme that optimizes prediction consistency between the test image (as target) and its deteriorated view. The key challenge lies in devising effective augmentations/deteriorations that: i) preserve the image’s geometric information, e.g., object sizes and locations, which is crucial for TTA on object/pixel-level tasks, and ii) provide sufficient learning signals for TTA. To this end, we analyze how common distribution shifts affect the image's information power across spatial frequencies in the Fourier domain, and reveal that low-frequency components carry high power and masking these components supplies more learning signals, while masking high-frequency components can not. In light of this, we randomly mask the low-frequency amplitude of an image in its Fourier domain for augmentation. Meanwhile, we also augment the image with noise injection to compensate for missing learning signals at high frequencies, by enhancing the information power there. Experiments show that, either independently or as a plug-and-play module, our method achieves superior results across classification, segmentation, and 3D monocular detection tasks with both transformer and CNN models.

As large language models (LMs) advance, there is an increasing need to control their outputs to align with human values (e.g., detoxification) or desired attributes (e.g., personalization, topic). However, autoregressive models focus on next-token predictions and struggle with global properties that require looking ahead. Existing solutions either post-train LMs for each new attribute—expensive and inflexible—or approximate the Expected Attribute Probability (EAP) of future sequences by sampling or training, which is slow and unreliable for rare attributes. We introduce **TRACE** (Tractable Probabilistic Reasoning for Adaptable Controllable gEneration), a novel framework that efficiently computes EAP and adapts to new attributes through tractable *probabilistic* reasoning and lightweight *control*. TRACE distills a Hidden Markov Model (HMM) from an LM and pairs it with a small classifier to estimate attribute probabilities, enabling exact EAP computation over the HMM’s predicted futures. This EAP is then used to reweigh the LM’s next-token probabilities for globally compliant continuations. Empirically, TRACE achieves state-of-the-art detoxification results with only 20% decoding overhead, yields 76 low-resource personalized LMs within seconds, and seamlessly extends to composite attributes.