ICML 2025 论文列表

While existing benchmarks probe the reasoning abilities of large language models (LLMs) across diverse domains, they predominantly assess passive reasoning, providing models with all the information needed to reach a solution. By contrast, active reasoning—where an LLM must interact with external systems to acquire missing evidence or data—has received little systematic attention. To address this shortfall, we present AR-Bench, a novel benchmark designed explicitly to evaluate an LLM’s active reasoning skills. AR-Bench comprises three task families—detective cases, situation puzzles, and guessing numbers—that together simulate real-world, agentic scenarios and measure performance across commonsense, logical, and symbolic reasoning challenges. Empirical evaluation on AR-Bench demonstrates that contemporary LLMs exhibit pronounced difficulties with active reasoning: they frequently fail to acquire or leverage the information needed to solve tasks. This gap highlights a stark divergence between their passive and active reasoning abilities. Moreover, ablation studies indicate that even advanced strategies, such as tree-based searching or post-training approaches, yield only modest gains and fall short of the levels required for real-world deployment. Collectively, these findings highlight the critical need to advance methodology for active reasoning, e.g., incorporating interactive learning, real-time feedback loops, and environment-aware objectives for training. The benchmark is publicly available at: https://github.com/tmlr-group/AR-Bench.

We revisit the problem of selecting $k$-out-of-$n$ elements with the goal of optimizing an objective function, and ask whether it can be solved approximately with sublinear query complexity. For objective functions that are monotone submodular, [Li, Feldman, Kazemi, Karbasi, NeurIPS'22; Kuhnle, AISTATS'21] gave an $\Omega(n/k)$ query lower bound for approximating to within any constant factor. We strengthen their lower bound to a nearly tight $\tilde{\Omega}(n)$. This lower bound holds even for estimating the value of the optimal subset. When the objective function is additive, we prove that finding an approximately optimal subset still requires near-linear query complexity, but we can estimate the value of the optimal subset in $\tilde{O}(n/k)$ queries, and that this is tight up to polylog factors.

The next-coordinate prediction paradigm has emerged as the de facto standard in current auto-regressive mesh generation methods. Despite their effectiveness, there is no efficient measurement for the various tokenizers that serialize meshes into sequences. In this paper, we introduce a new metric Per-Token-Mesh-Entropy (PTME) to evaluate the existing mesh tokenizers theoretically without any training. Building upon PTME, we propose a plug-and-play tokenization technique called coordinate merging. It further improves the compression ratios of existing tokenizers by rearranging and merging the most frequent patterns of coordinates. Through experiments on various tokenization methods like MeshXL, MeshAnything V2, and Edgerunner, we further validate the performance of our method. We hope that the proposed PTME and coordinate merging can enhance the existing mesh tokenizers and guide the further development of native mesh generation.

The selective classifier (SC) has been proposed for rank based uncertainty thresholding, which could have applications in safety critical areas such as medical diagnostics, autonomous driving, and the justice system. The Area Under the Risk-Coverage Curve (AURC) has emerged as the foremost evaluation metric for assessing the performance of SC systems. In this work, we present a formal statistical formulation of population AURC, presenting an equivalent expression that can be interpreted as a reweighted risk function. Through Monte Carlo methods, we derive empirical AURC plug-in estimators for finite sample scenarios. The weight estimators associated with these plug-in estimators are shown to be consistent, with low bias and tightly bounded mean squared error (MSE). The plug-in estimators are proven to converge at a rate of $\mathcal{O}(\sqrt{\ln(n)/n})$ demonstrating statistical consistency. We empirically validate the effectiveness of our estimators through experiments across multiple datasets, model architectures, and confidence score functions (CSFs), demonstrating consistency and effectiveness in fine-tuning AURC performance.

Medical image segmentation models often struggle to generalize across different domains due to various reasons. Domain Generalization (DG) methods overcome this either through representation learning or data augmentation (DA). While representation learning methods seek domain-invariant features, they often rely on ad-hoc techniques and lack formal guarantees. DA methods, which enrich model representations through synthetic samples, have shown comparable or superior performance to representation learning approaches. We propose LangDAug, a novel **Lang**evin **D**ata **Aug**mentation for multi-source domain generalization in 2D medical image segmentation. LangDAug leverages Energy-Based Models (EBMs) trained via contrastive divergence to traverse between source domains, generating intermediate samples through Langevin dynamics. Theoretical analysis shows that LangDAug induces a regularization effect, and for GLMs, it upper-bounds the Rademacher complexity by the intrinsic dimensionality of the data manifold. Through extensive experiments on Fundus segmentation and 2D MRI prostate segmentation benchmarks, we show that LangDAug outperforms state-of-the-art domain generalization methods and effectively complements existing domain-randomization approaches. The codebase for our method is available at https://github.com/backpropagator/LangDAug.

