AAAI 2026 论文列表

This paper investigates the reliable parking space search problem in structured environments, with the objective of minimizing the linear combination of mean and standard deviation (mean-std) parking space search time. While canonical parking space search algorithms usually target the minimal expected search time, we argue that risk-averse users would like to trade expectation with its variance, leading to the reliable parking space search problem, which minimizes the mean-std search time. However, the non-additive nature of standard deviation makes the reliable parking space search problem difficult to solve with canonical search algorithms. To address the challenge, we propose a model-free reinforcement learning algorithm, namely MS-PPO, which simultaneously estimates the mean and standard deviation of the current decision-making policy's search time, and performs policy optimization via clipped mean-std advantage function maximization. MS-PPO is compared with several baseline parking space search algorithms as well as canonical reinforcement learning algorithms in a range of representative parking lot networks, and achieves the best overall performance in terms of the mean-std parking space search time. We also validate the effectiveness of MS-PPO in a real parking garage by deploying it to an autonomous vehicle testbed.

Incomplete cross-modal retrieval (ICMR) requires models to recover missing modalities and robustly align heterogeneous ones for effective retrieval. Existing methods, however, fall short in both aspects. They often rely on limited semantic cues, such as single samples or coarse category prototypes, which compromises reconstruction quality. Moreover, these approaches are vulnerable to learning spurious cross-modal correlations, thereby impairing accurate alignment and hindering retrieval performance. To address these challenges, we propose Causality-Aligned Semantic Recovery (CASR), a novel method designed to both comprehensively restore missing modalities and mitigate spurious associations between vision and language. Our CASR involves two essential components: i) the Missing Modality Imagination (MMI) module, which combines category semantic priors with relevant contextual information to achieve high-quality semantic reconstruction; ii) the Explicit Causal Alignment (ECA) module, which explicitly learns environment-invariant attention, effectively eliminating the interference of spurious correlations and improving retrieval performance. Furthermore, we extend CASR to the challenging task of Partially Aligned Cross-Modal Retrieval, where we treat unlabeled unpaired data as a form of incomplete data. By leveraging MMI and ECA modules, we are able to learn robust representations in this setting. Extensive experiments on benchmark datasets under various missing rates demonstrate that CASR achieves superior robustness and retrieval performance.

Multimodal large language models (MLLMs) demonstrate strong capabilities in multimodal understanding, reasoning, and interaction but still face the fundamental limitation of hallucinations, where they generate erroneous or fabricated information. Most existing research induces hallucinations by manually perturbing visual or instruction inputs, then uses output differences or model-generated descriptions as references to mitigate hallucinations and improve responsevisual consistency. However, these methods are constrained by model capabilities and prone to hallucination propagation. We propose Visual Clue Guided Decoding (VCGD), a novel decoding strategy that introduces an auxiliary Caption Model to generate precise visual clues during decoding for guiding model generation. It further incorporates image confidence constraints to critically suppress hallucination propagation during generation, thereby significantly improving content reliability and visual consistency. Specifically, VCGD leverages high-quality visual descriptions to guide MLLMs in correcting perceptual biases while generating answers. Furthermore, we introduce a Reinforcement Learning-based training paradigm for the Caption Model, in which a Reward Agent provides feedback on the quality of visual clues, further enhancing the accuracy of visual information. Extensive experiments across multiple benchmark datasets and state-of-the-art MLLMs demonstrate that VCGD significantly reduces hallucination rates and improves cross-modal consistency. Our method exhibits strong generalizability and scalability, offering an effective decoding enhancement strategy that can be seamlessly integrated into existing multimodal frameworks.

The restless multi-armed bandit (RMAB) framework is a popular model with applications across a wide variety of fields. However, its solution is hindered by the exponentially growing state space (with respect to the number of arms) and the combinatorial action space, making traditional reinforcement learning methods infeasible for large-scale instances. In this paper, we propose GINO-Q, a three-timescale stochastic approximation algorithm designed to learn an asymptotically optimal index policy for RMABs. GINO-Q mitigates the curse of dimensionality by decomposing the RMAB into a series of subproblems, each with the same dimension as a single arm, ensuring that complexity increases linearly with the number of arms. Unlike recently developed Whittle-index-based algorithms, GINO-Q does not require RMABs to be indexable, enhancing its flexibility and applicability. Our experimental results demonstrate that GINO-Q consistently learns near-optimal policies, even for non-indexable RMABs where Whittle-index-based algorithms perform poorly, and it converges significantly faster than existing baselines.

