Safety alignment remains brittle under domain shift and noisy preference supervision. Existing robust alignment methods predominantly focus on data uncertainty in alignment data, while being less effective at addressing failures caused by optimization-induced fragility. In this work, we revisit robustness for safety alignment from an optimization geometry perspective, highlighting optimization-induced fragility as a complementary factor to data-space uncertainty. To address this, we propose ShaPO, a geometry-aware preference optimization framework that enforces worst-case alignment objectives via selective geometry control. We instantiate ShaPO at two levels: Token-level ShaPO stabilizes likelihood-based surrogate optimization, while Reward-level ShaPO enforces reward-consistent optimization and provides improved robustness under noisy supervision. Across diverse safety benchmarks and noisy preference settings, ShaPO consistently improves safety robustness over popular preference optimization methods. Moreover, ShaPO composes cleanly with data-robust objectives, yielding additional gains and empirically supporting the proposed optimization-geometry perspective. Our results highlight optimization geometry as a complementary and actionable axis for robust LLM safety alignment. The code is available at \url{https://anonymous.4open.science/r/ShaPO-D1B0}.

Machine unlearning aims to forget sensitive knowledge from Large Language Models (LLMs) while maintaining general utility. However, existing approaches typically treat all tokens in a response indiscriminately and enforce uncertainty over the entire vocabulary. This global treatment results in unnecessary utility degradation and extends optimization to content-agnostic regions. To address these limitations, we propose PALU (Prefix-Aware Localized Unlearning), a framework driven by a local entropy maximization objective across both temporal and vocabulary dimensions. PALU reveals that (i) suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and (ii) flattening only the top-K logits is adequate to maximize uncertainty in the critical subspace. These findings allow PALU to alleviate redundant optimization across the full vocabulary and parameter space while minimizing collateral damage to general model performance. Comprehensive evaluations validate that PALU achieves superior forgetting efficacy and utility preservation compared to state-of-the-art baselines. Our code is available at https://github.com/nxZhai/PALU.

Cognitive Diagnosis (CD) models are constrained by the data quality of students' response logs. Recent advancements in Large Language Model (LLM) based data augmentation show promise for enhancing CD. However, ensuring the reliability and accuracy of LLM-generated annotations remains a significant challenge. In this paper, we propose Multi-Agent based Concept Augmentation for Cognitive Diagnosis (MACA-CD), a novel approach that enhances CD by generating and fusing reliable concept descriptions and relations based solely on concept names. MACA-CD consists of two main components: (1) a Multi-Agent Debate (MAD) based concept augmentation process that generates diverse and reliable concept descriptions and relations, reducing reliance on behavioral data. For concept descriptions, two agents generate outputs that include definitions, core features, and real-world applications, and continue debating until a judge agent determines that consensus has been reached. Concept relations are then identified using a Breadth-First Search approach to efficiently and progressively uncover relationships based on concept descriptions, with each step carried out by MAD. (2) a concept augmentation-enhanced CD model that refines concept embeddings using a graph self-supervised learning fusion layer and a pairwise comparator-based Description Fusion Layer, leading to more reliable and accurate concept embeddings. Experimental results on three real-world datasets show that MACACD consistently outperforms existing methods under various realworld scenarios.

Deep Research (DR) Agents powered by advanced Large Language Models (LLMs) have fundamentally shifted the paradigm for completing complex research tasks. Yet, a comprehensive and live evaluation of their forecasting performance on real-world, research-oriented tasks in high-stakes domains (e.g., finance) remains underexplored. We introduce FinDeepForecast, the first live, end-to-end multi-agent system for automatically evaluating DR agents by continuously generating research-oriented fi- nancial forecasting tasks. This system is equipped with a dual-track taxonomy, enabling the dynamic generation of recurrent and non-recurrent forecasting tasks at both corporate and macro levels. With this system, we generate FinDeepForecastBench, a weekly evaluation benchmark over a ten-week horizon, encompassing 8 global economies and 1,314 listed companies, and evaluate 13 representative methods. Extensive experiments show that, while DR agents consistently outperform strong baselines, their performance still falls short of genuine forward-looking financial reasoning. We expect the proposed FinDeepForecast system to consistently facilitate future advancements of DR agents in research-oriented financial forecasting tasks. The benchmark and leaderboard are publicly available on the OpenFinArena Platform.

