Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. We present OmniScale, a modular and efficient training framework to accelerate the development of omni-modal LLMs. OmniScale introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. OmniScale also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. Using OmniScale, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Skill discovery has emerged as a popular route for unsupervised reinforcement learning (URL), offering agents a diverse, reusable set of behaviours learned before any task-specific reward is experienced. However, existing methodologies tend to favour either categorical codes or unimodal skill priors, which simplifies training at the cost of limiting the variety of behaviours they can represent. We introduce Discovery of Mixture Skills (DiMS), a URL algorithm that learns a latent Gaussian mixture by training a Gaussian Mixture Variational Autoencoder (GMVAE) in tandem with the unsupervised policy. In DiMS, a hierarchical GMVAE simultaneously discovers clusters of skills, while an auxiliary macro-latent dynamically positions mixture components to prevent mode collapse. A joint loss term combining log-likelihood and curiosity rewards enables simultaneous updates of representation and policy while improving exploration. Experiments on the Unsupervised Reinforcement Learning Benchmark (URLB) show that DiMS consistently outperforms a wide range of state-of-the-art baselines. Ablation studies confirm that the mixture prior is critical to these gains, and that DiMS is robust to alternative exploration bonuses. Overall, our results suggest that Gaussian mixture skill priors offer a compelling foundation for future unsupervised RL.
Video Large Language Models (VLLMs) excel in video understanding, but their excessive visual tokens pose a significant computational challenge for real-world applications. Current methods aim to enhance inference efficiency by visual token pruning. However, they do not consider the dynamic characteristics and temporal dependencies of video frames, as they perceive video understanding as a multi-frame task. To address these challenges, we propose MMG-Vid, a novel training-free visual token pruning framework that removes redundancy by Maximizing Marginal Gains at both segment-level and token-level. Specifically, we first divide the video into segments based on frame similarity, and then dynamically allocate the token budget for each segment to maximize the marginal gain of each segment. Subsequently, we propose a temporal-guided DPC algorithm that jointly models inter-frame uniqueness and intra-frame diversity, thereby maximizing the marginal gain of each token. By combining both stages, MMG-Vid can maximize the utilization of the limited token budget, significantly improving efficiency while maintaining strong performance. Extensive experiments demonstrate that MMG-Vid can maintain over 99.5% of the original performance, while effectively reducing 75% visual tokens and accelerating the prefilling stage by 3.9x on LLaVA-OneVision-7B.
Text-attributed graphs (TAGs), which combine structural and textual node information, are ubiquitous across many domains. Recent work integrates Large Language Models (LLMs) with Graph Neural Networks (GNNs) to jointly model semantics and structure, resulting in more general and expressive models that achieve state-of-the-art performance on TAG benchmarks. However, this integration introduces dual vulnerabilities: GNNs are sensitive to structural perturbations, while LLM-derived features are vulnerable to prompt injection and adversarial phrasing. While existing adversarial attacks largely perturb structure or text independently, we find that uni-modal attacks cause only modest degradation in LLM-enhanced GNNs. Moreover, many existing attacks assume unrealistic capabilities, such as white-box access or direct modification of graph data. To address these gaps, we propose GraphTextack, the first black-box, multi-modal node injection attack for LLM-enhanced GNNs. GraphTextack injects nodes with carefully crafted structure and semantics to degrade model performance, operating under a realistic threat model without relying on model internals or surrogate models. To navigate the combinatorial, non-differentiable search space of connectivity and feature assignments, GraphTextack introduces a novel evolutionary optimization framework with a multi-objective fitness function that balances local prediction disruption and global graph influence. Extensive experiments on five datasets and two state-of-the-art LLM-enhanced GNN models show that GraphTextack significantly outperforms 12 strong baselines.
TIMA: Text-Image Mutual Awareness for Balancing Zero-Shot Adversarial Robustness and Generalization Ability
PDF ↗Achieving zero-shot adversarial robustness without sacrificing generalization remains challenging for foundation models such as CLIP, especially under large adversarial perturbations. Through empirical analyses, we identify three critical yet overlooked issues: (1) Logit margins exhibit a stable offset between small and large adversarial perturbations, suggesting that explicitly adjusting margins could improve robustness against unseen large perturbations. (2) A significant negative correlation exists between logit margin and inter-class semantic similarity, indicating that semantic structures are insufficiently leveraged by existing methods. (3) Existing methods for adjusting text embeddings disrupt the intrinsic semantic consistency established by pre-trained models, undermining generalization capability. Motivated by these findings, we propose a novel Text-Image Mutual Awareness (TIMA) framework, including a Text-Aware Image (TAI) tuning module with an Adaptive Semantic-Aware Margin (ASAM) to explicitly calibrate logit margins, and an Image-Aware Text (IAT) tuning module with Semantic Consistent Minimum Hyperspherical Energy (SC-MHE) to preserve semantic consistency. Comprehensive experiments validate that TIMA significantly outperforms existing approaches by effectively addressing the identified limitations.
