AAAI 2026 论文列表

Visual autoregressive modeling (VAR) via next-scale prediction has emerged as a scalable image generation paradigm. While Key and Value (KV) caching in large language models (LLMs) has been extensively studied, next-scale prediction presents unique challenges, and KV caching design for next-scale based VAR transformers remains largely unexplored. A major bottleneck is the excessive KV memory growth with the increasing number of scales—severely limiting scalability. Our systematic investigation reveals that: (1) Attending to tokens from local scales significantly contributes to generation quality (2) Allocating a small amount of memory for the coarsest scales, termed as condensed scales, stabilizes multi-scale image generation (3) Strong KV similarity across finer scales is predominantly observed in cache-efficient layers, whereas cache-demanding layers exhibit weaker inter-scale similarity. Based on the observations, we introduce AMS-KV, a scale-adaptive KV caching policy for next-scale prediction in VAR models. AMS-KV prioritizes storing KVs from condensed and local scales, preserving the most relevant tokens to maintain generation quality. It further optimizes KV cache utilization and computational efficiency identifying cache-demanding layers through inter-scale similarity analysis. Compared to the vanilla next-scale prediction-based VAR models, AMS-KV reduces KV cache usage by up to 84.83% and self-attention latency by 60.48%. Moreover, when the baseline VAR-d30 model encounters out-of-memory failures at a batch size of 128, AMS-KV enables stable scaling to a batch size of 256 with improved throughput.

Clean-image backdoor attacks, which use only label manipulation in training datasets to compromise deep neural networks, pose a significant threat to security-critical applications. A critical flaw in existing methods is that the poison rate required for a successful attack induces a proportional, and thus noticeable, drop in Clean Accuracy (CA), undermining their stealthiness. This paper presents a new paradigm for clean-image attacks that minimizes this accuracy degradation by optimizing the trigger itself. We introduce Generative Clean-Image Backdoors (GCB), a framework that uses a conditional InfoGAN to identify naturally occurring image features that can serve as potent and stealthy triggers. By ensuring these triggers are easily separable from benign task-related features, GCB enables a victim model to learn the backdoor from an extremely small set of poisoned examples, resulting in a CA drop of less than 1%. Our experiments demonstrate GCB's remarkable versatility, successfully adapting to six datasets, five architectures, and four tasks, including the first demonstration of clean-image backdoors in regression and segmentation. GCB also exhibits resilience against most of the existing backdoor defenses.

The Combined Algorithm Selection and Hyperparameter Optimization (CASH) problem is fundamental in Automated Machine Learning (AutoML). Inspired by the success of ensemble learning, recent AutoML systems construct post-hoc ensembles for final predictions rather than relying on the best single model. However, while most CASH methods conduct extensive searches for the optimal single model, they typically employ fixed strategies during the ensemble phase that fail to adapt to specific task characteristics. To tackle this issue, we propose PSEO, a framework for post-hoc stacking ensemble optimization. First, we conduct base model selection through binary quadratic programming, with a trade-off between diversity and performance. Furthermore, we introduce two mechanisms to fully realize the potential of multi-layer stacking. Finally, PSEO builds a hyperparameter space and searches for the optimal post-hoc ensemble strategy within it. Empirical results on 80 public datasets show that PSEO achieves the best average test rank (2.96) among 16 methods, including post-hoc designs in recent AutoML systems and state-of-the-art ensemble learning methods.

Model-based reinforcement learning (MBRL) is a crucial approach to enhance the generalization capabilities and improve the sample efficiency of RL algorithms. However, current MBRL methods focus primarily on building world models for single tasks and rarely address generalization across different scenarios. Building on the insight that dynamics within the same simulation engine share inherent properties, we attempt to construct a unified world model capable of generalizing across different scenarios, named Meta-Regularized Contextual World-Model (MrCoM). This method first decomposes the latent state space into various components based on the dynamic characteristics, thereby enhancing the accuracy of world-model prediction. Further, MrCoM adopts meta-state regularization to extract unified representation of scenario-relevant information, and meta-value regularization to align world-model optimization with policy learning across diverse scenario objectives. We theoretically analyze the generalization error upper bound of MrCoM in multi-scenario settings. We systematically evaluate our algorithm's generalization ability across diverse scenarios, demonstrating significantly better performance than previous state-of-the-art methods.

