ICML 2025 Papers

Language models (LMs) have shown impressive performance on tasks within their training distribution, but often struggle with structurally novel tasks even when given a small number of in-context task examples. We investigate the effectiveness of test-time training (TTT)—temporarily updating model parameters during inference using a loss derived from input data—as a mechanism for improving LMs' reasoning and few-shot learning capabilities. On the Abstraction and Reasoning Corpus (ARC), performing TTT with in-context examples yields up to $6\times$ higher accuracy compared to fine-tuned baselines—reaching $53.0\%$ on the public validation set with an 8B-parameter LM and $61.9\%$ when ensembled with program-synthesis methods, matching average human performance. On BIG-Bench Hard (BBH), TTT on in-context examples surpasses standard few-shot prompting in the $10$-shot setting by $7.3$ percentage points ($50.5\%$ to $57.8\%$). Our findings highlight the limitations of in-context learning for novel tasks and demonstrate the potential of test-time training to enhance language model adaptability.

Transformers excel at discovering patterns in sequential data, yet their fundamental limitations and learning mechanisms remain crucial topics of investigation. In this paper, we study the ability of Transformers to learn pseudo-random number sequences from linear congruential generators (LCGs), defined by the recurrence relation $x_{t+1} = a x_t + c \\;\mathrm{mod}\\; m$. We find that with sufficient architectural capacity and training data variety, Transformers can perform in-context prediction of LCG sequences with unseen moduli ($m$) and parameters ($a,c$). By analyzing the embedding layers and attention patterns, we uncover how Transformers develop algorithmic structures to learn these sequences in two scenarios of increasing complexity. First, we investigate how Transformers learn LCG sequences with unseen ($a, c$) but fixed modulus; and demonstrate successful learning up to $m = 2^{32}$. We find that models learn to factorize $m$ and utilize digit-wise number representations to make sequential predictions. In the second, more challenging scenario of unseen moduli, we show that Transformers can generalize to unseen moduli up to $m_{\text{test}} = 2^{16}$. In this case, the model employs a two-step strategy: first estimating the unknown modulus from the context, then utilizing prime factorizations to generate predictions. For this task, we observe a sharp transition in the accuracy at a critical depth $d= 3$. We also find that the number of in-context sequence elements needed to reach high accuracy scales sublinearly with the modulus.

Recent years have seen significant progress in developing spiking neural networks (SNNs) as a potential solution to the energy challenges posed by conventional artificial neural networks (ANNs). However, our theoretical understanding of SNNs remains relatively limited compared to the ever-growing body of literature on ANNs. In this paper, we study a discrete-time model of SNNs based on leaky integrate-and-fire (LIF) neurons, referred to as discrete-time LIF-SNNs, a widely used framework that still lacks solid theoretical foundations. We demonstrate that discrete-time LIF-SNNs realize piecewise constant functions defined on polyhedral regions, and more importantly, we quantify the network size required to approximate continuous functions. Moreover, we investigate the impact of latency (number of time steps) and depth (number of layers) on the complexity of the input space partitioning induced by discrete-time LIF-SNNs. Our analysis highlights the importance of latency and contrasts these networks with ANNs that use piecewise linear activation functions. Finally, we present numerical experiments to support our theoretical findings.

Process reward models (PRMs) provide more nuanced supervision compared to outcome reward models (ORMs) for optimizing policy models, positioning them as a promising approach to enhancing the capabilities of LLMs in complex reasoning tasks. Recent efforts have advanced PRMs from step-level to token-level granularity by integrating reward modeling into the training of generative models, with reward scores derived from token generation probabilities. However, the conflict between generative language modeling and reward modeling may introduce instability and lead to inaccurate credit assignments. To address this challenge, we revisit token-level reward assignment by decoupling reward modeling from language generation and derive a token-level reward model through the optimization of a discriminative policy, termed the Q-function Reward Model (Q-RM). We theoretically demonstrate that Q-RM explicitly learns token-level Q-functions from preference data without relying on fine-grained annotations. In our experiments, Q-RM consistently outperforms all baseline methods across various benchmarks. For example, when integrated into PPO/REINFORCE algorithms, Q-RM enhances the average Pass@1 score by 5.85/4.70 points on mathematical reasoning tasks compared to the ORM baseline, and by 4.56/5.73 points compared to the token-level PRM counterpart. Moreover, reinforcement learning with Q-RM significantly enhances training efficiency, achieving convergence 12× faster than ORM on GSM8K and 11× faster than step-level PRM on MATH. Code and data are available at https://github.com/homzer/Q-RM.