Diversity has been one of the most crucial factors on the design of adversarial ensemble methods. This work focuses on the fundamental problems: How to define the diversity for the adversarial ensemble, and how to correlate with algorithmic performance. We first show that it is an NP-Hard problem to precisely calculate the diversity of two networks in adversarial ensemble learning, which makes it different from prior diversity analysis. We present the first diversity decomposition under the first-order approximation for the adversarial ensemble learning. Specifically, the adversarial ensemble loss can be decomposed into average of individual adversarial losses, gradient diversity, prediction diversity and cross diversity. Hence, it is not sufficient to merely consider the gradient diversity on the characterization of diversity as in previous adversarial ensemble methods. We present diversity decomposition for classification with cross-entropy loss similarly. Based on the theoretical analysis, we develop new ensemble method via orthogonal adversarial predictions to simultaneously improve gradient diversity and cross diversity. We finally conduct experiments to validate the effectiveness of our method.

While direct policy optimization methods exist, pioneering LLMs are fine-tuned with reinforcement learning from human feedback (RLHF) to generate better responses under the supervision of a reward model learned from preference data. One major challenge of RLHF is the inaccuracy of the intermediate reward model, especially in the tasks that requires complex reasoning for the reward model to score a response. We find that the reliability of the reward model varies across responses assigned with different rewards. This motivates us to filter the samples whose rewards may be unreliable to improve the signal-to-noise ratio during policy learning, resulting in Policy Filtration for Proximal Policy Optimization (PF-PPO). To choose a proper policy filtering strategy, we use the coefficient of determination ($R^2$) between the rewards and actual scores on filtered samples as the metrics to help us find promising strategies since it measures how well the rewards filtered by PF-PPO indicate real performance. We provide extensive experiments to validate the effectiveness of PF-PPO in code generation and math reasoning tasks. In code generation, PF-PPO achieves the state-of-the-art performance of 7-billion-parameter models on HumanEval (+7.9%), MBPP (+0.7%), and LeetCode Contest (+10.0%) which is a more challenging benchmark created by us. In math reasoning, PF-PPO yields performance increase using different reward models and benchmarks (Ape210K and CMATH).

Self-supervised learning (SSL) allows training data representations without a supervised signal and has become an important paradigm in machine learning. Most SSL methods employ the cosine similarity between embedding vectors and hence effectively embed data on a hypersphere. While this seemingly implies that embedding norms cannot play any role in SSL, a few recent works have suggested that embedding norms have properties related to network convergence and confidence. In this paper, we resolve this apparent contradiction and systematically establish the embedding norm's role in SSL training. Using theoretical analysis, simulations, and experiments, we show that embedding norms (i) govern SSL convergence rates and (ii) encode network confidence, with smaller norms corresponding to unexpected samples. Additionally, we show that manipulating embedding norms can have large effects on convergence speed. Our findings demonstrate that SSL embedding norms are integral to understanding and optimizing network behavior.

Strategic classification addresses a learning problem where a decision-maker implements a classifier over agents who may manipulate their features in order to receive favorable predictions. In the standard model of online strategic classification, in each round, the decision-maker implements and publicly reveals a classifier, after which agents perfectly best respond based on this knowledge. However, in practice, whether to disclose the classifier is often debated---some decision-makers believe that hiding the classifier can prevent misclassification errors caused by manipulation. In this paper, we formally examine how limiting the agents' access to the current classifier affects the decision-maker's performance. Specifically, we consider an extended online strategic classification setting where agents lack direct knowledge about the current classifier and instead manipulate based on a weighted average of historically implemented classifiers. Our main result shows that in this setting, the decision-maker incurs $(1-\gamma)^{-1}$ or $k_{\text{in}}$ times more mistakes compared to the full-knowledge setting, where $k_{\text{in}}$ is the maximum in-degree of the manipulation graph (representing how many distinct feature vectors can be manipulated to appear as a single one), and $\gamma$ is the discount factor indicating agents' memory of past classifiers. Our results demonstrate how withholding access to the classifier can backfire and degrade the decision-maker's performance in online strategic classification.