Graph Convolutional Networks (GCNs) are widely adopted backbones for collaborative filtering due to their simplicity and performance. To address data sparsity, Graph Contrastive Learning (GCL) has been integrated, with its effectiveness often attributed to promoting node representations' uniformity. However, we uncover an under-explored conflict: the messagepassing mechanism of GCNs inherently amplifies popular nodes, while the uniformity objective of GCL tends to dilute their discriminative power. This induces a “seesaw effect”: enhancing long-tail recommendations often comes at the expense of popular item performance, and an excessive pursuit of uniformity may blur essential semantic distinctions. To ameliorate this conflict, we propose UniGCL, a unified framework that harmonizes GCN and GCL by deliberately rebalancing the influence of popular nodes. First, we enhance the GCN backbone with inverse logarithmbased normalization, which adaptively strengthens the influence of popular nodes in representation propagation. Second, we propose a triple contrastive learning module grounded in inverse propensity scoring (IPS) to evaluate alignment and uniformity across users, items, and interaction pairs, thereby mitigating the over-emphasis on popular nodes during training. Consequently, UniGCL achieves a dual adjustment: it leverages the inherent aggregation of popular nodes during message passing, while counteracting their dominance in the contrastive phase. This strategy enables the model to excel in long-tail recommendations without compromising the discriminability of popular items, effectively mitigating the seesaw effect. The effectiveness of the proposed UniGCL framework is validated through extensive experiments on three public benchmarks. Specifically, UniGCL surpasses the strong baseline VGCL in terms of NDCG@20 by $9.2 \%, 2.7 \%$, and 3.1% on the Douban-Book, Yelp2018, and Amazon-Book datasets, respectively. Furthermore, we present a comprehensive group-wise evaluation to validate its success in mitigating the seesaw effect. Our code can be found at: https://github.com/ChanglongZheng/UniGCL.

Recommender systems serve as a foundational infrastructure in modern information ecosystems, helping users navigate the expanding digital content space and discover items aligned with their preferences. At their core, recommender systems address a fundamental research problem: matching users with items. Over the past decades, the field has experienced successive technological paradigm shifts, from collaborative filtering and matrix factorization in the machine learning era to sophisticated neural architectures in the deep learning era. Recently, the emergence of generative models, especially large language models (LLMs) and diffusion models have sparked a new paradigm: generative recommendation, which reconceptualizes the recommendation problem as a generation task rather than a discriminative scoring procedure. This survey provides a comprehensive examination of this paradigm through a unified tripartite framework spanning data, model, and task dimensions. Rather than simply categorizing works, we systematically decompose approaches into operational stages—data augmentation and unification, model alignment and training, task formulation and execution. At the data level, generative models enable knowledge-infused augmentation and agent-based simulation while unifying heterogeneous signals. At the model level, we taxonomize LLM-based methods, large recommendation models, and diffusion approaches, analyzing their alignment mechanisms and innovations. At the task level, we illuminate new capabilities including conversational interaction, explainable reasoning, and personalized content generation. We identify five key advantages: world knowledge integration, natural language understanding, reasoning capabilities, scaling laws, and creative generation. We critically examine challenges in benchmark design, model robustness, and deployment efficiency, while charting a roadmap toward intelligent recommendation assistants that fundamentally reshape human-information interaction.