Large language models (LLMs) face significant deployment challenges due to their massive computational demands. While pruning offers a promising compression solution, existing methods suffer from two critical limitations: (1) They neglect activation distribution shifts between calibration data and test data, resulting in inaccurate error estimations; (2) Overlooking the long-tail distribution characteristics of activations in the attention module. To address these limitations, this paper proposes a novel pruning method, D²Prune. First, we propose a dual Taylor expansion-based method that jointly models weight and activation perturbations for precise error estimation, leading to precise pruning mask selection and weight updating and facilitating error minimization during pruning. Second, we propose an attention-aware dynamic update strategy that preserves the long-tail attention pattern by jointly minimizing the KL divergence of attention distributions and the reconstruction error. Extensive experiments show that D²Prune consistently outperforms SOTA methods across various LLMs (e.g., OPT-125M, LLaMA2/3, Qwen3). Moreover, the dynamic attention update mechanism also generalizes well to ViT-based vision models like DeiT, achieving superior accuracy on ImageNet-1K.

In social impact optimization, AI decision systems often rely on solvers that optimize well-calibrated mathematical objectives. However, these solvers cannot directly accommodate evolving human preferences, typically expressed in natural language rather than formal constraints. Recent approaches address this by using large language models (LLMs) to generate new reward functions from preference descriptions. While flexible, they risk sacrificing the system's core utility guarantees. In this paper, we propose VORTEX, a language-guided reward shaping framework that preserves established optimization goals while adaptively incorporating human feedback. By formalizing the problem as multi-objective optimization, we use LLMs to iteratively generate shaping rewards based on verbal reinforcement and text-gradient prompt updates. This allows stakeholders to steer decision behavior via natural language without modifying solvers or specifying trade-off weights. We provide theoretical guarantees that VORTEX converges to Pareto-optimal trade-offs between utility and preference satisfaction. Empirical results in real-world allocation tasks demonstrate that VORTEX outperforms baselines in satisfying human-aligned coverage goals while maintaining high task performance. This work introduces a practical and theoretically grounded paradigm for human-AI collaborative optimization guided by natural language.

Effective multivariate time series forecasting often benefits from accurately modeling complex inter-variable dependencies. However, existing attention- or graph-based methods face three key issues: (a) strong temporal self-dependencies are often disrupted by irrelevant variables; (b) softmax normalization ignores and reverses negative correlations; (c) variables struggle to perceive their temporal positions. To address these, we propose **SEED**, a Spectral Entropy-guided evaluation framework for spatial-temporal dependency modeling. SEED introduces a Dependency Evaluator, a key innovation that leverages spectral entropy to dynamically provide a preliminary evaluation of the spatial and temporal dependencies of each variable, enabling the model to adaptively balance Channel Independence (CI) and Channel Dependence (CD) strategies. To account for temporal regularities originating from the influence of other variables rather than intrinsic dynamics, we propose Spectral Entropy-based Fuser to further refine the evaluated dependency weights, effectively separating this part. Moreover, to preserve negative correlations, we introduce a Signed Graph Constructor that enables signed edge weights, overcoming the limitations of softmax. Finally, to help variables perceive their temporal positions and thereby construct more comprehensive spatial features, we introduce the Context Spatial Extractor, which leverages local contextual windows to extract spatial features. Extensive experiments on 12 real-world datasets from various application domains demonstrate that SEED achieves state-of-the-art performance, validating its effectiveness and generality.

Multi-modal entity alignment aims to identify equivalent entities across different multi-modal knowledge graphs (MMKGs). While prior work has achieved notable progress through improved multi-modal encoding and cross-modal fusion techniques, two critical challenges remain unresolved. First, due to the heterogeneous and often inconsistent sources from which MMKGs are constructed, the quality and informativeness of modalities vary significantly across entities, leading to the modality weighting problem. Second, existing cross-modal fusion mechanisms predominantly emphasize modality-shared information, often at the expense of modality-specific signals that are also essential for precise alignment. To address these issues, we propose HUMEA, a novel framework that integrates hierarchical Mixture-of-Experts (MoE) with unimodal distillation. HUMEA consists of: (1) A hierarchical MoE module comprising intra-modal and inter-modal experts, which adaptively modulates modality contributions by capturing entity representations at fine-to-coarse semantic granularities. In addition, we introduce a contrastive mutual information loss to enhance expert diversity and reduce redundancy. (2) A unimodal distillation strategy that preserves modality-specific information in the fused representations through single-modality alignment and distillation, achieving a balanced integration of shared and unique modality features. Extensive experiments on two benchmark datasets, FB15K-DB15K and FB15K-YAGO15K, demonstrate state-of-the-art performance, validating the effectiveness of our approach.

