AAAI 2026 Papers

With the emergence of large multimodal models, dual-encoder alignment via contrastive learning has seen a resurgence. However, the escalating model size demands effective Parameter-Efficient Fine-Tuning (PEFT). While LoRA is a promising inference-free alternative to adapters, we find that its naive application to multimodal tasks causes a severe rank imbalance, favoring the text modality and FFN layers. To address this, we propose HALoRA (Hierarchical Allocation LoRA), which introduces a component-wise budget allocator to ensure balanced fine-tuning across both modalities and their internal components. This is complemented by a gradient-approximated initialization to accelerate convergence. With only half the parameters of adapters, HALoRA achieves superior or competitive performance in retrieval and zero-shot classification. Our work presents a more principled approach to multimodal LoRA, uncovering an intriguing asymmetry in vision-language alignment.

Cross-market recommendation (CMR) faces severe challenges from distribution shifts between data-rich source markets and sparse target markets. Existing methods rely on a pre-training and fine-tuning paradigm for knowledge transfer, yet suffer from two key limitations: i) the objective gap between pre-training and full-parameter fine-tuning causes loss of generalized knowledge from source markets; ii) the high computational costs of extensive fine-tuning hinder scalability. To this end, we propose DCMPT, a novel Distilled Cross-Market Prompt-Tuning approach. DCMPT reframes the problem under a more efficient pre-training and prompt-tuning paradigm. Instead of full fine-tuning, we adapt a pre-trained universal backbone by freezing its weights and injecting a minimal set of learnable prompts to form a "student" model. To effectively optimize these prompts on sparse data, we introduce a novel teacher-student architecture: a specialized "teacher" model, trained exclusively on the target market, provides dense, market-specific supervision. This guidance is delivered via a dual distillation strategy designed to transfer global ranking patterns and adapt to local consumer tastes. Extensive experiments on real-world market datasets demonstrate that DCMPT significantly outperforms state-of-the-art methods, achieving superior target market performance with substantial parameter-efficiency.

Multimodal learning has shown significant superiority on various tasks by integrating multiple modalities. However, the interdependencies among modalities increase the susceptibility of multimodal models to adversarial attacks. Existing methods mainly focus on attacks on specific modalities or indiscriminately attack all modalities. In this paper, we find that these approaches ignore the differences between modalities in their contribution to final robustness, resulting in suboptimal robustness performance. To bridge this gap, we introduce Vulnerability-Aware Robust Multimodal Adversarial Training (VARMAT), a probe-in-training adversarial training method that improves multimodal robustness by identifying the vulnerability of each modality. To be specific, VARMAT first explicitly quantifies the vulnerability of each modality, grounded in a first-order approximation of the attack objective (Probe). Then, we propose a targeted regularization term that penalizes modalities with high vulnerability, guiding robust learning while maintaining task accuracy (Training). We demonstrate the enhanced robustness of our method across multiple multimodal datasets involving diverse modalities. Finally, we achieve {12.73%, 22.21%, 11.19%} robustness improvement on three multimodal datasets, revealing a significant blind spot in multimodal adversarial training.

Peptide-based drug design targeting “undruggable” proteins remains one of the most critical challenges in modern drug discovery. Conventional peptide-discovery pipelines rely on low-throughput experimental screening, which is both time-consuming and prohibitively expensive. Moreover, existing computational approaches for designing peptides against target proteins typically depend on the availability of high-quality structural information. Although recent structure-prediction tools such as AlphaFold3 have achieved breakthroughs in protein modeling, their accuracy for functional interfaces remains limited. The acquisition of high-resolution structures is often expensive, time-intensive, and particularly challenging for targets with dynamic conformations, further restricting the efficient development of peptide therapeutics. Additionally, current sequence-based generative methods follow a paradigm that relies on known templates, which limits the exploration of sequence space and results in generated peptides lacking diversity and novelty. To address these limitations, we propose a contrastive conditioned diffusion framework for target-specific peptide generation, referred to as PepCCD. It employs a contrastive learning strategy between proteins and peptides to extract sequence-based conditioning representations of target proteins, which serve as precise conditions to guide a pre-trained diffusion model to generate peptide sequences with the desired target specificity. Extensive experiments on multiple benchmark target proteins demonstrate that the peptides designed by PepCCD exhibit strong binding affinity and outperform state-of-the-art methods in terms of diversity and generation efficiency.

