Sequential recommendation (SR) has received much attention and made promising progress in the past few years due to its high alignment with real recommendation scenarios. It models users' preferences and behavior patterns from their historical behavior sequences and provides personalized recommendations. However, the widespread problem of data sparsity limits the performance of sequential recommendation models. To tackle this, data augmentation (DA) provides a feasible solution by improving the quantity, quality, or diversity of the training samples without the need for additional data collection. In this survey, we present a systematic and timely review of research efforts on data augmentation for sequential recommendation. We start by providing a clear formulation of the problem and task. Then, we develop a unified taxonomy that categorizes existing augmentation methodologies regarding their augmentation objects and principles. Next, we conduct a comparative discussion on the advantages and disadvantages of different categories, supplemented with quantitative performance evaluations, time-complexity analyses, and visual case studies of representative methods, aiming to provide actionable guidance for the selection and development of augmentation methods in real-world scenarios. Finally, we present the future research directions and summarize this survey.

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.

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.

Test-time reinforcement learning generates multiple candidate answers via repeated rollouts and performs online updates using pseudo-labels constructed by majority voting. To reduce overhead and improve exploration, prior work introduces tree-structured rollouts, which share reasoning prefixes and branch at key nodes to improve sampling efficiency. However, this paradigm still faces two challenges: (1) high-entropy branching can trigger rollout collapse, where the branching budget concentrates on a few trajectories with consecutive high-entropy segments, rapidly reducing the number of effective branches; (2) early pseudo-labels are noisy and biased, which can induce self-reinforcing overfitting, causing the policy to sharpen prematurely and suppress exploration. To address these issues, we propose Entropy–Confidence Hybrid Group Relative Policy Optimization (ECHO). During rollout, ECHO jointly leverages local entropy and group-level confidence to adaptively control branch width, and further introduces online confidence-based pruning to terminate persistently low-confidence branches, avoiding high-entropy traps and mitigating collapse. During policy updates, ECHO employs confidence-adaptive clipping and an entropy–confidence hybrid advantage shaping approach to enhance training robustness and mitigate early-stage bias. Experiments demonstrate that ECHO achieves consistent gains on multiple mathematical and visual reasoning benchmarks, and generalizes more effectively under a limited rollout budget.

Logical reasoning-based recommendation methods formulate logical expressions to characterize user-item interaction patterns, incorporating regularization constraints to ensure consistency with logical rules. However, these methods face two critical challenges: (1) As sequence length increases, they cannot effectively capture the dynamic transfer of user interests across subsequences (i.e., subsequence interest drift), thereby degenerating logical expressions to single-subsequence inference. (2) The time complexity of logical reasoning and rule learning scales quadratically with the sequence length, severely constraining computational efficiency in long-sequence recommendation. To address these challenges, we propose ELECTOR, an intErest-shift-aware long-sequence Logical reasoning for EffiCienT lOng-sequence Recommendation method. Specifically, we design a Subsequence Interest Learning Module (SIL) to model cross-subsequence interest drifts in long sequences. SIL employs a local attention mechanism to extract subsequence interests effectively and a global attention mechanism to capture the correlations among subsequence interests. Subsequently, we propose an Interest-aware Logical Reasoning (ILR) mechanism that performs logical reasoning using a limited set of subsequence and short-term interests, rather than reasoning over the entire sequence, significantly reducing time complexity. Additionally, ILR employs interest logical reasoning contrastive loss to ensure the model simultaneously considers multiple interests. Experiments on four real-world datasets demonstrate that our method significantly outperforms all baselines regarding computational efficiency and recommendation accuracy, confirming its effectiveness.