To improve the training efficiency of federated learning (FL), previous research has employed low-rank decomposition techniques to reduce communication overhead. In this paper, we seek to enhance the performance of these low-rank decomposition methods. Specifically, we focus on three key issues related to decomposition in FL: what to decompose, how to decompose, and how to aggregate. Subsequently, we introduce three novel techniques: Model Update Decomposition (MUD), Block-wise Kronecker Decomposition (BKD), and Aggregation-Aware Decomposition (AAD), each targeting a specific issue. These techniques are complementary and can be applied simultaneously to achieve optimal performance. Additionally, we provide a rigorous theoretical analysis to ensure the convergence of the proposed MUD. Extensive experimental results show that our approach achieves faster convergence and superior accuracy compared to relevant baseline methods. The code is available at https://github.com/Leopold1423/fedmud-icml25.

When training classification models, it expects that the learned features are compact within classes, and can well separate different classes. As the dominant loss function for training classification models, minimizing cross-entropy (CE) loss maximizes the compactness and distinctiveness, i.e., reaching neural collapse (NC). The recent works show that binary CE (BCE) performs also well in multi-class tasks. In this paper, we compare BCE and CE in deep feature learning. For the first time, we prove that BCE can also maximize the intra-class compactness and inter-class distinctiveness when reaching its minimum, i.e., leading to NC. We point out that CE measures the relative values of decision scores in the model training, implicitly enhancing the feature properties by classifying samples one-by-one. In contrast, BCE measures the absolute values of decision scores and adjust the positive/negative decision scores across all samples to uniformly high/low levels. Meanwhile, the classifier biases in BCE present a substantial constraint on the decision scores to explicitly enhance the feature properties in the training. The experimental results are aligned with above analysis, and show that BCE could improve the classification and leads to better compactness and distinctiveness among sample features. The codes have be released.

Robotic manipulation systems operating in diverse, dynamic environments must exhibit three critical abilities: multitask interaction, generalization to unseen scenarios, and spatial memory. While significant progress has been made in robotic manipulation, existing approaches often fall short in generalization to complex environmental variations and addressing memory-dependent tasks. To bridge this gap, we introduce **SAM2Act**, a multi-view robotic transformer-based policy that leverages multi-resolution upsampling with visual representations from large-scale foundation model. SAM2Act achieves a state-of-the-art average success rate of **86.8% across 18 tasks** in the RLBench benchmark, and demonstrates robust generalization on *The Colosseum* benchmark, with only a **4.3% performance gap** under diverse environmental perturbations. Building on this foundation, we propose **SAM2Act+**, a memory-based architecture inspired by SAM2, which incorporates a memory bank, an encoder, and an attention mechanism to enhance spatial memory. To address the need for evaluating memory-dependent tasks, we introduce ***MemoryBench***, a novel benchmark designed to assess spatial memory and action recall in robotic manipulation. SAM2Act+ achieves an average success rate of **94.3% on memory-based tasks** in *MemoryBench*, significantly outperforming existing approaches and pushing the boundaries of memory-based robotic systems. Project page: [sam2act.github.io](https://sam2act.github.io/).

