ICML 2025 论文列表

While recent code-specific large language models (LLMs) have greatly enhanced their code generation capabilities, the safety of these models remains under-explored, posing potential risks as insecure code generated by these models may introduce vulnerabilities into real-world systems. Existing methods collect security-focused datasets from real-world vulnerabilities for instruction tuning in order to mitigate such issues. However, they are largely constrained by the data sparsity of vulnerable code, and have limited applicability in the multi-stage post-training workflows of modern LLMs. In this paper, we propose ProSec, a novel proactive security alignment approach designed to align code LLMs with secure coding practices. ProSec systematically exposes the vulnerabilities in a code LLM by synthesizing vulnerability-inducing coding scenarios from Common Weakness Enumerations (CWEs) and generates fixes to vulnerable code snippets, allowing the model to learn secure practices through preference learning objectives. The scenarios synthesized by ProSec trigger 25× more vulnerable code than a normal instruction-tuning dataset, resulting in a security-focused alignment dataset 7× larger than the previous work. Experiments show that models trained with ProSec are 25.2% to 35.4% more secure compared to previous work without degrading models' utility.

Reinforcement learning (RL) training is inherently unstable due to factors such as moving targets and high gradient variance. Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning from AI Feedback (RLAIF) introduce additional challenges. For instance, diverse preferences complicate the alignment process, and prediction errors in a trained reward model can become more severe as the LLM generates unseen outputs. These RL challenges create confusion about whether the probability of an action for a given state should be increased or decreased, similar to the noise in labels for classification tasks. In this work, we focus on RL algorithms that share learning difficulties with cross-entropy loss, especially for low-probability predictions. To enhance stability, we adapt reverse cross-entropy (RCE) from supervised learning for noisy data, defining a symmetric RL loss. We demonstrate performance improvements across various tasks and scales. We conduct experiments in discrete action tasks (Atari games) and continuous action space tasks (MuJoCo benchmark and Box2D) using Symmetric A2C (SA2C) and Symmetric PPO (SPPO). Notably, SPPO shows strong performance across different hyperparameters. Furthermore, we validate the symmetric RL loss in the RLHF framework using PPO for natural language processing tasks such as IMDB positive sentiment and TL;DR summarization.

Rigorous statistical evaluations of large language models (LLMs), including valid error bars and significance testing, are essential for meaningful and reliable performance assessment. Currently, when such statistical measures are reported, they typically rely on the Central Limit Theorem (CLT). In this position paper, we argue that while CLT-based methods for uncertainty quantification are appropriate when benchmarks consist of thousands of examples, they fail to provide adequate uncertainty estimates for LLM evaluations that rely on smaller, highly specialized benchmarks. In these small-data settings, we demonstrate that CLT-based methods perform very poorly, usually dramatically underestimating uncertainty (i.e. producing error bars that are too small). We give recommendations for alternative frequentist and Bayesian methods that are both easy to implement and more appropriate in these increasingly common scenarios.

We introduce Lexico, a novel KV cache compression method that leverages sparse coding with a universal dictionary. Our key finding is that key-value cache in modern LLMs can be accurately approximated using sparse linear combination from a small, input-agnostic dictionary of ~4k atoms, enabling efficient compression across different input prompts, tasks and models. Using orthogonal matching pursuit for sparse approximation, Lexico achieves flexible compression ratios through direct sparsity control. On GSM8K, across multiple model families (Mistral, Llama 3, Qwen2.5), Lexico maintains 90-95% of the original performance while using only 15-25% of the full KV-cache memory, outperforming both quantization and token eviction methods. Notably, Lexico remains effective in low memory regimes where 2-bit quantization fails, achieving up to 1.7x better compression on LongBench and GSM8K while maintaining high accuracy.

In-context learning with large language models (LLMs) delivers strong few-shot performance by choosing few-shot demonstrations from the entire training dataset. However, previous few-shot in-context learning methods, which calculate similarity scores for choosing demonstrations, incur high computational costs by repeatedly retrieving large-scale datasets for each query. This is due to their failure to recognize that not all demonstrations are equally informative, and many less informative demonstrations can be inferred from a core set of highly informative ones. To this end, we propose FEEDER (FEw yet Essential Demonstration prE-selectoR), a novel \emph{pre-selection} framework that identifies a core subset of demonstrations containing the most informative examples. This subset, referred to as the FEEDER set, consists of demonstrations that capture both the ''sufficiency'' and ''necessity'' information to infer the entire dataset. Notice that FEEDER is selected before the few-shot in-context learning, enabling more efficient few-shot demonstrations choosing in a smaller set. To identify FEEDER, we propose a novel effective tree based algorithm. Once selected, it can replace the original dataset, leading to improved efficiency and prediction accuracy in few-shot in-context learning. Additionally, FEEDER also benefit fine-tuning LLMs, we propose a bi-level optimization method enabling more efficient training without sacrificing performance when datasets become smaller. Our experiments are on 6 text classification datasets, 1 reasoning dataset, and 1 semantic-parsing dataset, across 6 LLMs (ranging from 335M to 7B parameters), demonstrate that: (i) In few-shot inference, FEEDER achieves superior (or comparable) performance while utilizing only half the input training data. (ii) In fine-tuning, FEEDER significantly boosts the performance of LLMs.