Multimodal Emotion Recognition in Conversation (MERC) aims to predict speakers’ emotions by integrating textual, acoustic, and visual cues. Existing approaches either struggle to capture complex cross‑modal interactions or experience gradient conflicts and unstable training when using deeper architectures. To address these issues, we propose Cross-Space Synergy (CSS), which couples a representation component with an optimization component. Synergistic Polynomial Fusion (SPF) serves the representation role, leveraging low-rank tensor factorization to efficiently capture high-order cross-modal interactions. Pareto Gradient Modulator (PGM) serves the optimization role, steering updates along Pareto-optimal directions across competing objectives to alleviate gradient conflicts and improve stability. Experiments show that CSS outperforms existing representative methods on IEMOCAP and MELD in both accuracy and training stability, demonstrating its effectiveness in complex multimodal scenarios.
Time series forecasting is fundamental to diverse applications, with recent approaches leverage large vision models (LVMs) to capture temporal patterns through visual representations. We reveal that while vision models enhance forecasting performance, 99% of their parameters are unnecessary for time series tasks. Through cross-modal analysis, we find that time series align with low-level textural features but not high-level semantics, which can impair forecasting accuracy. We propose OccamVTS, a knowledge distillation framework that extracts only the essential 1% of predictive information from LVMs into lightweight networks. Using pre-trained LVMs as privileged teachers, OccamVTS employs pyramid-style feature alignment combined with correlation and feature distillation to transfer beneficial patterns while filtering out semantic noise. Counterintuitively, this aggressive parameter reduction improves accuracy by eliminating overfitting to irrelevant visual features while preserving essential temporal patterns. Extensive experiments across multiple benchmark datasets demonstrate that OccamVTS consistently achieves state-of-the-art performance with only 1% of the original parameters, particularly excelling in few-shot and zero-shot scenarios.
Strategic machine learning investigates scenarios where agents manipulate their features to receive favorable decisions from predictive models. To address fairness concerns intrinsic to strategic classification, recent work has introduced group-specific fairness constraints. However, current fairness-aware approaches face a fundamental dilemma in the issue of fairness exposure: making these constraints public enables strategic manipulation and can lead to fairness reversal, while keeping them hidden may reduce social welfare and discourage genuine improvement. To fill this gap, we subsequently propose the problem of Partial Fairness Awareness (PFA), as our theoretical analysis informs that such a dilemma can be mitigated by releasing the candidate set of fairness constraints and concealing the grounding constraint. To be specific, we introduce a belief-guided strategic mechanism wherein agents iteratively interact with the decision system and maintain a belief distribution over the candidate set of fairness constraints. This belief-guided process enables agents, through iterative interaction and feedback, to update their belief distribution over the candidate set, thereby gradually aligning their belief with the grounding fairness constraint employed by the system. Extensive experiments on real-world and synthetic datasets demonstrate that PFA achieves lower group fairness gaps, higher acceptance of truly qualified individuals, and more stable outcomes compared to fully public or private fairness regimes.
In multi-view classification tasks (MVC), each view provides an unique perspective on the data, offering complementary information that can improve classification performance when properly integrated. However, traditional methods typically adopt a uniform processing strategy for all views before fusion, overlooking the fact that different views may require different treatments due to variations in their quality and informativeness. To address this limitation, we propose a novel framework called Uncertainty-Guided View-Strength-Aware Feature Utilization (UVF) for multi-view classification. Our approach introduces a view uncertainty estimation module to quantify the discriminative strength of each view. Based on this estimation, a Differentiated Feature Selector (DFS) adaptively selects features, retaining informative dimensions in weak views while preserving original features in strong views. Furthermore, we employ an uncertainty-guided fusion strategy that assigns dynamic weights to each view's contribution based on its uncertainty score, enhancing the robustness and reliability of the final decision. Experimental results on benchmark datasets demonstrate that our method significantly outperforms conventional approaches, achieving better classification accuracy and interpretability through strength-aware feature processing and fusion.
