Pretraining large language models (LLMs) with next-token prediction has led to remarkable advances, yet the context-dependent nature of token embeddings in such models results in high intra-class variance and inter-class similarity, thus hindering the efficiency of representation learning. While similarity-based regularization has demonstrated benefit in supervised fine-tuning and classification tasks, its application and efficacy in large-scale LLM pretraining remains underexplored. In this work, we propose the SimReg, an embedding similarity regularization loss that explicitly encourages token representations with the same ground-truth label within each sequence to be more similar, while enforcing separation from different-label tokens via a contrastive loss. Our analysis reveals that this mechanism introduces gains by enlarging multi-classification margins, thereby enabling more efficient classification. Extensive experiments across dense and Mixture-of-Experts (MoE) architectures demonstrate that SimReg consistently accelerates training convergence by over 30

Variance reduction has been shown to improve the performance of Stochastic Gradient Descent (SGD) in centralized machine learning. However, when it is extended to federated learning systems, many issues may arise, including (i) mega-batch size settings; (ii) additional noise introduced by the gradient difference between the current iteration and the snapshot point; and (iii) gradient (statistical) heterogeneity. In this paper, we propose a lightweight algorithm termed federated adaptive batch size time evolving variance reduction (FedATEVR) to tackle these issues, consisting of an adaptive batch size setting scheme and a time-evolving variance reduction gradient estimator. In particular, we use the historical gradient information to set an appropriate mega-batch size for each client, which can steadily accelerate the local SGD process and reduce the computation cost. The historical information involves both global and local gradient, which mitigates unstable varying in mega-batch size introduced by gradient heterogeneity among the clients. For each client, the gradient difference between the current iteration and the snapshot point is used to tune the time-evolving weight of the variance reduction term in the gradient estimator. This can avoid meaningless variance reduction caused by the out-of-date snapshot point gradient. We theoretically prove that our algorithm can achieve a linear speedup of of $\mathcal {O}(\frac{1}{\sqrt{SKT}})$O(1SKT) for non-convex objective functions under partial client participation. Extensive experiments demonstrate that our proposed method can achieve higher test accuracy than the baselines and decrease communication rounds greatly.

As a prominent research topic, multi-view multi-label classification (MvMlC) aims to assign multiple labels to samples by integrating information from various perspectives. However, in real-world scenarios, MvMlC frequently faces the learning challenge of data with missing views and labels, typically resulting from sensor malfunctions, or the costly and time-consuming process of manual annotation. In addition, learning robust representations that are both consistent across views and specific to individual views remains a challenge. To address these issues, we propose a novel double incomplete multi-view multi-label classification framework based on Disentangling Consistent and Specific Information (DCSI). Specifically, we employ a dual-channel encoder with identical architecture but distinct objectives to extract cross-view consistent information and view-specific unique information from all views, respectively. Meanwhile, a view discriminator is constructed to decouple these two types of information, facilitating the extraction of pure consistent and specific information. Moreover, we meticulously design fusion strategies tailored to each representation type. Regarding consistent representations, we propose a dynamic-confidence-aware fusion mechanism that assesses the reliability of each view's representations in relation to the classification task, enabling the model to prioritize information from trustworthy representations. For specific representations, in light of their complementary rather than redundant property, we suggest treating such representations from each view equally to ensure fairness. Through experimental validation on five datasets, the results demonstrate that our method outperforms existing state-of-the-art methods.

Harmful fine-tuning can invalidate safety alignment of large language models, exposing significant safety risks. In this paper, we utilize the attention sink mechanism to mitigate harmful fine-tuning. Specifically, we first measure a statistic named for each attention head and observe that . To understand its safety implications, we conduct experiments and find that the number of attention heads of positive sink divergence increases along with the increase of the model's harmfulness when undergoing harmful fine-tuning. Based on this finding, we propose a separable sink divergence hypothesis -- . Based on the hypothesis, we propose a fine-tuning-stage defense, dubbed Surgery. Surgery utilizes a regularizer for sink divergence suppression, which steers attention heads toward the negative sink divergence group, thereby reducing the model’s tendency to learn and amplify harmful patterns. Extensive experiments demonstrate that Surgery improves defense performance by 5.90\%, 11.25\%, and 9.55\% on the BeaverTails, HarmBench, and SorryBench benchmarks, respectively. Source code is available on https://anonymous.4open.science/r/Surgery-A69E.