We present TUMTraf VideoQA, a novel dataset and benchmark designed for spatio-temporal video understanding in complex roadside traffic scenarios. The dataset comprises 1,000 videos, featuring 85,000 multiple-choice QA pairs, 2,300 object captioning, and 5,700 object grounding annotations, encompassing diverse real-world conditions such as adverse weather and traffic anomalies. By incorporating tuple-based spatio-temporal object expressions, TUMTraf VideoQA unifies three essential tasks—multiple-choice video question answering, referred object captioning, and spatio-temporal object grounding—within a cohesive evaluation framework. We further introduce the TraffiX-Qwen baseline model, enhanced with visual token sampling strategies, providing valuable insights into the challenges of fine-grained spatio-temporal reasoning. Extensive experiments demonstrate the dataset’s complexity, highlight the limitations of existing models, and position TUMTraf VideoQA as a robust foundation for advancing research in intelligent transportation systems. The dataset and benchmark are publicly available to facilitate further exploration.

Graph neural networks (GNNs) have achieved remarkable success, yet most are developed under the in-distribution assumption and fail to generalize to out-of-distribution (OOD) environments. To tackle this problem, some graph invariant learning methods aim to learn invariant subgraph against distribution shifts, which heavily rely on predefined or automatically generated environment labels. However, directly annotating or estimating such environment labels from biased graph data is typically impractical or inaccurate for real-world graphs. Consequently, GNNs may become biased toward variant patterns, resulting in poor OOD generalization. In this paper, we propose to learn disentangled invariant subgraph via self-supervised contrastive variant subgraph estimation for achieving satisfactory OOD generalization. Specifically, we first propose a GNN-based invariant subgraph generator to disentangle the invariant and variant subgraphs. Then, we estimate the degree of the spurious correlations by conducting self-supervised contrastive learning on variant subgraphs. Thanks to the accurate identification and estimation of the variant subgraphs, we can capture invariant subgraphs effectively and further eliminate spurious correlations by inverse propensity score reweighting. We provide theoretical analyses to show that our model can disentangle the ground-truth invariant and variant subgraphs for OOD generalization. Extensive experiments demonstrate the superiority of our model over state-of-the-art baselines.

With the advance of diffusion models, today's video generation has achieved impressive quality. To extend the generation length and facilitate real-world applications, a majority of video diffusion models (VDMs) generate videos in an autoregressive manner, i.e., generating subsequent clips conditioned on the last frame(s) of the previous clip. However, existing autoregressive VDMs are highly inefficient and redundant: The model must re-compute all the conditional frames that are overlapped between adjacent clips. This issue is exacerbated when the conditional frames are extended autoregressively to provide the model with long-term context. In such cases, the computational demands increase significantly (i.e., with a quadratic complexity w.r.t. the autoregression step). In this paper, we propose **Ca2-VDM**, an efficient autoregressive VDM with **Ca**usal generation and **Ca**che sharing. For **causal generation**, it introduces unidirectional feature computation, which ensures that the cache of conditional frames can be precomputed in previous autoregression steps and reused in every subsequent step, eliminating redundant computations. For **cache sharing**, it shares the cache across all denoising steps to avoid the huge cache storage cost. Extensive experiments demonstrated that our Ca2-VDM achieves state-of-the-art quantitative and qualitative video generation results and significantly improves the generation speed. Code is available: https://github.com/Dawn-LX/CausalCache-VDM

Mesh-based 3D static analysis methods have recently emerged as efficient alternatives to traditional computational numerical solvers, significantly reducing computational costs and runtime for various physics-based analyses. However, these methods primarily focus on surface topology and geometry, often overlooking the inherent thickness of real-world 3D objects, which exhibits high correlations and similar behavior between opposing surfaces. This limitation arises from the disconnected nature of these surfaces and the absence of internal edge connections within the mesh. In this work, we propose a novel framework, the Thickness-aware E(3)-Equivariant 3D Mesh Neural Network (T-EMNN), that effectively integrates the thickness of 3D objects while maintaining the computational efficiency of surface meshes. Additionally, we introduce data-driven coordinates that encode spatial information while preserving E(3)-equivariance or invariance properties, ensuring consistent and robust analysis. Evaluations on a real-world industrial dataset demonstrate the superior performance of T-EMNN in accurately predicting node-level 3D deformations, effectively capturing thickness effects while maintaining computational efficiency.