Large language model pre-training has traditionally relied on human experts to craft heuristics for improving the corpora quality, resulting in numerous rules developed to date. However, these fixed rules lack the flexibility to address the unique characteristics of individual examples, yet crafting sample-wise rules is impractical for human experts. In this paper, we show that even small language models, with only 0.3B parameters, can exhibit substantial data refining capabilities. We propose Programming Every Example (ProX), a novel framework that treats data refinement as a programming task, and enables the model to refine corpora by generating and executing fine-grained operations, such as string normalization, for each individual example at scale. Experiments show that models trained on ProX-refined data consistently outperform other baselines across 10 benchmarks, demonstrating effectiveness across model sizes (up to 1.7B) and pre-training corpora (C4, RedPajama-V2, FineWeb, FineWeb-Edu, and DCLM). ProX also shows great potential in continual pre-training: on math domain, ProX boosts 7B models by up to 20% within 10B tokens—results typically achieved with much larger scale training (e.g., 200B tokens). We believe ProX offers a way to curate high-quality pre-training data, and finally contributes to efficient LLM development.

Dataset Distillation (DD) compresses large datasets into smaller, synthetic subsets, enabling models trained on them to achieve performance comparable to those trained on the full data. However, these models remain vulnerable to adversarial attacks, limiting their use in safety-critical applications. While adversarial robustness has been extensively studied in related fields, research on improving DD robustness is still limited. To address this, we propose ROME, a novel method that enhances the adversarial RObustness of DD by leveraging the InforMation BottlenEck (IB) principle. ROME includes two components: a performance-aligned term to preserve accuracy and a robustness-aligned term to improve robustness by aligning feature distributions between synthetic and perturbed images. Furthermore, we introduce the Improved Robustness Ratio (I-RR), a refined metric to better evaluate DD robustness. Extensive experiments on CIFAR-10 and CIFAR-100 demonstrate that ROME outperforms existing DD methods in adversarial robustness, achieving maximum I-RR improvements of nearly 40% under white-box attacks and nearly 35% under black-box attacks. Our code is available at https://github.com/zhouzhengqd/ROME.

Deep learning-based watermarking models play a crucial role in copyright protection across various applications. However, many high-performance models are limited in practical deployment due to their large number of parameters. Meanwhile, the robustness and invisibility performance of existing lightweight models are unsatisfactory. This presents a pressing need for a watermarking model that combines lightweight capacity with satisfactory performance. Our research identifies a key reason that limits the performance of existing watermarking frameworks: a mismatch between commonly used decoding losses (e.g., mean squared error and binary cross-entropy loss) and the actual decoding goal, leading to parameter redundancy. We propose two innovative solutions: (1) Decoding-oriented surrogate loss (DO), which redesigns the loss function to mitigate the influence of decoding-irrelevant optimization directions; and (2) Detachable projection head (PH), which incorporates a detachable redundant module during training to handle these irrelevant directions and is discarded during inference. Additionally, we propose a novel watermarking framework comprising five submodules, allowing for independent parameter reduction in each component. Our proposed model achieves better efficiency, invisibility, and robustness while utilizing only 2.2\% of the parameters compared to the state-of-the-art frameworks. By improving efficiency while maintaining robust copyright protection, our model is well suited for practical applications in resource-constrained environments. The DO and PH methods are designed to be plug-and-play, facilitating seamless integration into future lightweight models.

The message-passing mechanism of graph convolutional networks (i.e., GCNs) enables label information to reach more unlabeled neighbors, thereby increasing the utilization of labels. However, the additional label information does not always contribute positively to the GCN. To address this issue, we propose a new two-step framework called ELU-GCN. In the first stage, ELU-GCN conducts graph learning to learn a new graph structure (i.e., ELU-graph), which allows the additional label information to positively influence the predictions of GCN. In the second stage, we design a new graph contrastive learning on the GCN framework for representation learning by exploring the consistency and mutually exclusive information between the learned ELU graph and the original graph. Moreover, we theoretically demonstrate that the proposed method can ensure the generalization ability of GCNs. Extensive experiments validate the superiority of our method.