Continual learning capability is critical for Industrial LLMs, as deployed models must be continuously updated to meet evolving requirements and environments, rather than repeatedly retrained from scratch. However, most existing research focuses on improvements on static benchmarks, failing to capture real industrial needs. In this survey, we reformulate Industrial Continual Learning (ICL) for LLMs as a closed-loop update-and-release problem in a versioned ecosystem, where updates propagate hierarchically to industrial, application-specific models and LLM-powered applications, with capability inheritance and transfer across versions and model families. From this ecosystem perspective, we identify three core challenges: repeated adaptation erodes model plasticity, foundation-model upgrades break capability inheritance, and long-term sustainability is constrained by deployment requirements. We then organize the technical landscape of ICL around five lifecycle design principles: preserving plasticity headroom, treating upgrades as capability transfer, enabling trustworthy continual reinforcement learning, making training recipes self-optimizing, and building accountability as a base layer for long-term iteration. For each principle, we synthesize representative technical directions. Finally, we evaluate the maturity of each principle and its technical components via an evidence-based lens, identify key gaps hindering real-world deployment, and outline a practical ICL deployment blueprint and a pathway for feeding industrial realities back into academic research.

Graph clustering is a fundamental task in data analysis, aiming at grouping nodes with similar characteristics in the graph into clusters. This problem has been widely explored using graph neural networks (GNNs) due to their ability to leverage node attributes and graph topology for effective cluster assignments. However, representations learned through GNNs typically struggle to capture global relationships between nodes via local message-passing mechanisms. Moreover, the redundancy and noise inherently present in the graph data may easily result in node representations lacking compactness and robustness. To address the aforementioned issues, we propose a conjoint framework called CoCo, which captures compactness and consistency in the learned node representations for deep graph clustering. Technically, our CoCo leverages graph convolutional filters to learn robust node representations from both local and global views, and then encodes them into low-rank compact embeddings, thus effectively removing the redundancy and noise as well as uncovering the intrinsic underlying structure. To further enrich the node semantics, we develop a consistency learning strategy based on compact embeddings to facilitate knowledge transfer from the two perspectives. Our experimental findings indicate that our proposed CoCo outperforms state-of-the-art counterparts on various benchmark datasets.

With the increasing multimodal knowledge privatization requirements, multimodal knowledge graphs in different institutes are usually decentralized, lacking of effective collaboration system with both stronger reasoning ability and transmission safety guarantees. In this paper, we propose the Federated Multimodal Knowledge Graph Completion (FedMKGC) task, aiming at training over federated MKGs for better predicting the missing links in clients without sharing sensitive knowledge. We propose a framework named MMFeD3-HidE for addressing multimodal uncertain unavailability and multimodal client heterogeneity challenges of FedMKGC. (1) Inside the clients, our proposed Hyper-modal Imputation Diffusion Embedding model (HidE) recovers the complete multimodal distributions from incomplete entity embeddings constrained by available modalities. (2) Among clients, our proposed Multimodal FeDerated Dual Distillation (MMFeD3) transfers knowledge mutually between clients and the server with logit and feature distillation to improve both global convergence and semantic consistency. We propose a FedMKGC benchmark for a comprehensive evaluation, consisting of a general FedMKGC backbone named MMFedE, datasets with heterogeneous multimodal information, and three groups of constructed baselines. Experiments conducted on our benchmark validate the effectiveness, semantic consistency, and convergence robustness of MMFeD3-HidE.

Continual learning (CL) focuses on learning non-stationary data distribution without forgetting previous knowledge. The most widely used memory-replay approaches are often prone to memory overfitting due to the limited memory diversity and hardness. Existing work mitigating memory overfitting either lacks data diversity or hardness or is hard to train. To address the above limitations and release the memory buffer potential, we view the memory buffer transformation from a new dynamic system perspective and propose a continuous and reversible memory transformation method. We introduce an adversarial optimization objective that jointly learns the CL model and memory transformer. Specifically, we present a deterministic continuous memory transformer (DCMT) to generate diverse memory data. Furthermore, we inject uncertainty into the transformation function and develop a stochastic continuous memory transformer (SCMT), which substantially enhances the diversity of the transformed memory buffer. The presented neural transformation approaches have significant advantages over existing ones: (1) they significantly increase the memory buffer diversity and hardness to overfit; (2) they are memory efficient without needing to make a replica of the memory data. Extensive experiments show a significant improvement with our approach compared to strong baselines.