Generative recommendation autoregressively generates the semantic IDs of the target item, unifying preference modeling and index retrieval within the shared token space. Recent attempts have introduced Multi-Token Prediction (MTP) into this field, yet they primarily inherit its efficiency merit, leaving its potential as dense supervision unexplored. Unlocking this potential hinges on whether future behaviors qualify as informative supervision. Our analysis reveals that future behaviors carry a semantic echo of the current one far above that of random pairs, which nevertheless decays along horizons under intent transitions, making them informative yet order-dependent signals. Motivated by this, we propose EchoRec, which empowers MTP with cycle-consistent holistic preference alignment across multi-horizon for generative recommendation. It comprises two synergistic modules. Horizon-aware Preference Generation (HPG) sequentially chains lightweight auxiliary branches upon the base recommender, where each branch conditions on its predecessor to respect preference evolution. Verifiable Holistic-Preference Alignment (VHA) further consolidates them into the holistic preference and echoes it back through cycle-consistent projectors to suppress spurious alignment, with theoretical guarantees that exclude the rank-collapse form of spurious alignment under an invertible transport, enabling the holistic preference to be retained in the decoding representation. All auxiliary components serve as disposable scaffolding discarded at inference, introducing negligible online serving overhead. Extensive experiments on three datasets demonstrate the superiority of our EchoRec, together with its naturally acquired multi-item generation ability. Our code and datasets will be available upon acceptance.

Multimodal recommendation has emerged as an effective paradigm for enhancing collaborative filtering by incorporating heterogeneous content modalities. Existing multimodal recommenders predominantly focus on reinforcing cross-modal consistency to facilitate multimodal fusion. However, we observe that multimodal representations often exhibit substantial cross-modal redundancy, where dominant shared components overlap across modalities. Such redundancy can limit the effective utilization of complementary information, explaining why incorporating additional modalities does not always yield performance improvements. In this work, we propose CLEAR, a lightweight and plug-and-play cross-modal de-redundancy approach for multimodal recommendation. Rather than enforcing stronger cross-modal alignment, CLEAR explicitly characterizes the redundant shared subspace across modalities by modeling cross-modal covariance between visual and textual representations. By identifying dominant shared directions via singular value decomposition and projecting multimodal features onto the complementary null space, CLEAR reshapes the multimodal representation space by suppressing redundant cross-modal components while preserving modality-specific information. This subspace-level projection implicitly regulates representation learning dynamics, preventing the model from repeatedly amplifying redundant shared semantics during training. Notably, CLEAR can be seamlessly integrated into existing multimodal recommenders without modifying their architectures or training objectives. Extensive experiments on three public benchmark datasets demonstrate that explicitly reducing cross-modal redundancy consistently improves recommendation performance across a wide range of multimodal recommendation models.

Large language model-based web agents have demonstrated strong performance on realistic web interaction tasks. However, existing evaluations are predominantly conducted under relatively stable and well-behaved interaction conditions, which may overestimate agent robustness. High task success in such idealized settings does not necessarily reflect performance under realistic web interaction. To address this limitation, we introduce a diagnostic stress-testing benchmark for web agents. We first construct realistic and controllable web environments that provide clean and stable interaction workflows as reference baselines. We then introduce structured and controlled perturbations that emulate interaction variability, including shifting layouts, altered interaction semantics, and execution disruptions. By comparing agent behavior between clean and perturbed settings, our framework enables systematic diagnosis of robustness under what-if interaction scenarios. Through extensive evaluation of state-of-the-art multimodal web agents, we show that stress-based evaluation exposes failure modes and substantial robustness gaps that remain hidden under clean benchmark conditions.

Deepfake detection suffers from poor generalization across forgery methods, as existing models tend to rely on spurious method-specific shortcuts that fail to transfer to unseen manipulations. While recent approaches attempt to improve generalization, they lack an explicit mechanism to identify and suppress such shortcuts in learned representations. In this work, we propose Shortcut Subspace Suppression (S^3) framework that explicitly characterizes and suppresses method-specific shortcuts via subspace modeling. Our key insight is that variations distinguishing different forgery methods capture method-specific artifacts and thus serve as an effective proxy for method-specific shortcuts. To this end, we train a lightweight linear probe for forgery method classification and perform Singular Value Decomposition (SVD) to extract the dominant shortcut subspace. Building on this formulation, we develop two complementary strategies to reduce shortcut reliance. During training, we softly suppress the shortcut subspace in feature representations, encouraging the model to rely on more generalizable cues for real/fake discrimination. At inference time, we introduce a training-free counterpart that attenuates neurons aligned with the identified shortcut directions, enabling plug-and-play generalization enhancement with improved interpretability. Extensive experiments on multiple benchmarks demonstrate that our method significantly improves cross-method generalization while maintaining strong in-domain performance. The code will be released upon acceptance of the submission.

