The recent success of large language models (LLMs) has sparked a growing interest in training large-scale models. As the model size continues to scale, concerns are growing about the depletion of high-quality, well-curated training data. This has led practitioners to explore training approaches like Federated Learning (FL), which can leverage the abundant data on edge devices while maintaining privacy. However, the decentralization of training datasets in FL introduces challenges to scaling large models, a topic that remains under-explored. This paper fills this gap and provides qualitative insights on generalizing the previous model scaling experience to federated learning scenarios. Specifically, we derive a PAC-Bayes (Probably Approximately Correct Bayesian) upper bound for the generalization error of models trained with stochastic algorithms in federated settings and quantify the impact of distributed training data on the optimal model size by finding the analytic solution of model size that minimizes this bound. Our theoretical results demonstrate that the optimal model size has a negative power law relationship with the number of clients if the total training compute is unchanged. Besides, we also find that switching to FL with the same training compute will inevitably reduce the upper bound of generalization performance that the model can achieve through training, and that estimating the optimal model size in federated scenarios should depend on the average training compute across clients. Furthermore, we also empirically validate the correctness of our results with extensive training runs on different models, network settings, and datasets.
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
XLinear: A Lightweight and Accurate MLP-Based Model for Long-Term Time Series Forecasting with Exogenous Inputs
PDF ↗Despite the prevalent assumption of uniform variable importance in long-term time series forecasting models, real-world applications often exhibit asymmetric causal relationships and varying data acquisition costs. Specifically, cost‐effective exogenous data (e.g., local weather) can unilaterally influence dynamics of endogenous variables, such as lake surface temperature. Exploiting these links enables more effective forecasts when exogenous inputs are readily available. Transformer-based models capture long-range dependencies but incur high computation and suffer from permutation invariance. Patch-based variants improve efficiency yet can miss local temporal patterns. To efficiently exploit informative signals across both the temporal dimension and relevant exogenous variables, this study proposes XLinear, a lightweight time series forecasting model built upon Multi-Layer Perceptrons (MLPs). XLinear uses a global token derived from an endogenous variable as a pivotal hub for interacting with exogenous variables, and employs MLPs with sigmoid activation to extract both temporal patterns and variate-wise dependencies. Its prediction head then integrates these signals to forecast the endogenous series. We evaluate XLinear on seven standard benchmarks and five real-world datasets with exogenous inputs. Compared with state-of-the-art models, XLinear delivers superior accuracy and efficiency for both multivariate forecasts and univariate forecasts influenced by exogenous inputs.
Prune&Comp: Free Lunch for Layer-Pruned LLMs via Iterative Pruning with Magnitude Compensation
PDF ↗Layer pruning is a viable technique for compressing large language models while achieving acceleration proportional to the pruning ratio. In this work, we identify that removing any layer induces a magnitude gap in hidden states, and demonstrate that a simple compensation operation leads to superior performance in iterative layer pruning. This key observation motivates us to propose Prune&Comp, a novel, plug-and-play iterative layer pruning scheme that leverages magnitude compensation to mitigate such gaps in a training-free manner. Specifically, we first estimate the magnitude gap of layer removal and then eliminate it by rescaling the remaining weights offline. We further demonstrate the advantages of Prune&Comp in improving the stability of iterative pruning. When integrated with an iterative prune-and-compensate loop, Prune&Comp consistently enhances existing layer pruning metrics. For instance, when 5 layers of LLaMA-3-8B are pruned with the prevalent Taylor+ metric, Prune&Comp reduces PPL from 512.78 to 16.34 and retains 90.57% of the original performance across 9 question-answering tasks, outperforming the baseline by 24.72%.
Images are generally represented by pixel intensities or color values, which are usually used as direct inputs for learning. This study innovatively proposes a geometric image representation method and refreshes the general learning model (e.g., autoencoder) in the diffeomorphic space. Based on the theory of geometric optimal transport and quasiconformal mapping, we equivalently transform the intensity representation into a shape representation. The image space becomes a diffeomorphic space, where any image can be uniquely represented as a Beltrami coefficient function defined on a uniform grid reference, and vice versa. This innovative geometric image representation (G-IR) captures the fine-grained structure inherent in the entire image, which is different from the traditional feature extraction that focuses on the internal geometric objects of the image (such as boundaries and axes). The diffeomorphic property preserves structure in the generation process, which is very necessary in the field of real physics. It can be assembled into existing pipelines as a plug-in, providing structure-preserving properties for the entire framework. Experiments on image restoration and interpolation validated the high efficiency, efficacy and applicability of the G-IR method, demonstrating its superior performance compared to common pixel-level image appearance representations.