Infrared imaging plays a critical role in low-light and adverse weather conditions. However, due to the distinct characteristics of infrared images, existing foundation models such as Masked Autoencoder (MAE) trained on visible data perform suboptimal in infrared image interpretation tasks. To bridge this gap, an infrared foundation model known as InfMAE was developed and pre-trained on large-scale infrared datasets. Despite its effectiveness, InfMAE still faces several limitations, including the omission of informative tokens, insufficient modeling of global associations, and neglect of non-uniform noise. In this paper, we propose a Dual-domain Guided Infrared foundation model based on MAE (DuGI-MAE). First, we design a deterministic masking strategy based on token entropy, preserving only high-entropy tokens for reconstruction to enhance informativeness. Next, we introduce a Dual-Domain Guidance (DDG) module, which simultaneously captures global token relationships and adaptively filters non-uniform background noise commonly present in infrared imagery. To facilitate large-scale pretraining, we construct Inf-590K, a comprehensive infrared image dataset encompassing diverse scenes, various target types, and multiple spatial resolutions. Pretrained on Inf-590K, DuGI-MAE demonstrates strong generalization capabilities across various downstream tasks, including infrared object detection, semantic segmentation, and small target detection. Experimental results validate the superiority of the proposed method over both supervised and self-supervised comparison methods.

Time series anomaly detection (TSAD) is critical in various real-world applications. Due to the high cost of manual annotation, unsupervised methods are commonly employed to distinguish abnormal patterns from normal ones based on data or representation characteristics. However, the limited coverage of a single dataset often leads to misclassifying test-time normal patterns that deviate from the training distribution as anomalies. In view of this, we propose to introduce domain knowledge from auxiliary datasets (AuxSets) to enhance domain-level normality understanding in the target dataset (TargetSet). However, through in-depth analysis on the representation space of the TargetSet after incorporating AuxSets, we find that consistent knowledge about normality from homogeneous AuxSets do little help to TargetSet, while diverse knowledge from heterogeneous AuxSets can bring semantic confusion of normality for TargetSet, both of which can degrade TargetSet detection performance. To address the issue, we design DoKnowAD, a framework that introduces a Representation HyperVolume Estimation metric to identify helpful heterogeneous AuxSets, and further adopts contrastive learning to enforce loose coupling between datasets and high cohesion within single dataset to calibrate the TargetSet’s representation space, thus mitigating knowledge confusion. Extensive experiments on five popular datasets across different domains demonstrate that DoKnowAD consistently outperforms existing TSAD baselines in various metrics.

With the increasing application of high-stakes decisionmaking application in Federated Learning (FL), ensuring fairness across different populations to prevent biases against certain groups has become crucial. However, achieving group fairness (GF) in FL presents a formidable challenge due to its decentralization, which complicates the global GF estimation by the server. Moreover, distrust and fragility hinder the server from gathering GF values from unreliable clients. This challenge motivates our proposal of OursFed, a provable GF-aware FL framework that integrates a privacy pairbased contract and robust GF estimation method to address issues of distrust and fragility. Methodologically, we categorize client unreliability into two categories: active unreliability stemming from distrust and passive unreliability arising from fragility. To mitigate active unreliability, we design a privacy pair-based contract to guarantee truthful GF reporting, and enhance multivariate analysis by identifying relationships among multiple private data. To counteract passive unreliability, we develop a robust GF estimation using non-parametric techniques to smooth data and estimate probability densities and regression functions, improving per-client GF accuracy under multi-dimensional data perturbation. Theoretically, we demonstrate the efficacy of OursFed by analyzing its convergence, GF stability, and accuracy deviation. Experimentally, evaluations on two real datasets show that OursFed improves GF by 28.61% with at most 2.7% trade-off versus state-ofthe-art baselines, and synthetic experiments further confirm its effectiveness in handling fragility and distrust.