High-dimensional mediation analysis (HMA) seeks to uncover complex causal mechanisms involving numerous mediators and plays a crucial role in scientific and social sciences. In this work, we introduce the Generative Adversarial High-dimensional Mediation Network (GAHMN), a novel, scalable structured generative framework designed for causal analysis in high-dimensional settings. GAHMN formulates mediation analysis as dual conditional generative blocks, explicitly capturing mediators' dual roles as outcomes influenced by treatments and as predictors affecting outcomes. Each block integrates a high-dimensional partially linear structure with multi-channel convolutional layers, promoting effective parameter sharing and enhanced representation learning. To induce sparsity and accurate mediator selection, GAHMN employs customized min-max optimization problems with L1 penalties on generator parameters, alongside specially designed optimization algorithms for efficient computation. Unlike existing benchmark methods relying on restrictive parametric assumptions or random-effect specifications, GAHMN flexibly captures heterogeneity, complex distributions, and inter-mediator correlations. With careful design, the computational complexity of GAHMN scales linearly with the number of mediators p, rather than quadratically as in conventional approaches. Theoretical results rigorously ensures estimation consistency, convergence rate, and accurate sparse recovery. GAHMN also serves as a structured generative causal modeling framework, extending to causal decomposition, structural equation modeling, and counterfactual policy evaluation. Extensive experiments confirm GAHMN's superior performance and robustness in synthetic and real-world scenarios.

Backdoor attacks pose a severe threat to federated graph learning (FGL), where malicious clients can inject hidden triggers into the global model without being detected. Defending against such attacks is particularly challenging due to the complex graph structures and the stealthy nature of trigger patterns. In this work, we propose MultiKD, a novel backdoor mitigation method based on attention-guided multi-teacher distillation. Unlike existing defenses that focus on detecting suspicious clients or restricting backdoor activation, MultiKD directly purifies the global model on the server side by exploiting intermediate representations. It integrates knowledge from multiple client models and guides the global model to suppress backdoor behaviors by aligning attention maps and preserving inter-layer relational consistency. Our defensive intuition enables MultiKD to retain task-relevant information while mitigating malicious patterns, even when some teacher models are compromised. Extensive experiments on four real-world datasets demonstrate the effectiveness of our approach in significantly reducing attack success rate (≤ 8%) with minimal impact on utility (≤ 5%).

Continual Learning with Pre-trained Models holds great promise for efficient adaptation across sequential tasks. However, most existing approaches freeze PTMs and rely on auxiliary modules like prompts or adapters, limiting model plasticity and leading to suboptimal generalization when facing significant distribution shifts. While full fine-tuning can improve adaptability, it risks disrupting crucial pre-trained knowledge. In this paper, we propose Mutual Information-guided Sparse Tuning (MIST), a plug-and-play method that selectively updates a small subset of PTM parameters, less than 5%, based on sensitivity to mutual information objectives. MIST enables effective task-specific adaptation while preserving generalization. To further reduce interference, we introduce strong sparsity regularization by randomly dropping gradients during tuning, resulting in fewer than 0.5% of parameters being updated per step. Applied before standard freeze-based methods, MIST consistently boosts performance across diverse continual learning benchmarks. Experiments show that integrating our method into multiple baselines yields significant performance gains.

Time series forecasting is essential across diverse domains. While MLP-based methods have gained attention for achieving Transformer-comparable performance with fewer parameters and better robustness, they face critical limitations including loss of weak seasonal signals, capacity constraints in weight-sharing MLPs, and insufficient channel fusion in channel-independent strategies. To address these challenges, we propose MDMLP-EIA (Multi-domain Dynamic MLPs with Energy Invariant Attention) with three key innovations. First, we develop an adaptive fused dual-domain seasonal MLP that categorizes seasonal signals into strong and weak components. It employs an adaptive zero-initialized channel fusion strategy to minimize noise interference while effectively integrating predictions. Second, we introduce an energy invariant attention mechanism that adaptively focuses on different feature channels within trend and seasonal predictions across time steps. This mechanism maintains constant total signal energy to align with the decomposition-prediction-reconstruction framework and enhance robustness against disturbances. Third, we propose a dynamic capacity adjustment mechanism for channel-independent MLPs. This mechanism scales neuron count with the square root of channel count, ensuring sufficient capacity as channels increase. Extensive experiments across nine benchmark datasets demonstrate that MDMLP-EIA achieves state-of-the-art performance in both prediction accuracy and computational efficiency.