Current video-language models struggle with long-video understanding due to limited context lengths and reliance on sparse frame subsampling, which often leads to information loss. In this paper, we introduce $\infty$-Video, which is able to process arbitrarily long videos through a continuous-time long-term memory (LTM) consolidation mechanism. Our framework augments video Q-formers by making them able to process unbounded video contexts efficiently and without requiring additional training. Through continuous attention, our approach dynamically allocates higher granularity to the most relevant video segments, forming "sticky" memories which evolve over time. Experiments with Video-LLaMA and VideoChat2 demonstrate improved performance in video question-answering tasks, showcasing the potential of continuous-time LTM mechanisms to enable scalable and training-free comprehension of long videos.

Mixture of Experts (MoE) architecture improves Large Language Models (LLMs) with better scaling, but its higher parameter counts and memory demands create challenges for deployment. In this paper, we present MoE-SVD, a new decomposition-based compression framework tailored for MoE LLMs without any extra training. By harnessing the power of Singular Value Decomposition (SVD), MoE-SVD addresses the critical issues of decomposition collapse and matrix redundancy in MoE architectures. Specifically, we first decompose experts into compact low-rank matrices, resulting in accelerated inference and memory optimization. In particular, we propose selective decomposition strategy by measuring sensitivity metrics based on weight singular values and activation statistics to automatically identify decomposable expert layers. Then, we share a single V-matrix across all experts and employ a top-k selection for U-matrices. This low-rank matrix sharing and trimming scheme allows for significant parameter reduction while preserving diversity among experts. Comprehensive experiments on Mixtral, Phi-3.5, DeepSeek, and Qwen2 MoE LLMs show MoE-SVD outperforms other compression methods, achieving a 60\% compression ratio and 1.5× faster inference with minimal performance loss. Codes are available at: https://github.com/lliai/MoE-SVD.

Tabular data is widely utilized in various machine learning tasks. Current tabular learning research predominantly focuses on closed environments, while in real-world applications, open environments are often encountered, where distribution and feature shifts occur, leading to significant degradation in model performance. Previous research has primarily concentrated on mitigating distribution shifts, whereas feature shifts, a distinctive and unexplored challenge of tabular data, have garnered limited attention. To this end, this paper conducts the first comprehensive study on feature shifts in tabular data and introduces the first **tab**ular **f**eature-**s**hift **bench**mark (TabFSBench). TabFSBench evaluates impacts of four distinct feature-shift scenarios on four tabular model categories across various datasets and assesses the performance of large language models (LLMs) and tabular LLMs in the tabular benchmark for the first time. Our study demonstrates three main observations: (1) most tabular models have the limited applicability in feature-shift scenarios; (2) the shifted feature set importance has a linear relationship with model performance degradation; (3) model performance in closed environments correlates with feature-shift performance. Future research direction is also explored for each observation. Benchmark: [LAMDASZ-ML/TabFSBench](https://github.com/LAMDASZ-ML/TabFSBench).

Accurately predicting user watch-time is crucial for enhancing user stickiness and retention in video recommendation systems. Existing watch-time prediction approaches typically involve transformations of watch-time labels for prediction and subsequent reversal, ignoring both the natural distribution properties of label and the \textit{instance representation confusion} that results in inaccurate predictions. In this paper, we propose ProWTP, a two-stage method combining prototype learning and optimal transport for watch-time regression prediction, suitable for any deep recommendation model. Specifically, we observe that the watch-ratio (the ratio of watch-time to video duration) within the same duration bucket exhibits a multimodal distribution. To facilitate incorporation into models, we use a hierarchical vector quantised variational autoencoder (HVQ-VAE) to convert the continuous label distribution into a high-dimensional discrete distribution, serving as credible prototypes for calibrations. Based on this, ProWTP views the alignment between prototypes and instance representations as a Semi-relaxed Unbalanced Optimal Transport (SUOT) problem, where the marginal constraints of prototypes are relaxed. And the corresponding optimization problem is reformulated as a weighted Lasso problem for solution. Moreover, ProWTP introduces the assignment and compactness losses to encourage instances to cluster closely around their respective prototypes, thereby enhancing the prototype-level distinguishability. Finally, we conducted extensive offline experiments on two industrial datasets, demonstrating our consistent superiority in real-world application.