Existing sparse attention methods primarily target inference-time acceleration by selecting critical tokens under predefined sparsity patterns. However, they often fail to bridge the training–inference gap and lack the capacity for fine-grained token selection across multiple dimensions—such as queries, key-values (KV), and heads—leading to suboptimal performance and acceleration gains. In this paper, we introduce OmniSparse, a training-aware fine-grained sparse attention of long-video MLLMs, which is applied in both training and inference with dynamic token budget allocation. Specifically, OmniSparse contains three adaptive and complementary mechanisms: (1) query selection as lazy-active classification, aiming to retain active queries that capture broader semantic similarity, while discarding most of lazy ones that focus on limited local context and exhibit high functional redundancy with their neighbors, (2) KV selection with head-level dynamic budget allocation, where a shared budget is determined based on the flattest head and applied uniformly across all heads to ensure attention recall after selection, and (3) KV cache slimming to alleviate head-level redundancy, which selectively fetches visual KV cache according to the head-level decoding query pattern. Experimental results demonstrate that OmniSparse can achieve comparable performance with full attention, achieving 2.7x speedup during prefill and 2.4x memory reduction for decoding.

Vision-language models (VLMs) pre-trained on natural image and language data, such as CLIP, have exhibited significant potential in few-shot image recognition tasks, leading to development of various efficient transfer learning methods. These methods exploit inherent pre-learned knowledge in VLMs and have achieved strong performance on standard image datasets. However, their effectiveness is often limited when confronted with cross-domain tasks where imaging domains differ from natural images. To address this limitation, we propose Consistency-guided Multi-view Collaborative Optimization (CoMuCo), a novel fine-tuning strategy for VLMs. This strategy employs two functionally complementary expert modules to extract multi-view features, while incorporating prior knowledge-based consistency constraints and information geometry-based consensus mechanisms to enhance the robustness of feature learning. Additionally, a new cross-domain few-shot benchmark is established to help comprehensively evaluate methods on imaging domains distinct from natural images. Extensive empirical evaluations on both existing and newly proposed benchmarks suggest CoMuCo consistently outperforms current methods.

Recent studies have shown that unsupervised graph contrastive learning (GCL) is vulnerable to adversarial attacks. Automatic adversarial augmentation techniques are proposed to improve both the effectiveness and robustness of GCL. Existing methods typically regard unsupervised contrastive loss as the adversarial goal, essentially aiming to maximize inter-view instance-wise discrepancies between adversarial and original views. However, such attacks overlook intra-view neighborhood inconsistency, which hinders the robustness of GCL models against local neighborhood noises, resulting in performance degradation on low-homophily graphs. To tackle this issue, we propose a novel adversarial contrastive paradigm, named Edge self-aDversarial Augmentation for Graph Contrastive Learning (EDA-GCL). We theoretically establish that the adversarial objective of the intra-view neighborhood is equivalent to maximizing the discrepancy between bidirectional edge features. Hence, we build our adversarial framework based on edge self-adversarial learning. It generates pairwise adversarial augmentations from the original view by learning distinct neighborhood connectivity structures. The learned pairwise adversarial views are utilized for GCL model training in the minimization stage. Notably, this edge-level adversarial approach reduces the computational complexity to the level of the edge number. Experiments on various graph tasks and complex noise scenarios demonstrate the superiority and robustness of our EDA-GCL.

Long-horizon planning is crucial in complex environments, but diffusion-based planners like Diffuser are limited by the trajectory lengths observed during training. This creates a dilemma: long trajectories are needed for effective planning, yet they degrade model performance. In this paper, we introduce this extendable long-horizon planning challenge and propose a two-phase solution. First, Progressive Trajectory Extension incrementally constructs longer trajectories through multi-round compositional stitching. Second, the Hierarchical Multiscale Diffuser enables efficient training and inference over long horizons by reasoning across temporal scales. To avoid the need for multiple separate models, we propose Adaptive Plan Pondering and the Recursive HM-Diffuser, which unify hierarchical planning within a single model. Experiments show our approach yields strong performance gains, advancing scalable and efficient decision-making over long-horizons.

Dropout is a widely used regularization technique which improves the generalization ability of a model by randomly dropping neurons. In light of this, we propose Dropout Prompt Learning, which aims for applying dropout to improve the robustness of the vision-language models. Different from the vanilla dropout, we apply dropout on the tokens of the textual and visual branches, where we evaluate the token significance considering both intra-modal context and inter-modal alignment, enabling flexible dropout probabilities for each token. Moreover, to maintain semantic alignment for general knowledge transfer while encouraging the diverse representations that dropout introduces, we further propose residual entropy regularization. Experiments on 11 benchmarks show our method's effectiveness in challenging scenarios like low-shot learning, long-tail classification, and out-of-distribution generalization. Notably, our method surpasses regularization-based methods including KgCoOp by 5.10% and PromptSRC by 2.13% in performance on base-to-novel generalization.