Data augmentation is an effective technique for tackling data sparsity in sequential recommendation (SR). Existing methods generate new data during the model training to improve the performance. However, deploying them on a backbone model requires retraining, architecture modification, or introducing additional modules and learnable parameters. These processes are time-consuming and costly for well-trained models, especially when the model and data scales become large. In this work, we explore the test-time augmentation (TTA) for SR, which augments the input sequences during the inference phase and then fuses the model's predictions to improve final accuracy. It avoids the significant overhead associated with training-time augmentation. We first experimentally examine the potential of existing augmentation operators for TTA and find that the Substitute and Mask consistently achieve better performance. Further analysis reveals that these two operators retain the original sequential pattern while adding appropriate perturbations. Moreover, the random selection of augmentation positions creates suitable augmented samples from both semantic and temporal perspectives. Meanwhile, we find that the fixed operation ratio limits the diversity of augmented data, and the TTA may impair the model's performance on long sequences. In addition, the two operators still face time-consuming similarity-based item selection or interference from mask tokens. Based on the analysis and limitations, we present TNoise and TMask. The former injects uniform noise into the representation, avoiding the computational overhead of item selection. The latter blocks mask tokens from participating in model calculations (TMask-B) or directly removes interactions that should have been replaced with mask tokens (TMask-R). Further, we sample the augmentation ratio from a uniform distribution to improve the data diversity. For short sequences, we introduce a sequence smoothing and lengthening method based on inter-item interpolation. For long sequences, we set a threshold to avoid the negative effects of TTA. Comprehensive experiments demonstrate the effectiveness, efficiency, and generalizability of our method.

Padding is a standard technique for training sequential recommendation models. It fills the short sequences to the same length so they can be fed into the model in batches. Traditional zero-padding uses the special value 0 as the padding content. However, zero-padding results in a significant waste of input space due to widespread data sparsity and the fact that the special value 0 is not involved in the model calculations. To tackle that, researchers propose repeated padding (RepPad) [1], which uses the original sequence as the padding content during training. Although this method turns out to be effective on many different types of sequential recommendation models, what underlies the performance gains is still a mystery. In this paper, we first review the operations of RepPad for input, positive, and negative sequences. Based on the operations, we decompose RepPad into twelve variants ranging from zero padding to full RepPad. Through an empirical study, we disclose that RepPad's effectiveness derives from performing repeated padding on input, positive sequences, and the extended random negative sequence. These three operations must exist simultaneously. Furthermore, we derive and analyze the twelve variants step-by-step from a loss function perspective, which helps clarify the changes brought about by RepPad compared to zero-padding during loss calculations. Based on these analyses, we refine the previous three operations into three reasons why RepPad is effective: maintaining sequence alignment and causality, increasing information density and diversity, and generating more negative samples in the single propagation. Based on these findings, we put forward the Repeated Padding with Extended Negative Sequence (RepPad-ENS) for sequential recommendation, which samples multiple extended random negative sequences based on RepPad. It unlocks the potential of RepPad by improving the quantity and diversity of negative samples in the single prop agation, thereby facilitating more accurate preference learning. Comprehensive experiments on various categories of baselines with real-world datasets demonstrate the effectiveness, efficiency, and generalizability of our method. Our code is available at https://github.com/KingGugu/RepPad-ENS.

Multi-task learning (MTL) leverages a shared model to accomplish multiple tasks and facilitate knowledge transfer. Recent research on task arithmetic-based MTL demonstrates that merging the parameters of independently fine-tuned models can effectively achieve MTL. However, existing merging methods primarily seek a static optimal solution within the original model parameter space, which often results in performance degradation due to the inherent diversity among tasks and potential interferences. To address this challenge, in this paper, we propose a Weight-Ensembling Mixture of Experts (WEMoE) method for multi-task model merging. Specifically, we first identify critical (or sensitive) modules by analyzing parameter variations in core modules of Transformer-based models before and after finetuning. Then, our WEMoE statically merges non-critical modules while transforming critical modules into a mixture-of-experts (MoE) structure. During inference, expert modules in the MoE are dynamically merged based on input samples, enabling a more flexible and adaptive merging approach. Building on WEMoE, we further introduce an efficient-and-effective WEMoE (E-WEMoE) method, whose core mechanism involves eliminating non-essential elements in the critical modules of WEMoE and implementing shared routing across multiple MoE modules, thereby significantly reducing both the trainable parameters, the overall parameter count, and computational overhead of the merged model by WEMoE. Experimental results across various architectures and tasks demonstrate that both WEMoE and E-WEMoE outperform state-of-the-art (SOTA) model merging methods in terms of MTL performance, generalization, and robustness.