As Large Language Models (LLMs) increasingly shape online content, removing targeted information from well-trained LLMs (also known as LLM unlearning) has become critical for web governance. A key challenge lies in sample-wise imbalance within the forget set: different samples exhibit widely varying unlearning difficulty, leading to asynchronous forgetting where some knowledge remains insufficiently erased while others become over-forgotten. To address this, we propose BalDRO, a novel and efficient framework for balanced LLM unlearning. BalDRO formulates unlearning as a min-sup process: an inner step identifies a worst-case data distribution that emphasizes hard-to-unlearn samples, while an outer step updates model parameters under this distribution. We instantiate BalDRO via two efficient variants: BalDRO-G, a discrete GroupDRO-based approximation focusing on high-loss subsets, and BalDRO-DV, a continuous Donsker-Varadhan dual method enabling smooth adaptive weighting within standard training pipelines. Experiments on TOFU and MUSE show that BalDRO significantly improves both forgetting quality and model utility over existing methods, and we release code for reproducibility.

Aligning large language models (LLMs) with human values has become increasingly important as their influence on human behavior and decision-making expands. However, existing steering-based alignment methods suffer from limited controllability: steering a target value often unintentionally activates other, non-target values. To characterize this limitation, we introduce value leakage, a diagnostic notion that captures the unintended activation of non-target values during value steering, along with a normalized leakage metric grounded in Schwartz's value theory. In light of this analysis, we propose NeVA, a neuron-level editing framework for controllable value alignment in LLMs. NeVA identifies sparse, value-relevant neurons and performs inference-time activation editing, enabling fine-grained control without parameter updates or retraining. Experiments show that NeVA achieves stronger target value alignment while incurring smaller performance degradation on general capability. Moreover, NeVA significantly reduces the average leakage, with residual effects largely confined to semantically related value classes. Overall, NeVA offers a more controllable and interpretable mechanism for value alignment.

Multilingual safety remains significantly imbalanced, leaving non-high-resource (NHR) languages vulnerable compared to robust high-resource (HR) ones. Moreover, the neural mechanisms driving safety alignment remain unclear despite observed cross-lingual representation transfer.In this paper, we find that LLMs contain a set of cross-lingual shared safety neurons (SS-Neurons), a remarkably small yet critical neuronal subset that jointly regulates safety behavior across languages. We first identify monolingual safety neurons (MS-Neurons) and validate their causal role in safety refusal behavior through targeted activation and suppression. Our cross-lingual analyses then identify SS-Neurons as the subset of MS-Neurons shared between HR and NHR languages, serving as a bridge to transfer safety capabilities from HR to NHR domains. We observe that suppressing these neurons causes concurrent safety drops across NHR languages, whereas reinforcing them improves cross-lingual defensive consistency. Building on these insights, we propose a simple neuron-oriented training strategy that targets SS-Neurons based on language resource distribution and model architecture. Experiments demonstrate that fine-tuning this tiny neuronal subset outperforms state-of-the-art methods, significantly enhancing NHR safety while maintaining the model's general capabilities.