Multimodal Large Language Models (MLLMs) for tabular understanding have made significant progress in tasks such as financial report analysis and public data tests. However, our comprehensive analysis shows that these models are still limited in certain simple scenarios, particularly when handling compositional conditions in QA. Further investigation reveals that the poor performance can be attributed to two main challenges: the visual encoder's inability to accurately recognize the content of a row, and the model's tendency to overlook conditions in the question. To address these, we introduce a new Compositional Condition Tabular Understanding method, called {\sc CoCoTab}. Specifically, to capture the structural relationships within tables, we enhance the visual encoder with additional row and column patches. Moreover, we introduce the conditional tokens between the visual patches and query embeddings, ensuring the model focuses on relevant parts of the table according to the conditions specified in the query. Additionally, we also introduce the Massive Multimodal Tabular Understanding (MMTU) benchmark, which comprehensively assesses the full capabilities of MLLMs in tabular understanding. Our proposed method achieves state-of-the-art performance on both existing tabular understanding benchmarks and MMTU. Our code can be available at \url{https://github.com/LAMDA-Tabular/MMTU}.

As large language models (LLMs) are becoming more capable and widespread, the study of their failure cases is becoming increasingly important. Recent advances in standardizing, measuring, and scaling test-time compute suggest new methodologies for optimizing models to achieve high performance on hard tasks. In this paper, we apply these advances to the task of model jailbreaking: eliciting harmful responses from aligned LLMs. We develop an adversarial reasoning approach to automatic jailbreaking that leverages a loss signal to guide the test-time compute, achieving SOTA attack success rates against many aligned LLMs, even those that aim to trade inference-time compute for adversarial robustness. Our approach introduces a new paradigm in understanding LLM vulnerabilities, laying the foundation for the development of more robust and trustworthy AI systems.

Learning lighting adaptation is a crucial step in achieving good visual perception and supporting downstream vision tasks. Current research often addresses individual light-related challenges, such as high dynamic range imaging and exposure correction, in isolation. However, we identify shared fundamental properties across these tasks: i) different color channels have different light properties, and ii) the channel differences reflected in the spatial and frequency domains are different. Leveraging these insights, we introduce the channel-aware Learning Adaptive Lighting Network (LALNet), a multi-task framework designed to handle multiple light-related tasks efficiently. Specifically, LALNet incorporates color-separated features that highlight the unique light properties of each color channel, integrated with traditional color-mixed features by Light Guided Attention (LGA). The LGA utilizes color-separated features to guide color-mixed features focusing on channel differences and ensuring visual consistency across all channels. Additionally, LALNet employs dual domain channel modulation for generating color-separated features and a mixed channel modulation and light state space module for producing color-mixed features. Extensive experiments on four representative light-related tasks demonstrate that LALNet significantly outperforms state-of-the-art methods on benchmark tests and requires fewer computational resources. We provide an anonymous online demo at [LALNet](https://xxxxxx2025.github.io/LALNet/).

Text-to-song generation, the task of creating vocals and accompaniment from textual inputs, poses significant challenges due to domain complexity and data scarcity. Existing approaches often employ multi-stage generation procedures, leading to cumbersome training and inference pipelines, as well as suboptimal overall generation quality due to error accumulation across stages. In this paper, we propose **SongGen**, a fully open-source, single-stage auto-regressive transformer designed for controllable song generation. The proposed model facilitates fine-grained control over diverse musical attributes, including lyrics and textual descriptions of instrumentation, genre, mood, and timbre, while also offering an optional three-second reference clip for voice cloning. Within a unified auto-regressive framework, SongGen supports two output modes: **mixed mode**, which generates a mixture of vocals and accompaniment directly, and **dual-track mode**, which synthesizes them separately for greater flexibility in downstream applications. We explore diverse token pattern strategies for each mode, leading to notable improvements and valuable insights. Furthermore, we design an automated data preprocessing pipeline with effective quality control. To foster community engagement and future research, we will release our model weights, training code, annotated data, and preprocessing pipeline. The code is available at https://github.com/LiuZH-19/SongGen.