Model merging (MM) offers an efficient mechanism for integrating multiple specialized models without access to original training data or costly retraining. While MM has demonstrated success in domains like computer vision, its role in recommender systems (RSs) remains largely unexplored. Recently, Generative Recommendation (GR) has emerged as a new paradigm in RSs, characterized by rapidly growing model scales and substantial computational costs, making MM particularly appealing for cost-sensitive deployment scenarios. In this work, we present the first systematic study of MM in GR through a contextual lens. We focus on a fundamental yet underexplored challenge in real-world: how to merge generative recommenders specialized to different real-world contexts, arising from temporal evolving user behaviors and heterogeneous application domains. To this end, we propose a unified framework MMGRid, a structured contextual grid of GR checkpoints that organizes models trained under diverse contexts induced by temporal evolution and domain diversity. All checkpoints are derived from a shared base LLM but fine-tuned on context-specific data, forming a realistic and controlled model space for systematically analyzing MM across GR paradigms and merging algorithms. Our investigation reveals several key insights. First, training GR models from LLMs can introduce parameter conflicts during merging due to token distribution shifts and objective disparities; such conflicts can be alleviated by disentangling task-aware and context-specific parameter changes via base model replacement. Second, incremental training across contexts induces recency bias, which can be effectively balanced through weighted contextual merging. Notably, we observe that optimal merging weights correlate with context-dependent interaction characteristics, offering practical guidance for weight selection in real-world deployments.

Direct Preference Optimization (DPO) guides large language models (LLMs) to generate recommendations aligned with user historical behavior distributions by minimizing preference alignment loss. However, our systematic empirical research and theoretical analysis reveal that DPO tends to amplify spurious correlations caused by environmental confounders during the alignment process, significantly undermining the generalization capability of LLM-based generative recommendation methods in out-of-distribution (OOD) scenarios. To mitigate this issue, we propose CausalDPO, an extension of DPO that incorporates a causal invariance learning mechanism. This method introduces a backdoor adjustment strategy during the preference alignment phase to eliminate interference from environmental confounders, explicitly models the latent environmental distribution using a soft clustering approach, and enhances robust consistency across diverse environments through invariance constraints. Theoretical analysis demonstrates that CausalDPO can effectively capture users' stable preference structures across multiple environments, thereby improving the OOD generalization performance of LLM-based recommendation models. We conduct extensive experiments under four representative distribution shift settings to validate the effectiveness of CausalDPO, achieving an average performance improvement of 24.10\% across four evaluation metrics.

Continual learning with large pre-trained models offers significant potential for cross-task knowledge accumulation, but faces critical challenges such as catastrophic forgetting and parameter interference, especially when historical data is unavailable. Existing approaches typically rely on sequential fine-tuning or model merging strategies, yet often overlook the impact of loss landscape sharpness and dominant singular value directions, which leads to subspace misalignment and severe knowledge forgetting. In this paper, we propose the Sharpness-Aware Isotropic Merging (SAIM) framework, which introduces targeted optimizations in both the fine-tuning and merging stages to address these issues. Specifically, SAIM consists of two synergistic modules: (1) a Sharpness-Aware Block Coordinate Descent (SA-BCD) optimizer that guides the model toward flatter minima and selectively updates the most task-sensitive parameters, thereby mitigating parameter interference and enhancing robustness; (2) an adaptive isotropic merging algorithm that dynamically balances the singular value spectrum across tasks, effectively preventing the model from overemphasizing any single task direction, maintaining balanced knowledge representation, and improving subspace alignment. Extensive experiments on vision and language benchmarks demonstrate that SAIM achieves 5-10% higher accuracy than existing methods and maintains robust performance as the number of tasks increases. Ablation studies further validate the effectiveness of the SA-BCD fine-tuning strategy in promoting flat minima and reducing parameter interference, as well as its compatibility with various merging approaches.