address the communication challenges associated with Federated Learning (FL), Decentralized Federated Learning (DFL) eliminates the central server and trains the model with decentralized method, enabling each client to only communicate with its neighbors. However, per our analysis, model trained with DFL experiences performance degradation because of data-heterogeneity and time-varying topologies. To address these issues, we propose a Dynamic K-step Gradient Tracking (DKGT) method to enhance the performance of DFL over time varying networks. Specifically, DKGT employs K-step local updates and gradient tracking to reduce the communication cost and the variance from heterogeneous data distribution, and we use dynamic gradient tracking parameter to correct gradient over time varying graph. Theoretically, we derive a universal convergence rate for smooth and non-convex problem at the rate of O( (f(x(0))-f(x(& lowast;)))/root T(L root KN )(-1)-tau(pKL root TKN)(-1) + sigma(2)/KTN tau(pK-tau)), that tau and p and p respectively represent the time window length and the connectivity of time-varying networks. Experimentally, we illustrate the robustness and effectiveness of this heterogeneity correction on extensive non-convex neural network training tasks over different topologies and dynamic network settings.

Industrial anomaly detection demands precise reasoning over fine-grained defect patterns. However, existing multimodal large language models (MLLMs), pretrained on general-domain data, often struggle to capture category-specific anomalies, thereby limiting both detection accuracy and interpretability. To address these limitations, we propose Reason-IAD, a knowledge-guided dynamic latent reasoning framework for explainable industrial anomaly detection. Reason-IAD comprises two core components. First, a retrieval-augmented knowledge module incorporates category-specific textual descriptions into the model input, enabling context-aware reasoning over domain-specific defects. Second, an entropy-driven latent reasoning mechanism conducts iterative exploration within a compact latent space using optimizable latent think tokens, guided by an entropy-based reward that encourages confident and stable predictions. Furthermore, a dynamic visual injection strategy selectively incorporates the most informative image patches into the latent sequence, directing the reasoning process toward regions critical for anomaly detection. Extensive experimental results demonstrate that Reason-IAD consistently outperforms state-of-the-art methods across multiple tasks. The code will be publicly available at https://github.com/chenpeng052/Reason-IAD.

Model merging is an efficient empowerment technique in the machine learning community that does not require the collection of raw training data and does not require expensive computation. As model merging becomes increasingly prevalent across various fields, it is crucial to understand the available model merging techniques comprehensively. However, there is a significant gap in the literature regarding a systematic and thorough review of these techniques. This survey provides a comprehensive overview of model merging methods and theories, their applications in various domains and settings, and future research directions. Specifically, we first propose a new taxonomic approach that exhaustively discusses existing model merging methods. Secondly, we discuss the application of model merging techniques in large language models, multimodal large language models, and more than ten machine learning subfields, including continual learning, multi-task learning, few-shot learning, and so on. Finally, we highlight the remaining challenges of model merging and discuss future research directions. A comprehensive list of papers about model merging is available at https://github.com/ EnnengYang/Awesome-Model-Merging-Methods-Theories-Applications.

Reinforcement finetuning (RFT) has garnered significant attention in recent years, particularly for enhancing large reasoning models such as OpenAI o1 and Deepseek R1. The appeal of RFT largely stems from its ability to refine model knowledge, better align outputs with user intent, and address challenges associated with limited finetuning data. Despite these advantages, the application of RFT in large Transformer-based generative agents remains relatively underexplored. Although these agents are designed to address multiple tasks through large-scale autoregressive pretraining and share many properties with large reasoning models, current adaptation strategies predominantly rely on supervised finetuning (SFT). In this work, we conduct a systematic investigation of several RFT techniques across a variety of finetuning parameter configurations and meta-reinforcement learning (meta-RL) environments, employing few-shot offline datasets. We provide a comprehensive analysis of RFT algorithm performance under diverse experimental conditions and, based on our empirical findings, introduce a lightweight enhancement to existing RFT methods. This enhancement consistently improves outcomes by combining the strengths of both SFT and RFT. Our findings provide valuable insights for advancing the effectiveness of RFT approaches and broadening their applicability to meta-RL tasks with large Transformer-based generative agents, motivating further research in broader domains.

Integrating Large Language Models (LLMs) with external tools via multi-agent systems offers a promising new paradigm for decomposing and solving complex problems. However, training these systems remains notoriously difficult due to the credit assignment challenge, as it is often unclear which specific functional agent is responsible for the success or failure of decision trajectories. Existing methods typically rely on sparse or globally broadcast rewards, failing to capture individual contributions and leading to inefficient reinforcement learning. To address these limitations, we introduce the Shapley-based Hierarchical Attribution for Reinforcement Policy (SHARP), a novel framework for optimizing multi-agent reinforcement learning via precise credit attribution. SHARP effectively stabilizes training by normalizing agent-specific advantages across trajectory groups, primarily through a decomposed reward mechanism comprising a global broadcast-accuracy reward, a Shapley-based marginal-credit reward for each agent, and a tool-process reward to improve execution efficiency. Extensive experiments across various real-world benchmarks demonstrate that SHARP significantly outperforms recent state-of-the-art baselines, achieving average match improvements of 23.66