Answering questions with Chain-of-Thought (CoT) has significantly enhanced the reasoning capabilities of Large Language Models (LLMs), yet its impact on Large Multimodal Models (LMMs) still lacks a systematic assessment and in-depth investigation. In this paper, we introduce **MME-CoT**, a specialized benchmark evaluating the CoT reasoning performance of LMMs, spanning six domains: math, science, OCR, logic, space-time, and general scenes. As the first comprehensive study in this area, we propose a thorough evaluation suite incorporating three novel metrics that assess the reasoning quality, robustness, and efficiency at a fine-grained level. Leveraging curated high-quality data and a unique evaluation strategy, we conduct an in-depth analysis of state-of-the-art LMMs, uncovering several key insights: *1)* Models with reflection mechanism demonstrate a superior CoT quality, with Kimi k1.5 outperforming GPT-4o and demonstrating the highest quality results; *2)* CoT prompting often degrades LMM performance on perception-heavy tasks, suggesting a potentially harmful overthinking behavior; and *3)* Although the CoT quality is high, LMMs with reflection exhibit significant inefficiency in both normal response and self-correction phases. We hope MME-CoT serves as a foundation for advancing multimodal reasoning in LMMs.

On-device learning has emerged as a promising direction for AI development, particularly because of its potential to reduce latency issues and mitigate privacy risks associated with device-server communication, while improving energy efficiency. Despite these advantages, significant memory and computational constraints still represent major challenges for its deployment. Drawing on previous studies on low-rank decomposition methods that address activation memory bottlenecks in backpropagation, we propose a novel shortcut approach as an alternative. Our analysis and experiments demonstrate that our method can reduce activation memory usage, even up to $120.09\times$ compared to vanilla training, while also reducing overall training FLOPs up to $1.86\times$ when evaluated on traditional benchmarks.

Long-range interactions are essential for the correct description of complex systems in many scientific fields. The price to pay for including them in the calculations, however, is a dramatic increase in the overall computational costs. Recently, deep graph networks have been employed as efficient, data-driven models for predicting properties of complex systems represented as graphs. These models rely on a message passing strategy that should, in principle, capture long-range information without explicitly modeling the corresponding interactions. In practice, most deep graph networks cannot really model long-range dependencies due to the intrinsic limitations of (synchronous) message passing, namely oversmoothing, oversquashing, and underreaching. This work proposes a general framework that \textit{learns to mitigate} these limitations: within a variational inference framework, we endow message passing architectures with the ability to adapt their depth and filter messages along the way. With theoretical and empirical arguments, we show that this strategy better captures long-range interactions, by competing with the state of the art on five node and graph prediction datasets.

Large language models are pre-trained on ever-growing token budgets under the assumption that better pre-training performance translates to improved downstream models. In this work, we challenge this assumption and show that extended pre-training can make models harder to fine-tune, leading to degraded final performance. We term this phenomenon \textbf{catastrophic overtraining}. For example, the instruction-tuned OLMo-1B model pre-trained on 3T tokens leads to over 2\% worse performance on multiple standard LLM benchmarks than its 2.3T token counterpart. Through controlled experiments and theoretical analysis, we show that catastrophic overtraining arises from a systematic increase in the broad sensitivity of pre-trained parameters to modifications, including but not limited to fine-tuning. Our findings call for a critical reassessment of pre-training design that considers the downstream adaptability of the model.

The real world naturally has dimensions of time and space. Therefore, estimating the counterfactual outcomes with spatial-temporal attributes is a crucial problem. However, previous methods are based on classical statistical models, which still have limitations in performance and generalization. This paper proposes a novel framework for estimating counterfactual outcomes with spatial-temporal attributes using the Transformer, exhibiting stronger estimation ability. Under mild assumptions, the proposed estimator within this framework is consistent and asymptotically normal. To validate the effectiveness of our approach, we conduct simulation experiments and real data experiments. Simulation experiments show that our estimator has a stronger estimation capability than baseline methods. Real data experiments provide a valuable conclusion to the causal effect of conflicts on forest loss in Colombia. The source code is available at this [URL](https://github.com/lihe-maxsize/DeppSTCI_Release_Version-master).