Model merging aims to integrate multiple independently fine-tuned expert models into a single model while preserving the knowledge of all experts. However, existing approaches mainly address parameter conflicts at the merging stage and overlook the role of the fine-tuning process, which often leads to significant post-merge performance degradation. To address this limitation, we propose a novel merging-aware optimizer (abbreviated as MergOPT) that injects principled merge-induced parameter shifts into the weight update steps so that the fine-tuned model exhibits a more stable loss landscape under subsequent merging operations. Specifically, we first formulate model merging as a distributionally robust optimization problem in the weight space: the parameters of other experts to be merged are viewed as adversarial merge-offsets, and fine-tuning adapts to the worst-case merging scenario. Building on this formulation, we analyze the distribution of parameter updates and the effects of merging hyperparameters, from which we derive a merging-guided feasible region for weight shifts. Finally, extensive experiments across four large language models (LLMs) and one vision model show that our approach consistently outperforms standard fine-tuning, yielding an average relative gain of 3.5\% and a maximum gain of 9.5\% across four merging strategies when merging seven experts.

Custom diffusion models (CDMs) have demonstrated impressive success in visual personalization tasks by enabling the generation of user-specific concepts. However, existing CDMs typically assume that personalized concepts are static and rely on costly model merging or sequential updates that are prone to catastrophic forgetting as new concepts are introduced. To address these limitations, we propose a Submodular Concept Neuron Selection method (SCNS), to solve CDMs with continual personalized concepts, which formulates continual personalization as a constrained submodular optimization problem to select a minimal yet sufficient set of concept-specific neurons under diminishing returns. SCNS combines a Facility Location-based coverage objective to suppress semantic redundancy, a Fisher-weighted risk proxy to protect previously learned concepts, and a cost-aware greedy rule to balance stability and plasticity with extreme sparsity. Extensive experiments demonstrate that SCNS achieves state-of-the-art performance in image alignment and anti-forgetting, while enabling fusion-free continual personalization by modifying only 0.41% of the total parameters for each concept on average. Our implementation is available at SCNS.

Data augmentation mitigates data sparsity in sequential recommendations by generating new yet effective data. Most existing work focuses on a single original sequence with item-level augmentations. It ignores the correlations between different users (sequences) and struggles to produce diverse yet reasonable new data across sequences, leading to limited performance improvements. Also, the item-level operation may destroy the integrity of the preference knowledge contained in the original sequences, resulting in incomplete preference learning. In this work, we propose a novel user correlation guided cross-sequence mixing plug-in for sequential recommendation (UCMRec). Our core idea is to select the appropriate sequence based on the user correlation graph and perform cross-sequence mixup operations at the representation level to generate high-quality samples. Specifically, we construct a user-user graph based on joint interactions and perform two types of searches to get the candidate users with different correlations. Then, we introduce a mixup operation at the sequence representation level to generate diverse yet reasonable samples across sequences. Furthermore, we propose a topology-aware reweighting module to enable the model to adjust the learning intensity based on preference similarity. Comprehensive experiments demonstrate the superiority of our method.