Few-shot image classification remains challenging due to the scarcity of labeled training examples. Augmenting them with synthetic data has emerged as a promising way to alleviate this issue, but models trained on synthetic samples often face performance degradation due to the inherent gap between real and synthetic distributions. To address this limitation, we develop a theoretical framework that quantifies the impact of such distribution discrepancies on supervised learning, specifically in the context of image classification. More importantly, *our framework suggests practical ways to generate good synthetic samples and to train a predictor with high generalization ability*. Building upon this framework, we propose a novel theoretical-based algorithm that integrates prototype learning to optimize both data partitioning and model training, effectively bridging the gap between real few-shot data and synthetic data. Extensive experiments results show that our approach demonstrates superior performance compared to state-of-the-art methods, outperforming them across multiple datasets.

Training a state-of-the-art Large Language Model (LLM) is an increasingly expensive endeavor due to growing computational, hardware, energy, and engineering demands. Yet, an often-overlooked (and seldom paid) expense is the human labor behind these models' training data. Every LLM is built on an unfathomable amount of human effort: trillions of carefully written words sourced from books, academic papers, codebases, social media, and more. This position paper aims to assign a monetary value to this labor and argues that the most expensive part of producing an LLM \emph{should} be the compensation provided to training data producers for their work. To support this position, we study 64 LLMs released between 2016 and 2024, estimating what it would cost to pay people to produce their training datasets from scratch. Even under highly conservative estimates of wage rates, the costs of these models' training datasets are $10$-$1000$ times larger than the costs to train the models themselves, representing a significant financial liability for LLM providers. In the face of the massive gap between the value of training data and the lack of compensation for its creation, we highlight and discuss research directions that could enable fairer practices in the future.

Large Language Models (LLMs) are often asked to generate structured outputs that obey precise syntactic rules, such as code snippets or formatted data. Grammar-constrained decoding (GCD) can guarantee that LLM outputs matches such rules by masking out tokens that will provably lead to outputs that do not belong to a specified context-free grammar (CFG). To guarantee soundness, GCD algorithms have to compute how a given LLM subword tokenizer can ``align'' with the tokens used by a given context-free grammar and compute token masks based on this information. Doing so efficiently is challenging and existing GCD algorithms require tens of minutes to preprocess common grammars. We present a new GCD algorithm together with an implementation that offers 17.71x faster offline preprocessing than existing approaches while preserving state-of-the-art efficiency in online mask computation.

Estimating the unknown reward functions driving agents' behavior is a central challenge in inverse games and reinforcement learning. This paper introduces a unified framework for reward function recovery in two-player zero-sum matrix games and Markov games with entropy regularization. Given observed player strategies and actions, we aim to reconstruct the underlying reward functions. This task is challenging due to the inherent ambiguity of inverse problems, the non-uniqueness of feasible rewards, and limited observational data coverage. To address these challenges, we establish reward function identifiability using the quantal response equilibrium (QRE) under linear assumptions. Building on this theoretical foundation, we propose an algorithm to learn reward from observed actions, designed to capture all plausible reward parameters by constructing confidence sets. Our algorithm works in both static and dynamic settings and is adaptable to incorporate other methods, such as Maximum Likelihood Estimation (MLE). We provide strong theoretical guarantees for the reliability and sample-efficiency of our algorithm. Empirical results demonstrate the framework’s effectiveness in accurately recovering reward functions across various scenarios, offering new insights into decision-making in competitive environments.

Neural network training is inherently sensitive to initialization and the randomness induced by stochastic gradient descent. However, it is unclear to what extent such effects lead to meaningfully different networks, either in terms of the models' weights or the underlying functions that were learned. In this work, we show that during the initial "chaotic" phase of training, even extremely small perturbations reliably causes otherwise identical training trajectories to diverge-an effect that diminishes rapidly over training time. We quantify this divergence through (i) $L^2$ distance between parameters, (ii) the loss barrier when interpolating between networks, (iii) $L^2$ and barrier between parameters after permutation alignment, and (iv) representational similarity between intermediate activations; revealing how perturbations across different hyperparameter or fine-tuning settings drive training trajectories toward distinct loss minima. Our findings provide insights into neural network training stability, with practical implications for fine-tuning, model merging, and diversity of model ensembles.