Large Vision-Language Models (VLMs) exhibit impressive multi-modal capabilities but suffer from prohibitive computational and memory demands, due to their long visual token sequences and massive parameter sizes. To address these issues, recent works have proposed training-free compression methods. However, existing efforts often suffer from three major limitations: (1) Current approaches do not decompose techniques into comparable modules, hindering fair evaluation across spatial and temporal redundancy. (2) Evaluation confined to simple single-turn tasks, failing to reflect performance in realistic scenarios. (3) Isolated use of individual compression techniques, without exploring their joint potential. To overcome these gaps, we introduce LLMC+, a comprehensive VLM compression benchmark with a versatile, plug-and-play toolkit. LLMC+ supports over 20 algorithms across five representative VLM families and enables systematic study of token-level and model-level compression. Our benchmark reveals that: (1) Spatial and temporal redundancies demand distinct technical strategies. (2) Token reduction methods degrade significantly in multi-turn dialogue and detail-sensitive tasks. (3) Combining token and model compression achieves extreme compression with minimal performance loss. We believe LLMC+ will facilitate fair evaluation and inspire future research in efficient VLM.
We propose new quantum algorithms for estimating spectral sums of positive semi-definite (PSD) matrices. For a matrix A and a function f, the spectral sum is the trace of f(A), equivalently the sum over eigenvalues of A of f applied to each eigenvalue. Typical examples of spectral sums are the von Neumann entropy, the trace of the inverse of A, the log-determinant, and the Schatten p-norm, where the latter does not require the matrix to be PSD. The current best classical randomized algorithms estimating these quantities have a runtime that is at least linearly in the number of nonzero entries of the matrix and quadratic in the estimation error. Assuming access to a block-encoding of a matrix, our algorithms are sub-linear in the matrix size, and depend at most quadratically on other parameters, like the condition number and the approximation error, and thus can compete with most of the randomized and distributed classical algorithms proposed in the literature, and polynomially improve the runtime of other quantum algorithms proposed for the same problems. We show how the algorithms and techniques used in this work can be applied to three problems in spectral graph theory: approximating the number of triangles, the effective resistance, and the number of spanning trees in a graph.
This paper studies the problem of dimension reduction, tailored to improving time series forecasting with high-dimensional predictors. We propose a novel Supervised Deep Dynamic Principal component analysis (SDDP) framework that incorporates the target variable and lagged observations into the factor extraction process. Assisted by a temporal neural network, we construct target-aware predictors by scaling the original predictors in a supervised manner, with larger weights assigned to predictors with stronger forecasting power. A principal component analysis is then performed on the target-aware predictors to extract the estimated SDDP factors. This supervised factor extraction not only improves predictive accuracy in the downstream forecasting task but also yields more interpretable and target-specific latent factors. Building upon SDDP, we propose a factor-augmented nonlinear dynamic forecasting model that unifies a broad family of factor-model-based forecasting approaches. To further demonstrate the broader applicability of SDDP, we extend our studies to a more challenging scenario when the predictors are only partially observable. We validate the empirical performance of the proposed method on several real-world public datasets. The results show that our algorithm achieves notable improvements in forecasting accuracy compared to state-of-the-art methods.
Recent advances in long video understanding typically mitigate visual redundancy through visual token pruning based on attention distribution. However, while existing methods employ post-hoc low-response token pruning in decoder layers, they overlook the input-level semantic correlation between visual tokens and instructions (query). In this paper, we propose QuoTA, an ante-hoc training-free modular that extends existing large video-language models (LVLMs) for visual token assignment based on query-oriented frame-level importance assessment. The query-oriented token selection is crucial as it aligns visual processing with task-specific requirements, optimizing token budget utilization while preserving semantically relevant content. Specifically, (i) QuoTA strategically allocates frame-level importance scores based on query relevance, enabling one-time visual token assignment before cross-modal interactions in decoder layers, (ii) we decouple the query through Chain-of-Thoughts reasoning to facilitate more precise LVLM-based frame importance scoring, and (iii) QuoTA offers a plug-and-play functionality that extends to existing LVLMs. Extensive experimental results demonstrate that implementing QuoTA with LLaVA-Video-7B yields an average performance improvement of 3.2% across six benchmarks (including Video-MME and MLVU) while operating within an identical visual token budget as the baseline.