Knowledge Graph Question Answering (KGQA) aims to improve factual accuracy by leveraging structured knowledge. However, real-world Knowledge Graphs (KGs) are often incomplete, leading to the problem of Incomplete KGQA (IKGQA). A common solution is to incorporate external data to fill knowledge gaps, but existing methods lack the capacity to adaptively and contextually fuse multiple sources, failing to fully exploit their complementary strengths. To this end, we propose Debate over Mixed-knowledge (DoM), a novel framework that enables dynamic integration of structured and unstructured knowledge for IKGQA. Built upon the Multi-Agent Debate paradigm, DoM assigns specialized agents to perform inference over knowledge graphs and external texts separately, and coordinates their outputs through iterative interaction. It decomposes the input question into sub-questions, retrieves evidence via dual agents (KG and Retrieval-Augmented Generation, RAG), and employs a judge agent to evaluate and aggregate intermediate answers. This collaboration exploits knowledge complementarity and enhances robustness to KG incompleteness. In addition, existing IKGQA datasets simulate incompleteness by randomly removing triples, failing to capture the irregular and unpredictable nature of real-world knowledge incompleteness. To address this, we introduce a new dataset, Incomplete Knowledge Graph WebQuestions, constructed by leveraging real-world knowledge updates. These updates reflect knowledge beyond the static scope of KGs, yielding a more realistic and challenging benchmark. Through extensive experiments, we show that DoM consistently outperforms state-of-the-art baselines.

Recent advances in Large Language Models (LLMs) have opened new avenues for sequential recommendation by enabling natural language reasoning over user behavior sequences. A common approach formulates recommendation as a language modeling task, where interaction histories are transformed into prompts and user preferences are learned via supervised fine-tuning. However, these methods operate solely in the textual modality and often miss users' fine-grained interests, especially when shaped by rich visual signals such as product images or movie posters. Multimodal Large Language Models (MLLMs) offer a promising alternative by aligning text and vision in a shared semantic space. A prevalent training paradigm applies Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO) to model user preferences. Yet, two core challenges remain: 1) Imbalanced sample hardness, where random negative sampling causes overfitting on easy examples and under-training on hard ones; 2) Cross-modal semantic bias, where the fixed reference model in DPO prevents the policy model from correcting modality misalignments–especially over long sequences. To address these issues, we propose a Multimodal LLM framework that integrates Hardness-aware and Noise-regularized preference optimization for Recommendation (HaNoRec). Specifically, HaNoRec dynamically adjusts optimization weights based on both the estimated hardness of each training sample and the policy model's real-time responsiveness, prioritizing harder examples. It further introduces Gaussian-perturbed distribution optimization on output logits to enhance cross-modal semantic consistency and reduce modality bias inherited from the reference model.

Micro-video recommendation aims to capture user preferences from the collaborative and context information of the interacted micro-videos, thereby predicting the appropriate videos. This target is often hindered by the inherent noise within multimodal content and unreliable implicit feedback, which weakens the correspondence between behaviors and underlying interests. While conventional works have predominantly approached such scenario through behavior-augmented modeling and content-centric multimodal analysis, these paradigms can inadvertently give rise to two non-trivial challenges: preference-irrelative video representation extraction and inherent modality conflicts. To address these issues, we propose a Multi-granularity sequential modeling method via hierarchical diffusion models for micro-video Recommendation (MealRec), which simultaneously considers temporal correlations during preference modeling from intra- and inter-video perspectives. Specifically, we first propose Temporal-guided Content Diffusion (TCD) to refine video representations under intra-video temporal guidance and personalized collaborative signals to emphasize salient content while suppressing redundancy. To achieve the semantically coherent preference modeling, we further design the Noise-unconditional Preference Denoising (NPD) to recovers informative user preferences from corrupted states under the blind denoising. Extensive experiments and analyses on four micro-video datasets from two platforms demonstrate the effectiveness, universality, and robustness of our MealRec, further uncovering the effective mechanism of our proposed TCD and NPD. The source code and corresponding dataset will be available upon acceptance.