GCA: Geometry-aware Conditional Alignment for Partial Domain Adaptation with Coding Rate Reduction
PDF ↗Partial Domain Adaptation (PDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain, where the target label space is a subset of the source label space. In PDA scenario, existing methods typically achieve transferability through distribution alignment in a statistical framework, and discriminability through geometric modeling. These two aspects are often treated as separate frameworks, which severs the intrinsic connection between them. To bridge this gap, we propose a unified framework termed Geometry-aware Conditional Alignment (GCA), which is derived from theoretical insights of Maximum Coding Rate Reduction. GCA collaboratively achieves conditional alignment and orthogonal discriminability in a unified framework, making the learned features more interpretable in both statistical and geometric aspects. As a result, GCA effectively enhances both the transferability and discriminability of features. Extensive experiments on four benchmark datasets validate the effectiveness of GCA.
Incomplete multi-view clustering (IMVC) aims to group data into meaningful clusters when each sample is only partially observed across multiple views. Most existing methods either rely on imputation strategies that may introduce noise and distort the underlying data distribution, or adopt cross-view alignment techniques that focus on pairwise relationships, often resulting in suboptimal representations and unstable clustering performance. In this paper, we propose Geometry-Aware Variational Information Maximization for Deep Incomplete Multi-view Clustering (GAVIM), a novel imputation-free variational framework that enables robust and coherent incomplete multi-view clustering. Specifically, GAVIM leverages mutual information maximization to preserve the high mutual information between the available multi-view data and the shared embedding. Moreover, we explicitly retain local geometric consistency within each view-specific latent space under the guidance of an adaptive global supervision signal. Lastly, GAVIM aligns all views simultaneously using a Gramian representation alignment measure, ensuring coherent structure across modalities and promoting unified, semantically meaningful representations. Extensive experiments on five benchmark IMVC datasets with varying levels of view incompleteness demonstrate that GAVIM consistently outperforms state-of-the-art methods in clustering accuracy and representation quality.
Federated causal discovery aims to uncover causal relationships while protecting data privacy, with significant real-world applications. Existing methods focus on horizontal federated settings where clients share the same variables but have different samples. However, in practice, clients may have different variables, leading to spurious causal relationships. To address this issue, we comprehensively consider causal structure learning methods under both horizontal and vertical federated settings. Interestingly, we find that, higher-order cumulants rely solely on the joint distribution of the relevant variables and are useful to solve the above problem in the linear non-Gaussian case. This motivates us to provide the identification theories for determining the causal order over observed variables, leveraging the difference in the product of the (cross) cumulants of the specific variables. Based on these theories, we develop a method for learning causal order in the horizontal and vertical federated scenarios. Specifically, we first obtain local (cross) cumulant matrices of observed variables from all participating clients to construct a global cumulant matrix. This global cumulant matrix is then used for recursive source variable identification, ultimately yielding a causal strength matrix of the union of variables from all clients. Our algorithm demonstrates superior performance in experiments on both synthetic and real-world data.
In natural scenarios, vision models often encounter the challenge of complex degradation scenarios(e.g., rain, snow, fog, or motion blur). These degradations severely corrupt image features, causing existing models to treat rarely seen or unseen degraded images as “unfamiliar”, thereby losing their inherent recognition and perception capabilities. To address this challenge, we propose a novel degradation disentanglement model (DDM) aimed at precisely disentangling degraded features from the image. The model enhances its perception of various degradations by controlling the matching of features across different degradation types and further strengthens the cross-correlation of target features by introducing a degradation suppression module. This enables the model to re-identify and re-localize targets while removing degradations. We validated the effectiveness of our method on more challenging few-shot segmentation datasets Degraded-Pascal and Degraded-COCO. Results on them outperform SOTA with 3.71% and 3.69% improvement respectively. The experimental results show that our method significantly improves the performance of vision models in various degradation scenarios and provides new ideas and solutions for visual understanding tasks in complex environments.