Foundation models update slowly due to resource-intensive training, whereas domain-specific models evolve rapidly between releases. Model merging seeks to combine multiple expert models into a single, more capable model, reducing storage and serving costs while supporting decentralized development. Despite its potential, previous studies have primarily focused on merging visual classification models or Large Language Models (LLMs) for code and math tasks. Recently, Multimodal LLMs (MLLMs) that extend LLMs through large-scale multimodal training have gained traction. However, no benchmark exists for model merging research that clearly divides the tasks of MLLM training and evaluation. In this paper, $(i)$ we introduce a model merging benchmark for MLLMs, which includes multiple tasks such as VQA, Geometry, Chart, OCR, and Grounding, studying both LoRA and full fine-tuning models. Moreover, we explore how model merging can combine different modalities (e.g., vision-language, audio-language, and video-language models), moving toward the Omni-language model. $(ii)$ We implement 10 model merging algorithms on the benchmark. Furthermore, we propose a novel method that removes noise from task vectors and robustly optimizes the merged vector based on a loss defined over task vector interactions, achieving an average performance gain of 2.48\%. $(iii)$ We find that model merging offers a promising way for building improved MLLMs without requiring training data. Our results also demonstrate that the complementarity among multiple modalities outperforms individual modalities.

Medication recommender is to suggest appropriate medication combinations based on a patient’s health history, e.g., diagnoses and procedures. Existing works represent different diagnoses/procedures well separated by one-hot encodings. However, they ignore the latent hierarchical structures of these medical terms, undermining the generalization performance of the model. For example, “Respiratory Diseases”, “Chronic Respiratory Diseases” and “Chronic Bronchiti” have a hierarchical relationship, progressing from general to specific. To address this issue, we propose a novel hierarchical encoder named HIER to hierarchically represent diagnoses and procedures, which is based on standard medical codes and compatible with any existing methods. Specifically, the proposed method learns relation embedding with a self-supervised objective for incorporating the neighbor hierarchical structure. Additionally, we develop the position encoding to explicitly introduce global hierarchical position. Extensive experiments demonstrate significant and consistent improvements in recommendation accuracy across four baselines and two real-world clinical datasets. The code is available at https://github.com/yuliang-liang/HEIR .

Safety alignment incurs safety tax that perturbs a large reasoning model's (LRM) general reasoning ability. Existing datasets used for safety alignment for an LRM are usually constructed by distilling safety reasoning traces and answers from an external LRM or human labeler. However, such reasoning traces and answers exhibit a distributional gap with the target LRM that needs alignment, and we conjecture such distributional gap is the culprit leading to significant degradation of reasoning ability of the target LRM. Driven by this hypothesis, we propose a safety alignment dataset construction method, dubbed DGR. DGR transforms and refines an existing out-of-distributional safety reasoning dataset to be aligned with the target's LLM inner distribution. Experimental results demonstrate that i) DGR effectively mitigates the safety tax while maintaining safety performance across all baselines, i.e., achieving +30.2% on DirectRefusal and +21.2% on R1-ACT improvement in average reasoning accuracy compared to Vanilla SFT; ii) the degree of reasoning degradation correlates with the extent of distribution shift, suggesting that bridging this gap is central to preserving capabilities. Furthermore, we find that safety alignment in LRMs may primarily function as a mechanism to activate latent knowledge, as a mere 10 samples are sufficient for activating effective refusal behaviors. These findings not only emphasize the importance of distributional consistency but also provide insights into the activation mechanism of safety in reasoning models.

Medical visual question answering (Med-VQA) is a task that aims to answer clinical questions given a medical image. Existing literature generally treats it as a classic classification task based on interaction features of the image and question. However, such a paradigm ignores the valuable semantics of candidate answers as well as their relations. From the real-world dataset, we observe that: 1) The text of candidate answers has a strong intrinsic correlation with medical images; 2) Subtle differences among multiple candidate answers are crucial for identifying the correct one. Therefore, we propose an answer semantics enhanced (ASE) method to integrate the semantics of answers and capture their subtle differences. Specifically, we enhance the semantic correlation of image-question-answer triplets by aligning images and question-answer tuples within the feature fusion module. Then, we devise a contrastive learning loss to highlight the semantic differences between the correct answer and other answers. Finally, extensive experiments demonstrate the effectiveness of our method.