We examine the problem of learning sequential tasks from a single visual demonstration. A key challenge arises when demonstrations are temporally misaligned due to variations in timing, differences in embodiment, or inconsistencies in execution. Existing approaches treat imitation as a distribution-matching problem, aligning individual frames between the agent and the demonstration. However, we show that such frame-level matching fails to enforce temporal ordering or ensure consistent progress. Our key insight is that matching should instead be defined at the level of sequences. We propose that perfect matching occurs when one sequence successfully covers all the subgoals in the same order as the other sequence. We present ORCA (ORdered Coverage Alignment), a dense per-timestep reward function that measures the probability of the agent covering demonstration frames in the correct order. On temporally misaligned demonstrations, we show that agents trained with the ORCA reward achieve $4.5$x improvement ($0.11 \rightarrow 0.50$ average normalized returns) for Meta-world tasks and $6.6$x improvement ($6.55 \rightarrow 43.3$ average returns) for Humanoid-v4 tasks compared to the best frame-level matching algorithms. We also provide empirical analysis showing that ORCA is robust to varying levels of temporal misalignment. The project website is at https://portal-cornell.github.io/orca/

Satellite imagery is essential for Earth observation, enabling applications like crop yield prediction, environmental monitoring, and climate change assessment. However, integrating satellite imagery with climate data remains a challenge, limiting its utility for forecasting and scenario analysis. We introduce a novel dataset of 2.9 million Sentinel-2 images spanning 15 land cover types with corresponding climate records, forming the foundation for two satellite image generation approaches using fine-tuned Stable Diffusion 3 models. The first is a text-to-image generation model that uses textual prompts with climate and land cover details to produce realistic synthetic imagery for specific regions. The second leverages ControlNet for multi-conditional image generation, preserving spatial structures while mapping climate data or generating time-series to simulate landscape evolution. By combining synthetic image generation with climate and land cover data, our work advances generative modeling in remote sensing, offering realistic inputs for environmental forecasting and new possibilities for climate adaptation and geospatial analysis.

We derive a controlled generation objective within the framework of Variational Flow Matching (VFM), which casts flow matching as a variational inference problem. We demonstrate that controlled generation can be implemented two ways: (1) by way of end-to-end training of conditional generative models, or (2) as a Bayesian inference problem, enabling post hoc control of unconditional models without retraining. Furthermore, we establish the conditions required for equivariant generation and provide an equivariant formulation of VFM tailored for molecular generation, ensuring invariance to rotations, translations, and permutations. We evaluate our approach on both uncontrolled and controlled molecular generation, achieving state-of-the-art performance on uncontrolled generation and outperforming state-of-the-art models in controlled generation, both with end-to-end training and in the Bayesian inference setting. This work strengthens the connection between flow-based generative modeling and Bayesian inference, offering a scalable and principled framework for constraint-driven and symmetry-aware generation.

Recent advancements in data-driven weather forecasting models have delivered deterministic models that outperform the leading operational forecast systems based on traditional, physics-based models. However, these data-driven models are typically trained with a mean squared error loss function, which causes smoothing of fine scales through a ``double penalty'' effect. We develop a simple, parameter-free modification to this loss function that avoids this problem by separating the loss attributable to decorrelation from the loss attributable to spectral amplitude errors. Fine-tuning the GraphCast model with this new loss function results in sharp deterministic weather forecasts, an increase of the model's effective resolution from 1,250km to 160km, improvements to ensemble spread, and improvements to predictions of tropical cyclone strength and surface wind extremes.

Addressing reward design complexities in deep reinforcement learning is facilitated by knowledge transfer across different domains. To this end, we define \textit{reward translation} to describe the cross-domain reward transfer problem. However, current methods struggle with non-pairable and non-time-alignable incompatible MDPs. This paper presents an adaptable reward translation framework \textit{neural reward translation} featuring \textit{semi-alignable MDPs}, which allows efficient reward translation under relaxed constraints while handling the intricacies of incompatible MDPs. Given the inherent difficulty of directly mapping semi-alignable MDPs and transferring rewards, we introduce an indirect mapping method through reward machines, created using limited human input or LLM-based automated learning. Graph-matching techniques establish links between reward machines from distinct environments, thus enabling cross-domain reward translation within semi-alignable MDP settings. This broadens the applicability of DRL across multiple domains. Experiments substantiate our approach's effectiveness in tasks under environments with semi-alignable MDPs.