Instruction tuning plays a critical role in enhancing the performance and efficiency of Large Language Models (LLMs). Its success depends not only on the quality of the instruction data but also on the inherent capabilities of the LLM itself. Some studies suggest that even a small amount of high-quality data can achieve instruction fine-tuning results that are on par with, or even exceed, those from using a full-scale dataset. However, rather than focusing solely on calculating data quality scores to evaluate instruction data, there is a growing need to select high-quality data that maximally enhances the performance of instruction tuning for a given LLM. In this paper, we propose the Model Instruction Weakness Value (MIWV) as a novel metric to quantify the importance of instruction data in enhancing model's capabilities. The MIWV metric is derived from the discrepancies in the model's responses when using In-Context Learning (ICL), helping identify the most beneficial data for enhancing instruction tuning performance. Our experimental results demonstrate that selecting only the top 1% of data based on MIWV can outperform training on the full dataset. Furthermore, this approach extends beyond existing research that focuses on data quality scoring for data selection, offering strong empirical evidence supporting the effectiveness of our proposed method.

Push-Sum-based decentralized learning enables optimization over directed communication networks, where information exchange may be asymmetric. While convergence properties of such methods are well understood, their finite-iteration stability and generalization behavior remain unclear due to structural bias induced by column-stochastic mixing and asymmetric error propagation. In this work, we develop a unified uniform-stability framework for the Stochastic Gradient Push (SGP) algorithm that captures the effect of directed topology. A key technical ingredient is an imbalance-aware consistency bound for Push-Sum, which controls consensus deviation through two quantities: the stationary distribution imbalance parameter δ and the spectral gap (1-λ) governing mixing speed. This decomposition enables us to disentangle statistical effects from topology-induced bias. We establish finite-iteration stability and optimization guarantees for both convex objectives and non-convex objectives satisfying the Polyak–Łojasiewicz condition. For convex problems, SGP attains excess generalization error of order 𝒪̃(1/√(mn)+δ(1-λ)+γ) under step-size schedules, and we characterize the corresponding optimal early stopping time that minimizes this bound. For PŁ objectives, we obtain convex-like optimization and generalization rates with dominant dependence proportional to κ(1+1/δ(1-λ)), revealing a multiplicative coupling between problem conditioning and directed communication topology. Our analysis clarifies when Push-Sum correction is necessary compared with standard decentralized SGD and quantifies how imbalance and mixing jointly shape the best attainable learning performance.

Supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) are two widely used post-training paradigms for improving the reasoning ability of large language models (LLMs). Recent methods attempt to integrate SFT and RLVR in a single stage by reweighting or scheduling their objectives. However, such coupling can be counterproductive because supervised updates are not uniformly beneficial for reward optimization, which can diminish reward gains. To address this, we propose \textsc{BRIDGE}, a scalable framework in which SFT learns to supervise RL by selectively transferring knowledge that improves reward optimization. Specifically, \textsc{BRIDGE} employs two nested optimization loops during meta-training: the inner loop updates base model parameters using a fused SFT--RL gradient. Concurrently, the outer loop updates a lightweight low-rank adapter (LoRA) to coordinate the two objectives by maximizing a reward-gap signal, defined as the reward of joint SFT--RL training over an RL-only baseline. Across three model scales and five reasoning benchmarks, \textsc{BRIDGE} consistently outperforms two-stage cold start, naive mixing, and representative single-stage integration baselines, yielding over three points average absolute improvement and more stable training dynamics.

Compared with general Federated Learning (FL), Decentralized FL (DFL) has diminished central communication burdens and lower risks of disruption. To be compatible with real-world scenarios, some existing works produce several local personalized models rather than a universal model for all edge devices. However, they still suffer from inferior performance due to the full model aggregation in heterogeneous datasets. Therefore, we propose a DFL framework DFedMDC through partial model personalization, which can adapt to resource-heterogeneous environments (e.g., the Internet of Things (IoT)). Specifically, it personalizes the "right" components of local models and trains the shared model u(i) and personal model v(i) alternately in each client i. To further accelerate the convergence process, we propose DFedSMDC with properly directed noise perturbation into the gradient update. Theoretically, we provide convergence analysis of both algorithms in the general non-convex setting. It can shed light on how vital factors affect the convergence rate, such as these alternate updates of partial models u(i), v(i), data heterogeneity delta(2) as well as various communication topologies (characterized by the spectral gap lambda). Empirically, we confirm the state-of-the-art (SOTA) superiority of the proposed methods on several real-world datasets with various data distributions, relative to both SOTA personalized FL (PFL) and DFL baselines.