When Top-ranked Recommendations Fail: Modeling Multi-Granular Negative Feedback for Explainable and Robust Video Recommendation
PDF ↗Existing video recommendation systems, relying mainly on ID-based embedding mapping and collaborative filtering, often fail to capture in-depth video content semantics. Moreover, most struggle to address biased user behaviors (e.g., accidental clicks, fast skips), leading to inaccurate interest modeling and frequent negative feedback in top recommendations with unclear causes. To tackle this issue, we collect real-world user video-watching sequences, annotate the reasons for users' dislikes, and construct a benchmark dataset for personalized explanations. We then introduce the Agentic Explainable Negative Feedback (ENF) framework, which integrates three core components: (1) the Profile Agent, extracting behavioral cues from users' historical data to derive psychological and personality profiles; (2) the Video Agent, performing comprehensive multimodal video analysis; and (3) the Reason Agent, synthesizing information from the other two agents to predict user engagement and generate explanations. Additionally, we propose the S-GRPO algorithm, enabling the model to progressively address complex tasks during reinforcement fine-tuning. Experimental results on the collected dataset show that our method significantly outperforms state-of-the-art baselines in negative feedback prediction and reason explanation. Notably, it achieves an 8.6% improvement over GPT-4o in reason classification. Deployment on the business platform further validates its benefits: increasing average user watch time by 6.2%, reducing the fast-skip rate by 9.4% , and significantly enhancing user satisfaction.
One global model in federated learning (FL) might not be sufficient to serve many clients with non-IID tasks and distributions. Despite recent advances in FL to train multiple global models for better personalization, they only provide limited model choices to clients, so local finetuning of multiple models is still indispensable. This paper proposes a novel ``FedMerge'' approach that can create a single personalized model per client by simply merging multiple global models with automatically optimized and customized weights. We formulate this problem as a joint optimization of global models and the merging weights per client. Unlike existing FL approaches, where the server broadcasts one or multiple global models to all clients, the server only needs to send a customized, merged model to each client. Moreover, instead of periodically interrupting the local training and re-initializing it to a global model, the merged model aligns better with each client's task and data distribution, smoothening the local-global gap between consecutive rounds caused by client drift. We evaluate FedMerge on different non-IID settings applied to various domains with diverse tasks and data types, in which FedMerge consistently outperforms existing FL approaches, including clustering-based and mixture-of-experts (MoE) based methods.
Large Language Models (LLMs) fine-tuning techniques not only improve the adaptability to diverse downstream tasks, but also mitigate adverse effects of model quantization. Despite this, conventional quantization suffers from its structural limitation that hinders flexibility during the fine-tuning and deployment stages. Practical on-device tasks demand different quantization precisions (i.e. different bit-widths), e.g., understanding tasks tend to exhibit higher tolerance to reduced precision compared to generation tasks. Conventional quantization, typically relying on scaling factors that are incompatible across bit-widths, fails to support the on-device switching of precisions when confronted with complex real‑world scenarios. To overcome the dilemma, we propose OTARo, a novel method that enables on-device LLMs to flexibly switch quantization precisions while maintaining performance robustness through once fine-tuning. OTARo introduces Shared Exponent Floating Point (SEFP), a distinct quantization mechanism, to produce different bit-widths through simple mantissa truncations of a single model. Moreover, to achieve bit-width robustness in downstream applications, OTARo performs a learning process toward losses induced by different bit-widths. The method involves two critical strategies: (1) Exploitation-Exploration Bit-Width Path Search (BPS), which iteratively updates the search path via a designed scoring mechanism; (2) Low-Precision Asynchronous Accumulation (LAA), which performs asynchronous gradient accumulations and delayed updates under low bit-widths. Experiments on popular LLMs, e.g., LLaMA3.2-1B, LLaMA3-8B, demonstrate that OTARo achieves consistently strong and robust performance for all precisions.
CLIP (Contrastive Language-Image Pre-training) has attracted widespread attention for its multimodal generalizable knowledge, which is significant for downstream tasks. However, the computational overhead of a large number of parameters and large-scale pre-training poses challenges of pre-training a different scale of CLIP. Learngene extracts the generalizable components termed as learngene from an ancestry model and initializes diverse descendant models with it. Previous Learngene paradigms fail to handle the generalizable knowledge in multimodal scenarios. In this paper, we put forward the idea of utilizing a multimodal block to extract the multimodal generalizable knowledge, which inspires us to propose MM-LG (Multimodal Learngene), a novel framework designed to extract and leverage generalizable components from CLIP. Specifically, we first establish multimodal and unimodal blocks to extract the multimodal and unimodal generalizable knowledge in a weighted-sum manner. Subsequently, we employ these components to numerically initialize descendant models of varying scales and modalities. Extensive experiments demonstrate MM-LG's effectiveness, which achieves performance gains over existing learngene approaches (e.g.,+3.1% on Oxford-IIIT PET and +4.13% on Flickr30k) and comparable or superior results to the pre-training and fine-tuning paradigm (e.g.,+1.9% on Oxford-IIIT PET and +3.65% on Flickr30k). Notably, MM-LG requires only around 25% of the parameter storage while reducing around 2.8× pre-training costs for diverse model scales compared to the pre-training and fine-tuning paradigm, making it particularly suitable for efficient deployment across diverse downstream tasks.