ViTCoP: Accelerating Large Vision-Language Models via Visual and Textual Semantic Collaborative Pruning
PDF ↗Large Vision-Language Models (LVLMs) incur high computational costs due to significant redundancy in their visual tokens. To effectively reduce this cost, researchers have proposed various visual token pruning methods. However, existing methods are generally limited, either losing critical visual information prematurely due to pruning in the vision encoder, or leading to information redundancy among the selected tokens due to pruning in the Large Language Models (LLMs). To address these challenges, we propose a Visual and Textual Semantic Collaborative Pruning framework (ViTCoP) that combines redundancy filtering in the vision encoder with step-wise co-pruning within the LLM based on its hierarchical characteristics, to efficiently preserve critical and informationally diverse visual tokens. Meanwhile, to ensure compatibility with acceleration techniques like FlashAttention, we introduce the L2 norm of K-vectors as the token saliency metric in the LLM. Extensive experiments on various Large Vision-Language Models demonstrate that ViTCoP not only achieves state-of-the-art performance surpassing existing methods on both image and video understanding tasks, but also significantly reduces model inference latency and GPU memory consumption. Notably, its performance advantage over other methods becomes even more pronounced under extreme pruning rates.
Towards Effective, Stealthy, and Persistent Backdoor Attacks Targeting Graph Foundation Models
PDF ↗Graph Foundation Models (GFMs) are pre-trained on diverse source domains and adapted to unseen targets, enabling broad generalization for graph machine learning. Despite that GFMs have attracted considerable attention recently, their vulnerability to backdoor attacks remains largely underexplored. A compromised GFM can introduce backdoor behaviors into downstream applications, posing serious security risks. However, launching backdoor attacks against GFMs is non-trivial due to three key challenges. (1) Effectiveness: Attackers lack knowledge of the downstream task during pre-training, complicating the assurance that triggers reliably induce misclassifications into desired classes. (2) Stealthiness: The variability in node features across domains complicates trigger insertion that remains stealthy. (3) Persistence: Downstream fine-tuning may erase backdoor behaviors by updating model parameters. To address these challenges, we propose GFM-BA, a novel Backdoor Attack model against Graph Foundation Models. Specifically, we first design a label-free trigger association module that links the trigger to a set of prototype embeddings, eliminating the need for knowledge about downstream tasks to perform backdoor injection. Then, we introduce a node-adaptive trigger generator, dynamically producing node-specific triggers, reducing the risk of trigger detection while reliably activating the backdoor. Lastly, we develop a persistent backdoor anchoring module that firmly anchors the backdoor to fine-tuning-insensitive parameters, enhancing the persistence of the backdoor under downstream adaptation. Extensive experiments demonstrate the effectiveness, stealthiness, and persistence of GFM-BA.
Continual learning (CL) empowers AI systems to progressively acquire knowledge from non-stationary data streams. However, catastrophic forgetting remains a critical challenge. In this work, we identify attention drift in Vision Transformers as a primary source of catastrophic forgetting, where the attention to previously learned visual concepts shifts significantly after learning new tasks. Inspired by neuroscientific insights into the selective attention in the human visual system, we propose a novel attention-retaining framework to mitigate forgetting in CL. Our method constrains attention drift by explicitly modifying gradients during backpropagation through a two-step process: 1) extracting attention maps of the previous task using a layer-wise rollout mechanism and generating instance-adaptive binary masks, and 2) when learning a new task, applying these masks to zero out gradients associated with previous attention regions, thereby preventing disruption of learned visual concepts. For compatibility with modern optimizers, the gradient masking process is further enhanced by scaling parameter updates proportionally to maintain their relative magnitudes. Experiments and visualizations demonstrate the effectiveness of our method in mitigating catastrophic forgetting and preserving visual concepts. It achieves state-of-the-art performance and exhibits robust generalizability across diverse CL scenarios.
SkipCat: Rank-Maximized Low-Rank Compression of Large Language Models via Shared Projection and Block Skipping
PDF ↗Large language models (LLM) have achieved remarkable performance across a wide range of tasks. However, their substantial parameter sizes pose significant challenges for deployment on edge devices with limited computational and memory resources. Low-rank compression is a promising approach to address this issue, as it reduces both computational and memory costs, making LLM more suitable for resource-constrained environments. Nonetheless, naïve low-rank compression methods require a significant reduction in the retained rank to achieve meaningful memory and computation savings. For a low-rank model, the ranks need to be reduced by more than half to yield efficiency gains. Such aggressive truncation, however, typically results in substantial performance degradation. To address this trade-off, we propose SkipCat, a novel low-rank compression framework that enables the use of higher ranks while achieving the same compression rates. First, we introduce an intra-layer shared low-rank projection method, where multiple matrices that share the same input use a common projection. This reduces redundancy and improves compression efficiency. Second, we propose a block skipping technique that omits computations and memory transfers for selected sub-blocks within the low-rank decomposition. These two techniques jointly enable our compressed model to retain more effective ranks under the same compression budget. Experimental results show that, without any additional fine-tuning, our method outperforms previous low-rank compression approaches by 7% accuracy improvement on zero-shot tasks under the same compression rate. These results highlight the effectiveness of our rank-maximized compression strategy in preserving model performance under tight resource constraints.