Autonomous agents have recently achieved remarkable progress across diverse domains, yet most evaluations focus on short-horizon, fully observable tasks. In contrast, many critical real-world tasks, such as large-scale software development, commercial investment, and scientific discovery, unfold in long-horizon and partially observable scenarios where success hinges on sustained reasoning, planning, memory management, and tool use. Existing benchmarks rarely capture these long-horizon challenges, leaving a gap in systematic evaluation. To bridge this gap, we introduce \textbf{UltraHorizon} a novel benchmark that measures the foundational capabilities essential for complex real-world challenges. We use exploration as a unifying task across three distinct environments to validate these core competencies. Agents are designed in long-horizon discovery tasks where they must iteratively uncover hidden rules through sustained reasoning, planning, memory and tools management, and interaction with environments. Under the heaviest scale setting, trajectories average \textbf{200k+} tokens and \textbf{400+} tool calls, whereas in standard configurations they still exceed \textbf{35k} tokens and involve more than \textbf{60} tool calls on average. Our extensive experiments reveal that LLM-agents consistently underperform in these settings, whereas human participants achieve higher scores, underscoring a persistent gap in agents' long-horizon abilities. We also observe that simple scaling fails in our task. To better illustrate the failure of agents, we conduct an in-depth analysis of collected trajectories. We identify eight types of errors and attribute them to two primary causes: in-context locking and functional fundamental capability gaps. \href{https://github.com/StarDewXXX/UltraHorizon}{Our code will be available here.}

Zeroth-Order optimization presents a promising memory-efficient paradigm for fine-tuning Large Language Models by relying solely on forward passes. However, its practical adoption is severely constrained by slow wall-clock convergence and high estimation variance. In this work, we dissect the runtime characteristics of ZO algorithms and identify a critical system bottleneck where the generation of perturbations and parameter updates accounts for over 40% of the training latency. We argue that the standard uniform exploration strategy is fundamentally flawed as it fails to account for the heterogeneous sensitivity of layers in deep networks, resulting in computationally wasteful blind searches. To address this structural mismatch, we propose AdaLeZO, an Adaptive Layer-wise ZO optimization framework. By formulating the layer selection process as a non-stationary Multi-Armed Bandit problem, AdaLeZO dynamically allocates the limited perturbation budget to the most sensitive parameters. We further introduce an Inverse Probability Weighting mechanism based on sampling with replacement, which guarantees unbiased gradient estimation while effectively acting as a temporal denoiser to reduce variance. Extensive experiments on LLaMA and OPT models ranging from 6.7B to 30B parameters demonstrate that AdaLeZO achieves 1.7x to 3.0x wall-clock acceleration compared to state-of-the-art methods. Crucially, AdaLeZO functions as a universal plug-and-play module that seamlessly enhances the efficiency of existing ZO optimizers without incurring additional memory overhead.

Recent progress in reasoning capabilities of Multimodal Large Language Models(MLLMs) has highlighted their potential for performing complex video understanding tasks. However, in the domain of Video Anomaly Detection and Understanding (VAD U), existing MLLM-based methods are largely limited to anomaly localization or post-hoc description, lacking explicit reasoning processes, risk awareness, and decision-oriented interpretation. To address this gap, we define a new task termed Video Anomaly Reasoning (VAR), which elevates video anomaly analysis from descriptive understanding to structured, multi-stage reasoning. VAR explicitly requires models to perform progressive reasoning over anomalous events before answering anomaly-related questions, encompassing visual perception, causal interpretation, and risk-aware decision making. To support this task, we present a new dataset with 8,641 videos, where each video is annotated with diverse question types corresponding to different reasoning depths, totaling more than 50,000 samples, making it one of the largest datasets for video anomaly. The annotations are based on a structured Perception-Cognition-Action Chain-of-Thought (PerCoAct-CoT), which formalizes domain-specific reasoning priors for video anomaly understanding. This design enables systematic evaluation of multi-stage and adaptive anomaly reasoning. In addition, we propose Anomaly-Aware Group Relative Policy Optimization to further enhance reasoning reliability under weak supervision. Building upon the proposed task and dataset, we develop an end-to-end MLLM-based VAR model termed Vad-R1-Plus, which supports adaptive hierarchical reasoning and risk-aware decision making. Extensive experiments demonstrate that the proposed benchmark and method effectively advance the reasoning capabilities of MLLMs on VAR tasks, outperforming both open-source and proprietary baselines.