Connectivity-Guided Sparsification of 2-FWL GNNs: Preserving Full Expressivity with Improved Efficiency
PDF ↗Higher-order Graph Neural Networks (HOGNNs) based on the 2-FWL test achieve superior expressivity by modeling 2-node and 3-node interactions, but incur cubic computational cost. Existing efficiency methods typically reduce this burden at the expense of expressivity. We propose Co-Sparsify, a connectivity-aware sparsification framework that eliminates provably redundant computations while preserving full 2-FWL expressive power. Our key insight is that 3-node interactions are expressively necessary only within biconnected components, namely, maximal subgraphs where every node pair lies on a cycle. Outside these components, structural relationships are fully captured via 2-node message passing and graph readouts, rendering higher-order modeling unnecessary. Co-Sparsify restricts 2-node message passing to connected components and 3-node interactions to biconnected components, eliminating redundant computation without approximation or sampling. We prove that Co-Sparsified GNNs match the expressivity of the 2-FWL test. Empirically, when applied to PPGN, Co-Sparsify matches or exceeds accuracy on synthetic substructure counting tasks and achieves state-of-the-art performance on real-world benchmarks (ZINC, QM9 and TUD). This study demonstrates that high expressivity and scalability are not mutually exclusive: principled, topology-guided sparsification enables powerful, efficient GNNs with theoretical guarantees.
Underwater object detection presents significant challenges due to the unique visual degradations in underwater environments, such as low contrast, poor visibility, and blurry object boundaries. While ANNs have achieved impressive detection accuracy, their high computational cost and power consumption limit their deployment in resource-constrained underwater platforms. In this work, we propose a Spatial-Frequency Spiking Neural Network (SFSNN) that combines the energy-efficient and event-driven nature of Spiking Neural Networks (SNNs) with the discriminative power of spatial-frequency analysis. SFSNN introduces a novel spatial-frequency spiking module that integrates spatial and frequency-domain representations, enhancing edge and texture features crucial for object detection in murky waters. Furthermore, we adapt the YOLOX architecture into a spike-based detector via ANN-to-SNN conversion using signed spiking neurons. Extensive experiments on the RUOD dataset demonstrate that SFSNN achieves superior performance over both SNN- and ANN-based detection models, offering a compelling solution for low-power underwater object detection.
Graph Neural Networks (GNNs) have received increasing attention due to their ability to handle graph-structured data, yet their explainability remains a significant challenge. An effective solution is to provide the GNN models with counterfactual explanations, which aim to answer “How should the input instance be perturbed to change the model's prediction?". However, existing works mainly focus on generating explanations that can effectively alter model predictions, while neglecting whether the explanations remain aligned with the original data distribution, leading to the distribution shift problem. To address this problem, we propose a novel method called ICExplainer for generating explanations within the original distribution. Specifically, we introduce graph diffusion-based generative model into the counterfactual reasoning, treating it as an optimization objective for graph distribution learning. Taking insights from variational inference, we use it to estimate the true distribution of the input graphs to retain essential structural and semantic information. The inferred distribution is then utilized as prior knowledge to guide the reverse process, ensuring that generated explanations are both counterfactual and distributionally coherent. Extensive experiments conducted on both synthetic and real-world datasets demonstrate the superior performance of ICExplainer over existing methods.
Chart editing reduces manual effort in visualization design. Typical benchmarks assume access to complete chart code, which is unrealistic for real-world applications. In this paper, we present ChartEditVista, a comprehensive benchmark consisting of 7,964 samples spanning 31 chart categories. It encompasses diverse editing instruction types and covers nearly all editable chart elements. The inputs in ChartEditVista include only the original chart image and natural language editing instructions, without the original chart codes. ChartEditVista is generated through a fully automated pipeline that produces, edits, and verifies charts, ensuring high-quality data. Besides, we introduce two novel fine-grained, rule-based evaluation metrics: the layout metric, which evaluates the position, size; and color of graphical components, and the text metric, which jointly assesses textual content and font styling. Building on top of ChartEditVista, we present ChartEditor, a model trained using a reinforcement learning framework that incorporates a novel rendering reward to simultaneously enforce code executability and visual fidelity. Through extensive experiments and human evaluations, we demonstrate that ChartEditVista provides a robust evaluation, while ChartEditor consistently outperforms models with similar-scale and larger-scale on chart editing tasks.