Federated recommendations (FRs), facilitating multiple local clients to collectively learn a global model without disclosing user private data, have emerged as a prevalent on-device service. In conventional FRs, a dominant paradigm is to utilize discrete identities to represent clients and items, which are then mapped to domain-specific embeddings to participate in model training. Despite considerable performance, we reveal three inherent limitations that can not be ignored in federated settings, i.e., non-transferability across domains, ineffectiveness in cold-start settings, and potential privacy violations during federated training. To this end, we propose a transferable federated recommendation model, TransFR, which delicately incorporates the general capabilities empowered by pre-trained models and the personalized abilities by fine-tuning local private data. Specifically, it first learns domain-agnostic representations of items by exploiting pre-trained models with public textual corpora. To tailor for FR tasks, we further introduce efficient federated adapter-tuning and post-adaptation personalization, which facilitate personalized adapters for each client by fitting local private data. We theoretically prove the advantages of incorporating adapter tuning in FRs regarding both effectiveness and privacy. Through extensive experiments, we show that our TransFR surpasses state-of-the-art FRs on transferability.

Graphs effectively model interactions in real-world applications such as social and trade networks, where Graph Neural Networks (GNNs) excel at tasks such as link prediction to enhance user experiences. Despite these benefits, users raise privacy concerns as user data can be exploited to improve GNN performance without consent. Accordingly, various graph unlearning methods have been developed. Prior work shows that comparing models before and after unlearning enables attackers to launch former membership inference attacks (FMIA) on unlearned data. However, the imprint of unlearned data left in the unlearned model itself remains underexplored, and existing membership inference methods mainly exploit overfitting, making them ineffective for identifying unlearned data. To address this, we conducted theoretical analysis and proposed an attack framework targeting unlearned GNNs by learning the distribution patterns of unlearned data to distinguish them from normal test data. Extensive experiments on four real-world datasets and GNN architectures confirm our framework's effectiveness and reveal significant vulnerabilities in current graph unlearning methods.

The widespread adoption of graph neural networks (GNNs) has brought increased attention to fairness issues related to sensitive attributes, such as gender and race, in practical scenarios. However, this concern remains largely unexplored in the context of graph clustering. Conventional fair graph clustering methods primarily depend on spectral clustering approaches. Meanwhile, we argue that existing graph learning works mainly focus on a single type of fairness, whereas graph clustering should achieve group equality-informed individual fairness. In this paper, we introduce for the first time a fairness-aware framework termed FairGC for deep graph clustering, which integrates the dual objectives of individual and group fairness while maintaining accurate clustering results. Specifically, we construct two views with distinct semantics using Siamese encoders. Then, we apply multi-step random walks on view-specific affinity graphs to capture high-order affinities of node pairs, thereby reformulating the contrastive learning with a focus on individual similarity. Besides, we utilize adversarial learning by making node representations independent of the estimated sensitive attributes to further eliminate group biases of clustering results. Extensive experiments on four benchmarks demonstrate the effectiveness and superiority of our proposed framework FairGC.

Large Language Models (LLMs) increasingly leverage Federated Learning (FL) to utilize private, task-specific datasets for fine-tuning while preserving data privacy. However, while federated LLM frameworks effectively enable collaborative training without raw data sharing, they critically lack built-in mechanisms for regulatory compliance like GDPR’s right to be forgotten. Integrating private data heightens concerns over data quality and long-term governance, yet existing distributed training frameworks offer no principled way to selectively remove specific client contributions post-training. Due to distributed data silos, stringent privacy constraints, and the intricacies of interdependent model aggregation, federated LLM unlearning is significantly more complex than centralized LLM unlearning. To address this gap, we introduce Oblivionis, a lightweight learning and unlearning framework that enables clients to selectively remove specific private data during federated LLM training, enhancing trustworthiness and regulatory compliance. By unifying FL and unlearning as a dual optimization objective, we incorporate 6 FL and 5 unlearning algorithms for comprehensive evaluation and comparative analysis, establishing a robust pipeline for federated LLM unlearning. Extensive experiments demonstrate that Oblivionis outperforms local training, achieving a robust balance between forgetting efficacy and model utility, with cross-algorithm comparisons providing clear directions for future LLM development.