Group relative policy optimization (GRPO) has become a standard post-training paradigm for improving reasoning and preference alignment in large language models (LLMs), and has recently shown strong effectiveness in LLM-based recommender systems. However, extending GRPO-based reasoning pipelines to multimodal sequential recommendation (MSR) with multimodal large language models (MLLMs) faces fundamental obstacles. First, MSR requires jointly encoding visual content for both historical interactions and multiple candidate items, causing visual tokens to dominate the input and making the cost of group-based rollout scale with history length and candidate set size, which renders GRPO-based training prohibitively expensive. Second, existing Chain-of-Thought (CoT) supervision suffers from reward inflation in recommendation scenarios, where higher training rewards do not reliably translate into improved ranking performance and may induce shortcut learning. To address these challenges, we propose MLLMRec-R1, an efficient and stable GRPO-based reasoning framework for multimodal sequential recommendation. MLLMRec-R1 textualizes visual signals offline to eliminate expensive visual tokens while preserving multimodal semantics, and constructs high-quality multimodal CoT supervision through refinement and confidence-aware assessment. Furthermore, a mixed-grained data augmentation strategy selectively injects reliable CoT samples while retaining standard training data, mitigating reward inflation and improving generalization stability. Extensive experiments on three benchmark datasets demonstrate that MLLMRec-R1 consistently outperforms state-of-the-art methods, establishing a practical and effective GRPO-based reasoning pipeline for multimodal sequential recommendation. The code is available at https://github.com/wangyu0627/MLLMRec-R1.

Recommender Systems (RS) have been shown to be vulnerable to injective attacks, where attackers inject limited fake user profiles to promote the exposure of target items to real users for unethical gains (e.g., economic or political advantages). Since attackers typically lack knowledge of the victim model deployed in the target RS, existing methods resort to using a fixed surrogate model to mimic the potential victim model. Despite considerable progress, we argue that the assumption that poisoned data generated for the surrogate model can be used to attack other victim models is wishful. When there are significant structural discrepancies between the surrogate and victim models, the attack transferability inevitably suffers. Intuitively, if we can identify the worst-case victim model and iteratively optimize the poisoning effect specifically against it, then the generated poisoned data would be better transferred to other victim models. However, exactly identifying the worst-case victim model during the attack process is challenging due to the large space of victim models. To this end, in this work, we propose a novel attack method called Sharpness-Aware Poisoning (SharpAP). Specifically, it employs the sharpness-aware minimization principle to seek the approximately worst-case victim model and optimizes the poisoned data specifically for this worst-case model. The poisoning attack with SharpAP is formulated as a min-max-min tri-level optimization problem. By integrating SharpAP into the iterative process for attacks, our method can generate more robust poisoned data which is less sensitive to the shift of model structure, mitigating the overfitting to the surrogate model. Comprehensive experimental comparisons on three real-world datasets demonstrate that SharpAP can significantly enhance the attack transferability.

Personalized sequential recommendation aims to predict the appropriate items to users from their behavioral sequences. To alleviate the data sparsity and interest drift issues, conventional approaches typically utilize the additional behaviors from other domains via cross-domain transition. However, existing cross-domain sequential recommendation (CDSR) algorithms follow the align-then-fusion paradigm which conducts the representation-level alignment across multiple domains and mechanically combine them for recommendation, overlooking the fine-grained multi-domain fusion. Inspired by the advancements of diffusion models (DMs) in distribution matching, we propose an align-for-fusion framework for CDSR to Harmonize triple preferences utilizing Dual-oriented DMs (HorizonRec). Specifically, we first investigate the uncertainty injection of DMs and attribute the fundamental factor of the instability in existing DMs recommenders to the stochastic noise and propose a Mixed-conditioned Distribution Retrieval strategy which leverages the retrieved distribution from users' authentic behavioral logic as a bridge across the triple domains, enabling consistent multi-domain preference modeling. To suppress the potential noise and emphasize target-relevant interests during multi-domain user representation fusion, we further propose a Dual-oriented Preference Diffusion method to guide the extraction of preferences aligned with users' authentic interests from each domain under the supervision of the mixed representation. We conduct extensive experiments and analyses on four CDSR datasets from two distinct platforms to verify the effectiveness and robustness of our HorizonRec and its effective mechanism in fine-grained fusion of triple domains. Our code and datasets are available in https://github.com/YongfuZha/HorizonRec.