Mixture-of-Experts (MoE) is a sparse neural architecture that significantly increases model capacity while maintaining low computational complexity. However, deploying MoE-based large language models (LLMs) on memory-constrained edge devices remains challenging due to their substantial memory requirements. To address this issue, we propose FIRM-MoE, a fine-grained expert offloading framework designed to enable flexible and efficient MoE inference. The core insight of our approach is to reduce the risk of inaccurate expert loading by decomposing each expert into fine-grained sub-experts and then dynamically allocating them through a fine-grained scheduling strategy. To further reduce the error in expert loading, we introduce a multi-layer expert prediction mechanism and a resource-adaptive expert pre-loading algorithm to enable more robust expert allocation. This design allows our model to achieve more efficient expert utilization and improved resilience to prediction errors. We conduct extensive experiments to demonstrate the superiority of FIRM-MoE across diverse memory constraints. The results show that FIRM-MoE achieves up to 1.5× speedup and 2.8× memory savings in decoding, compared to state-of-the-art MoE offloading strategies.
Existing self-evolution methods overlook the influence of fine-grained reasoning steps, which leads to the reasoner-verifier gap. The computational inefficiency of Monte Carlo (MC) process supervision further exacerbates the difficulty in mitigating the gap. Motivated by the Error-Related Negativity (ERN), which the reasoner can localize error following incorrect decisions, guiding rapid adjustments, we propose a Self-Adaptive Process Optimization (SAPO) method for self-improvement in Small Language Models (SLMs). SAPO adaptively and efficiently introduces process supervision signals by actively minimizing the reasoner-verifier gap rather than relying on inefficient MC estimations. Extensive experiments demonstrate that the proposed method outperforms most existing self-evolution methods on two challenging task types: mathematics and code. Additionally, to further investigate SAPO's impact on verifier performance, this work introduces two new benchmarks for process reward models in both mathematical and coding tasks.
Recently, Omni-modal large language models (OLLMs) have sparked a new wave of research, achieving impressive results in tasks such as audio-video understanding and real-time environment perception. However, hallucination issues still persist. Similar to the bimodal setting, the priors from the text modality tend to dominate, leading OLLMs to rely more heavily on textual cues while neglecting visual and audio information. In addition, fully multimodal scenarios introduce new challenges. Most existing models align visual or auditory modalities with text independently during training, while ignoring the intrinsic correlations between video and its corresponding audio. This oversight results in hallucinations when reasoning requires interpreting hidden audio cues embedded in video content. To address these challenges, we propose OmniDPO, a preference-alignment framework designed to mitigate hallucinations in OLLMs. Specifically, OmniDPO incorporates two strategies: (1) constructing text-preference sample pairs to enhance the model’s understanding of audio-video interactions; and (2) constructing multimodal-preference sample pairs to strengthen the model’s attention to visual and auditory information. By tackling both challenges, OmniDPO effectively improves multimodal grounding and reduces hallucination. Experiments conducted on two OLLMs demonstrate that OmniDPO not only effectively mitigates multimodal hallucinations but also significantly enhances the models' reasoning capabilities across modalities.
Zero-shot Recommendation: Towards Class Semantic Relation Learning for Inferring Labels of Unseen Micro-videos
PDF ↗Micro-video label prediction plays a pivotal role on contemporary video-sharing platforms, such as Kwai and Tiktok. The emergence of video content lacking labels presents a formidable challenge for conventional user interest prediction methods. This paper addresses the challenge of micro-video label prediction, particularly for unseen videos, by proposing a zero-shot method called Class Semantic Relation Learning (CSRL). Unlike traditional user interest prediction models, CSRL leverages the pre-trained Large Language Model (LLM) to enhance prediction accuracy for unlabeled videos. The novelty of CSRL lies in its integration of three key components: a raw feature autoencoder, LLM-enhanced features, and a decomposed graph network. The decomposed graph network is specifically designed to disentangle the relationships between labeled and unlabeled videos, offering a significant improvement over previous methods. By fusing hidden topics with LLM-enhanced text, CSRL effectively handles sparse video features. Experiments on large-scale datasets from the Kwai platform show that CSRL achieves state-of-the-art results, with up to 44.64% improvement in Hit Ratio (HR), highlighting its superiority over existing zero-shot recommendation models in predicting user interests within the user-video network.