Prompt tuning has emerged as a lightweight strategy for adapting foundation models to downstream tasks, particularly for resource-constrained systems. As pre-trained prompts become valuable assets, combining multiple source prompts offers a promising approach to enhance generalization for new tasks by leveraging complementary knowledge. However, naive aggregation often overlooks different source prompts have different contribution potential to the target task. To address this, we propose HGPrompt, a dynamic framework that learns optimal ensemble weights. These weights are optimized by jointly maximizing an information-theoretic metric for transferability and minimizing gradient conflicts via a novel regularization strategy. Specifically, we propose a differentiable prompt transferability metric to captures the discriminability of prompt-induced features on the target task. Meanwhile, HGPrompt match the gradient variances with respect to different source prompts based on Hessian and Fisher Information, ensuring stable and coherent knowledge transfer while suppressing gradient conflicts among them. Extensive experiments on the large-scale VTAB benchmark demonstrate the state-of-the-art performance of HGPrompt, validating its effectiveness in learning an optimal ensemble for effective multi-source prompt transfer.

Deep neural networks are often over-parameterized, resulting in prohibitive storage and computational costs. A fundamental question is whether a complex network can be re-expressed in terms of a compact set of basis functions without sacrificing accuracy. Motivated by this perspective, we aim to approximate a dense model by decomposing it into a small number of lightweight components that capture the essential functional structure of the network. To this end, we propose a series expansion framework that rewrites a neural network as a linear combination of low-bit basis models. Within the post-training quantization setting, the full-precision model is expanded hierarchically at the tensor, layer, and model levels into a structured set of basis functions. We theoretically prove that this expansion converges exponentially to the original model. Furthermore, we design AbelianAdd and AbelianMul operations between isomorphic basis models, endowing the expansion with an Abelian group structure that naturally supports commutative and parallel computation. Experimental results across diverse architectures show that our series expansion method leverages a set of ultra-low-bit basis functions, not only preserving full-precision performance without the need for calibration data or fine-tuning, but also featuring a parallel-friendly design that enables efficient and scalable deployment.

Mining time-frequency features is critical for time series forecasting. Existing research has predominantly focused on modeling low-frequency patterns, where most time series energy is concentrated. The overlooking of mid to high frequency continues to limit further performance gains in deep learning models. We propose FreqCycle, a novel framework integrating: (i) a Filter-Enhanced Cycle Forecasting (FECF) module to extract low-frequency features by explicitly learning shared periodic patterns in the time domain, and (ii) a Segmented Frequency-domain Pattern Learning (SFPL) module to enhance mid to high frequency energy proportion via learnable filters and adaptive weighting. Furthermore, time series data often exhibit coupled multi-periodicity, such as intertwined weekly and daily cycles. To address coupled multi-periodicity as well as long lookback window challenges, we extend FreqCycle hierarchically into MFreqCycle, which decouples nested periodic features through cross-scale interactions. Extensive experiments on seven diverse domain benchmarks demonstrate that FreqCycle achieves state-of-the-art accuracy while maintaining faster inference speeds, striking an optimal balance between performance and efficiency.

Large Language Models (LLMs) demonstrate impressive capabilities in natural language understanding and generation, but incur high communication overhead and privacy risks in cloud deployments, while facing compute and memory constraints when confined to edge devices.Cloud–edge inference has emerged as a promising paradigm for improving privacy in LLM services by retaining sensitive computations on local devices.However, existing cloud–edge inference approaches apply uniform privacy protection without considering input sensitivity, resulting in unnecessary perturbation and degraded utility even for non-sensitive tokens. To address this limitation, we propose Privacy-aware Routing for Inference with Semantic Modulation (PRISM), a context-aware framework that dynamically balances privacy and inference quality. PRISM executes in four stages: (1) the edge device profiles entity-level sensitivity; (2) a soft gating module, also on the edge, selects an execution mode -cloud, edge, or collaboration; (3) for collaborative paths, the edge applies adaptive two-layer local differential privacy based on entity risks; and (4) the cloud LLM generates a semantic sketch from the perturbed prompt, which is then refined by the edge-side small language model (SLM) using local context.Our results show that PRISM consistently achieves superior privacy-utility trade-offs in various scenarios, reducing energy consumption and latency to 40–50% of baseline methods such as Uniform and Selective LDP, while maintaining high output quality under strong privacy constraints. These findings are validated through comprehensive evaluations involving realistic prompts, actual energy measurements, and heterogeneous cloud–edge model deployments.