In continual learning scenarios, catastrophic forgetting of previously learned tasks is a critical issue, making it essential to effectively measure such forgetting. Recently, there has been growing interest in focusing on representation forgetting, the forgetting measured at the hidden layer. In this paper, we provide the first theoretical analysis of representation forgetting and use this analysis to better understand the behavior of continual learning. First, we introduce a new metric called representation discrepancy, which measures the difference between representation spaces constructed by two snapshots of a model trained through continual learning. We demonstrate that our proposed metric serves as an effective surrogate for the representation forgetting while remaining analytically tractable. Second, through mathematical analysis of our metric, we derive several key findings about the dynamics of representation forgetting: the forgetting occurs more rapidly to a higher degree as the layer index increases, while increasing the width of the network slows down the forgetting process. Third, we support our theoretical findings through experiments on real image datasets, including Split-CIFAR100 and ImageNet1K.

Through recognizing causal subgraphs, causal graph learning (CGL) has risen to be a promising approach for improving the generalizability of graph neural networks under out-of-distribution (OOD) scenarios. However, the empirical successes of CGL techniques are mostly exemplified in classification settings, while regression tasks, a more challenging setting in graph learning, are overlooked. We thus devote this work to tackling causal graph regression (CGR); to this end we reshape the processing of confounding effects in existing CGL studies, which mainly deal with classification. Specifically, we reflect on the predictive power of confounders in graph-level regression, and generalize classification-specific causal intervention techniques to regression through a lens of contrastive learning. Extensive experiments on graph OOD benchmarks validate the efficacy of our proposals for CGR. The model implementation and the code are provided on https://github.com/causal-graph/CGR.

We introduce a family of algorithms for online conformal prediction with coverage guarantees for both adversarial and stochastic data. In the adversarial setting, we establish the standard guarantee: over time, a pre-specified target fraction of confidence sets cover the ground truth. For stochastic data, we provide a guarantee at every time instead of just on average over time: the probability that a confidence set covers the ground truth—conditioned on past observations—converges to a pre-specified target when the conditional quantiles of the errors are a linear function of past data. Complementary to our theory, our experiments spanning over $15$ datasets suggest that the performance improvement of our methods over baselines grows with the magnitude of the data’s dependence, even when baselines are tuned on the test set. We put these findings to the test by pre-registering an experiment for electricity demand forecasting in Texas, where our algorithms achieve over a $10$\% reduction in confidence set sizes, a more than a $30$\% improvement in quantile and absolute losses with respect to the observed errors, and significant outcomes on all $78$ out of $78$ pre-registered hypotheses. We provide documentation for the pypi package implementing our algorithms here: \url{https://conformalopt.readthedocs.io/}.

Scaling laws predict the loss of a target machine learning model by extrapolating from easier-to-train models with fewer parameters or smaller training sets. This provides an efficient way for practitioners and researchers alike to compare pre- training decisions involving optimizers, datasets, and model architectures. Despite the widespread use of scaling laws to model the dynamics of language model training, there has been little work on understanding how to best estimate and interpret them. We collect (and release) a large-scale dataset containing losses and downstream evaluations for 485 previously published pretrained models. We use these to estimate more than 1000 scaling laws, then derive a set of best practices for estimating scaling laws in new model families. We find that fitting scaling laws to intermediate checkpoints of training runs (and not just their final losses) substan- tially improves accuracy, and that—all else equal—estimates of performance are generally most accurate when derived from other models of similar sizes. However, because there is a significant degree of variability across model seeds, training multiple small models is sometimes more useful than training a single large one. Moreover, while different model families differ in scaling behavior, they are often similar enough that a target model’s behavior can be predicted from a single model with the same architecture, along with scaling parameter estimates derived from other model families.

We present a novel self-supervised feature learning method using Vision Transformers (ViT) as the backbone, specifically designed for object detection and instance segmentation. Our approach addresses the challenge of extracting features that capture both class and positional information, which are crucial for these tasks. The method introduces two key components: (1) a positional encoding tied to the cropping process in contrastive learning, which utilizes a novel vector field representation for positional embeddings; and (2) masking and prediction, similar to conventional Masked Image Modeling (MIM), applied in parallel to both content and positional embeddings of image patches. These components enable the effective learning of intertwined content and positional features. We evaluate our method against state-of-the-art approaches, pre-training on ImageNet-1K and fine-tuning on downstream tasks. Our method outperforms the state-of-the-art SSL methods on the COCO object detection benchmark, achieving significant improvements with fewer pre-training epochs. These results suggest that better integration of positional information into self-supervised learning can improve performance on the dense prediction tasks.