Beyond the Mean: Fisher-Orthogonal Projection for Natural Gradient Descent in Large Batch Training
PDF ↗Modern GPUs are equipped with large amounts of high-bandwidth memory, enabling them to support mini-batch sizes of up to tens of thousands of training samples. However, most existing optimizers struggle to perform effectively at such a large batch size. As batch size increases, gradient noise decreases due to averaging over many samples, limiting the ability of first-order methods to escape sharp or suboptimal minima and reach the global minimum. Meanwhile, second-order methods like the natural gradient with Kronecker-Factored Approximate Curvature (KFAC) often require excessively high damping to remain stable at large batch sizes. This high damping effectively ``washes out" the curvature information that gives these methods their advantage, reducing their performance to that of simple gradient descent. In this paper, we introduce Fisher-Orthogonal Projection (FOP), a novel technique that restores the effectiveness of the second-order method at very large batch sizes, enabling scalable training with improved generalization and faster convergence. FOP constructs a variance-aware update direction by leveraging gradients from two sub-batches, enhancing the average gradient with a component of the gradient difference that is orthogonal to the average under the Fisher-metric. Through extensive benchmarks, we show that FOP accelerates convergence by ×1.2–1.3 over K-FAC and ×1.5–1.7 over SGD/AdamW at the same moderate batch sizes, while at extreme scales it achieves up to a ×7.5 speedup. Unlike other methods, FOP maintains small-batch accuracy when scaling to extremely large batch sizes. Moreover, it reduces Top-1 error by 2.3–3.3% on long-tailed CIFAR benchmarks, demonstrating robust generalization under severe class imbalance. Our lightweight, geometry-aware use of intra-batch variance makes natural-gradient optimization practical on modern data-centre GPUs. FOP is open-source and pip-installable, which can be integrated into existing training code with a single line and no extra configuration.
This paper introduces a spiking-aided wifi sensing network (SWS-Net), a novel hybrid neural architecture that integrates Spiking Neural Networks (SNNs) with conventional Artificial Neural Networks (ANNs) for robust WiFi-based indoor sensing. WiFi signals offer a low-cost and device-free solution for recognizing human activities, gestures, identities and etc. However, their susceptibility to multipath fading and environmental noise poses significant challenges. Inspired by the human brain’s capability to process noisy information, SWS-Net leverages the noise-resilient dynamics of spiking neurons alongside the feature extraction ability of ANNs. We present a theoretical analysis comparing the noise-handling capacities of SNNs and ANNs, and show how their combination yields both improved robustness and training efficiency. Experimental results across three WiFi sensing tasks demonstrate that SWS-Net consistently achieves higher accuracy and faster convergence compared to baseline models, validating its effectiveness in challenging indoor environments.
STEP-Nav: Spatial-Temporal Efficient Visual Token Pruning for Vision-and-Language Navigation with Large Language Models
PDF ↗Vision-and-Language Navigation (VLN) plays a critical role in tasks of embodied AI, particularly in unseen environments following natural language instructions. Recent advancements leverage large language models (LLMs) to improve the accuracy and generalizability of VLN systems by encoding image sequences as dense token representations. However, this tokenization approach incurs substantial computational overhead due to two key inefficiencies: 1) ego-centric camera views often include navigation-irrelevant re- gions (e.g., sky or distant backgrounds), and 2) high-frame-rate image sequences introduce temporal redundancy. To address these challenges, we propose Spatial-Temporal Efficient Visual Token Pruning (STEP-Nav), a unified frame- work that simultaneously prunes redundant visual tokens and fine-tunes VLN models to preserve navigation performance. In particular, STEP-Nav incorporates a distance- and content-aware token evaluation mechanism to remove irrelevant tokens at the spatial level, along with temporal level similarity-based filtering to reduce redundancy across sequential frames. To ensure pruning does not harm task performance, we introduce a distortion-aware fine-tuning strategy that aligns pruned-token representations with their full-token counterparts while maintaining navigation accuracy. Experiments on the R2R and RxR benchmarks using Navid-CE and NavGPT-2 as base models demonstrate that STEP-Nav preserves over 95% of the performance while reducing 66.7% of tokens, outperforming existing token pruning baselines.