Multiplex graphs are widely used to model multi-relational complex systems and play an important role in various real-world scenarios, such as financial systems and social networks. Hence, detecting anomalous samples in multiplex graph becomes crucial to ensure cybersecurity and stability. Although existing homogeneous graph anomaly detection (GAD) methods can be applied to deal with multiplex graphs, they still face two major challenges: 1) Due to the multiplicity and complexity of relations in multiplex graphs, homogeneous GAD models fail to effectively capture anomalous behaviors that correlate with diverse relational patterns. 2) In real-world applications, malicious entities usually disguise themselves through various camouflage strategies, making it difficult to capture subtle anomalous features via single-relation analysis. To address these challenges, we propose a novel unsupervised anomaly detection method for multiplex graphs based on Similarity-constrained Fusion Graph Autoencoder (SFGA). In SFGA, we design a multiplex graph autoencoder and introduced a cross-plex attention module at the model bottleneck to achieve comprehensive modeling of cross-relation anomaly patterns. Then, a similarity balancing strategy is proposed to constrain node representations at the bottleneck from both local and global perspectives, enhancing the discriminative power against camouflaged anomalies of autoencoder and enabling more effective identification of anomalous nodes with overlapping or deceptive patterns. Extensive experiments are conducted on both synthetic and real-world datasets at varying scales, and the results demonstrate that our proposed method outperforms state-of-the-art approaches by a large margin.

Recent vision-language models (VLMs) show strong reasoning capabilities through training with reinforcement learning from verifiable rewards (RLVR). Despite their impressive capabilities, current VLMs focus on a limited range of reasoning tasks, such as mathematical and logical reasoning, due to the lack of readily available verifiable reward data in broader domains. As a result, these models struggle to generalize their reasoning abilities to the wide variety of challenges encountered in real-world environments. To address this limitation, we collect and assemble a comprehensive RL-ready visual reasoning training dataset encompassing 46 datasets across 13 dimensions of 5 domains, covering a wide range of realistic scenarios such as infographic reasoning, mathematical reasoning, spatial reasoning, and general science reasoning. Based on this dataset, we propose an influence function-based data filtering strategy and a multi-round data curriculum method to iteratively strengthen general visual reasoning abilities. Using this approach, we train a general reasoning VLM, namely Vision-G1. Our 7B model achieves state-of-the-art performance across nine visual reasoning benchmarks, surpassing previous similar-sized VLMs and even GPT-4o and Gemini-1.5 Flash.

This paper presents FDC-Ground, a reinforcement learning framework that addresses the high-cost, low-signal challenge of GUI grounding training. The framework introduces two core contributions: (1) the Exponentially Decayed Distance Reward (EDDR), which provides resolution-robust and continuous feedback for position predictions, and (2) the Fact-Aligned Dynamic Completions Pruning (FDC-Pruning) strategy, which selectively retains completions whose advantage signs align with factual correctness, thereby reducing computational overhead while enhancing gradient quality and training stability. Using only 3.2K training samples and a single epoch, our 7B-parameter model achieves 88.3% and 91.0% accuracy on ScreenSpot and ScreenSpot-v2, outperforming several RL-based models such as UIShift and SE-GUI. Our 3B-parameter model based on Qwen2.5-VL-3B surpasses its original performance by +26.6%, demonstrating the effectiveness of our reward design and pruning strategy under low-resource conditions. Furthermore, the proposed FDC-Pruning strategy achieves a 1.18× training speedup and a +5.9% accuracy improvement over standard GRPO, and expanding the exploration space to 4× yields an additional +10.5% gain, confirming both the scalability and the training efficiency of our approach. These findings highlight that combining EDDR with FDC-Pruning offers a practical path toward scalable and efficient RL-based GUI grounding, even in low-resource settings.