Diffusion models have advanced from text-to-image (T2I) to image-to-image (I2I) generation by incorporating structured inputs such as depth maps, enabling fine-grained spatial control. However, existing methods either train separate models for each condition or rely on unified architectures with entangled representations, resulting in poor generalization and high adaptation costs for novel conditions. To this end, we propose DivControl, a decomposable pretraining framework for unified controllable generation and efficient adaptation. DivControl factorizes ControlNet via SVD into basic components—pairs of singular vectors—which are disentangled into condition-agnostic learngenes and condition-specific tailors through knowledge diversion during multi-condition training. Knowledge diversion is implemented via a dynamic gate that performs soft routing over tailors based on the semantics of condition instructions, enabling zero-shot generalization and parameter-efficient adaptation to novel conditions. To further improve condition fidelity and training efficiency, we introduce a representation alignment loss that aligns condition embeddings with early diffusion features. Extensive experiments demonstrate that DivControl achieves state-of-the-art controllability with 36.4× less training cost, while simultaneously improving average performance on basic conditions. It also delivers strong zero-shot and few-shot performance on unseen conditions, demonstrating superior scalability, modularity, and transferability.

Large Multimodal Models (LMMs) often face a modality representation gap during pretraining: while language embeddings remain stable, visual representations are highly sensitive to contextual noise (e.g., background clutter). To address this issue, we introduce a visual comprehension stage, which we call ViCToR (Visual Comprehension via Token Reconstruction), a novel pretraining framework for LMMs. ViCToR employs a learnable visual token pool and utilizes the Hungarian matching algorithm to select semantically relevant tokens from this pool for visual token replacement. Furthermore, by integrating a visual token reconstruction loss with dense semantic supervision, ViCToR can learn tokens which retain high visual detail, thereby enhancing the large language model's (LLM's) understanding of visual information. After pretraining on 3 million publicly accessible images and captions, ViCToR achieves state-of-the-art results, improving over LLaVA-NeXT-8B by 10.4%, 3.2%, and 7.2% on the MMStar, SEEDI, and RealWorldQA benchmarks, respectively.

We introduce MAVERIX (Multimodal Audio-Visual Evaluation and Recognition IndeX), a unified benchmark to probe video understanding in multimodal LLMs, encompassing video, audio, and text inputs with human performance baselines. Although recent advancements in audiovisual models have shown substantial progress, the field lacks a standardized evaluation framework to thoroughly assess their cross-modality comprehension performance. MAVERIX curates 2,556 questions from 700 videos, in the form of both multiple-choice and open-ended formats, explicitly designed to evaluate multimodal models through questions that necessitate tight integration of video and audio information, spanning a broad spectrum of agentic scenarios. MAVERIX uniquely provides models with questions that closely mimic the multimodal understanding experiences available to humans during decision-making processes. To our knowledge, MAVERIX is the first benchmark aimed explicitly at assessing comprehensive audiovisual integration in such granularity. Experiments with state-of-the-art models, including Qwen 2.5 Omni and Gemini 2.5 Flash-Lite, show performance around 64% accuracy, while human experts reach near-ceiling performance of 92.8%, exposing a substantial gap to human-level comprehension. With standardized evaluation protocols, a rigorously annotated pipeline, and a public toolkit, MAVERIX establishes a challenging testbed for advancing audiovisual multimodal intelligence, with the website publicly available below.

One-shot pruning efficiently compresses Large Language Models but produces coarse sparse weights, causing significant performance degradation. Traditional fine-tuning approaches to refine these weights are prohibitively expensive for large models. This highlights the need for a training-free weight refinement method that works seamlessly with one-shot pruning and can efficiently recover the lost performance. To tackle this problem, we propose Efficient Iterative Weight Refinement (EIWR), a lightweight, plug-and-play, and training-free method that refines pruned weights through layer-wise iterative optimization. EIWR achieves efficient weight refinement via three key components: a Global Soft Constraint that eliminates costly row-wise Hessian inversions and expands the solution space; a Historical Momentum Strategy that leverages one-shot pruning priors to accelerate convergence and enhance final performance; and Neumann Series Extrapolation that significantly speeds up per-iteration computation. As a result, EIWR enables effective weight refinement with minimal time and memory overhead. Extensive experiments on LLaMA2/3 and Qwen under different pruning strategies and sparsity levels demonstrate that our method can efficiently refine sparse weights and mitigate performance degradation. For example, on LLaMA2-7B under 70 percent sparsity, EIWR reduces perplexity by 15 percent compared with SparseGPT on the WikiText2 benchmark, with only 1.81 additional minutes of computation and 1GB of additional memory.

In many practical reinforcement learning tasks, feedback is only provided at the end of a long horizon, leading to sparse and delayed rewards. Existing reward redistribution methods typically assume that per-step rewards are independent, thus overlooking interdependencies among state–action pairs. In this paper, we propose a Gaussian Process-based Likelihood Reward Redistribution (GP-LRR) framework that addresses this issue by modeling the reward function as a sample from a Gaussian Process (GP), which explicitly captures dependencies between state–action pairs through the kernel function. By maximizing the likelihood of the observed episodic return via a leave-one-out strategy that leverages the entire trajectory, our framework inherently introduces uncertainty regularization. Moreover, we show that the conventional mean squared error (MSE)-based reward redistribution arises as a special case of our GP-LRR framework when using a degenerate kernel without observation noise. When integrated with an off-policy algorithm such as Soft Actor-Critic, GP-LRR yields dense and informative reward signals, resulting in superior sample efficiency and policy performance on several MuJoCo benchmarks.

Structured mesh generation serves as a crucial preprocessing step in numerical simulations and can be formulated as a mapping problem from geometry to structured mesh. Existing approaches typically establish an isolated mapping for each geometry. This geometry-specific paradigm fails to capture and leverage commonalities across geometries, inevitably requiring recomputation or costly retraining for new geometries. To overcome this limitation, we propose ICL-Mesh, a meta-learning framework based on in-context learning (ICL) for structured mesh generation. It treats learning one mapping as one task and trains a single neural network to extract commonalities across tasks and learn from in-context examples within each task, enabling rapid generalization to unseen tasks without parameter updates. Experimental results demonstrate that ICL-Mesh effectively generalizes to diverse geometries with only a few context examples, and even without examples. It also exhibits robustness to in-context example order sensitivity and can be extended to various mesh generation scenarios, including mesh refinement and coarsening.

Multi-View Clustering (MVC) aims to enhance clustering performance by integrating multi-source complementary information. However, existing deep MVC methods face inherent challenges in balancing the learning of shared consensus representations with the preservation of view-specific information: independent encoders hinder effective cross-view collaboration, while a single shared encoder tends to sacrifice representation diversity. Although the recently introduced Mixture-of-Experts (MoE) model offers a novel approach to facilitating view collaboration, its flattened expert pool design often leads to entanglement between shared and specific information, and its routing mechanism limits collaboration potential by neglecting cross-view context. To address these challenges, this paper proposes a novel deep multi-view clustering framework—Decoupled Mixture-of-Experts with Context-Aware Routing for Multi-View Clustering (DMCAR-MVC). At its core is an innovative Decoupled MoE (D-MoE) architecture. We establish a public expert pool to learn cross-view shared representations while equipping each view with an independent private expert pool to capture its unique information, thereby structurally enforcing the decoupling of shared and specific representations. Building on this, we further design a Context-Aware Hierarchical Routing (CAHR) mechanism. When routing for the public expert pool, this mechanism introduces a global context vector to guide expert selection, enabling more efficient and globally informed cross-view collaboration. Finally, to optimize the model, we adopt a multi-level contrastive learning paradigm: on one hand, a cross-view alignment loss ensures semantic consistency in shared representations; on the other, an orthogonality constraint is imposed to further enhance separability between shared and specific representations. Extensive experiments on multiple benchmark datasets demonstrate that DMCAR-MVC significantly outperforms state-of-the-art methods across key clustering metrics. Additionally, comprehensive ablation studies thoroughly validate the effectiveness and necessity of each proposed component.

Learning multimodal representation is a fundamental task that supports a wide range of applications such as visual-text retrieval. While pioneering approaches e.g., CLIP paves the way by learning separated encoders for different modalities, they struggle to model complex interactions between modalities, resulting in inferior vision and language representation. Recently, researchers have begun to leverage powerful Large Vision-Language Models (LVLMs) for unimodal or multimodal encoding, showing substantial improvement over separated encoder methods. However, we find that directly adapting LVLMs to embedding models suffers from insufficient visual representation and coarse multimodal alignment. To address these issues, we propose a simple yet effective Fine-grained Alignment Matters (FAM) method to achieve fine-grained vision-language embedding learning with LVLMs. First, to close the gap between the pure generation and multimodal embedding using LVLMs, we propose Multi-granularity Aligned Contrastive (MAC) to explicitly learn and align fine-grained modality representations at multiple granularity levels using image-text pairs. Second, to mitigate the insufficiency of visual representation during adapting LVLMs to downstream embedding tasks, we propose a Vision Embedding Inversion Training (VEIN) strategy to encourage the extracted embeddings to preserve fine-grained visual features. Extensive experiments demonstrate the effectiveness of our method, which achieves superior performance on various downstream multimodal datasets.