Continual Visual Instruction Tuning (CVIT) enables Multimodal Large Language Models (MLLMs) to incrementally learn new tasks over time. However, this process is challenged by catastrophic forgetting, where performance on previously learned tasks deteriorates as the model adapts to new ones. A common approach to mitigate forgetting is architecture expansion, which introduces task-specific modules to prevent interference. Yet, existing methods often expand entire layers for each task, leading to significant parameter overhead and poor scalability. To overcome these issues, we introduce LoRA in LoRA (LiLoRA), a highly efficient architecture expansion method tailored for CVIT in MLLMs. LiLoRA shares the LoRA matrix A across tasks to reduce redundancy, applies an additional low-rank decomposition to matrix B to minimize task-specific parameters, and incorporates a cosine-regularized stability loss to preserve consistency in shared representations over time. Extensive experiments on a diverse CVIT benchmark show that LiLoRA consistently achieves superior performance in sequential task learning while significantly improving parameter efficiency compared to existing approaches.

Mechanical ventilation is essential in intensive care units (ICUs), but prolonged use increases patient risk. Reinforcement learning (RL) offers potential for optimizing ventilator management, yet its clinical adoption is limited by the lack of interpretable and realistic simulation environments. We propose an interpretable and probabilistic patient environment simulator based on action-based k-nearest neighbors and empirical transition probabilities, modeling stochastic state transitions grounded in real ICU data (MIMIC-IV and eICU). The simulator supports anomaly detection and provides probabilistic next-state distributions to enhance transparency and safety. Within this environment, we benchmark seven offline RL algorithms under clinically guided reward designs, including five distinct reward function configurations to explore the impact of reward shaping on agent behavior. Our results show that RL agents such as Double DQN and NFQ outperform empirical physician policies in meeting extubation guidelines, especially for high-severity patients. This benchmark enables standardized, interpretable evaluation of RL-based decision support tools for critical care.

Diffusion-based multimodal large language models (Diffusion MLLMs) have recently demonstrated impressive non-autoregressive generative capabilities across vision-and-language tasks. However, Diffusion MLLMs exhibit substantially slower inference than autoregressive models: Each denoising step employs full bidirectional self-attention over the entire sequence, resulting in cubic decoding complexity that becomes computationally impractical with thousands of visual tokens. To address this challenge, we propose D³ToM, a Decider-guided dynamic token merging method that dynamically merges redundant visual tokens at different denoising steps to accelerate inference in Diffusion MLLMs. At each denoising step, D³ToM uses decider tokens—the tokens generated in the previous denoising step—to build an importance map over all visual tokens. Then it maintains a proportion of the most salient tokens and merges the remainder through similarity-based aggregation. This plug-and-play module integrates into a single transformer layer, physically shortening the visual token sequence for all subsequent layers without altering model parameters. Moreover, D³ToM employs a merge ratio that dynamically varies with each denoising step, aligns with the native decoding process of Diffusion MLLMs, achieving superior performance under equivalent computational budgets. Extensive experiments show that D³ToM accelerates inference while preserving competitive performance.

Medical Large Vision-Language Models (Med-LVLMs) have shown promising results in clinical applications, but often suffer from hallucinated outputs due to misaligned visual understanding. In this work, we identify two fundamental limitations contributing to this issue: insufficient visual representation learning and poor visual attention alignment. To address these problems, we propose MedAlign, a simple, lightweight alignment distillation framework that transfers visual alignment knowledge from a domain-specific Contrastive Language-Image Pre-training (CLIP) model to Med-LVLMs. MedAlign introduces two distillation losses: a spatial-aware visual alignment loss based on visual token-level similarity structures, and an attention-aware distillation loss that guides attention toward diagnostically relevant regions. Extensive experiments on medical report generation and medical visual question answering (VQA) benchmarks show that MedAlign consistently improves both performance and interpretability, yielding more visually grounded outputs.

Ad hoc teamwork (AHT) requires agents to collaborate with previously unseen teammates, which is crucial for many real-world applications. The core challenge of AHT is to develop an ego agent that can predict and adapt to unknown teammates on the fly. Conventional RL-based approaches optimize a single expected return, which often causes policies to collapse into a single dominant behavior, thus failing to capture the multimodal cooperation patterns inherent in AHT. In this work, we introduce PADiff, a diffusion-based approach that captures agent's multimodal behaviors, unlocking its diverse cooperation modes with teammates. However, standard diffusion models lack the ability to predict and adapt in non-stationary AHT scenarios. To address this limitation, we propose a novel diffusion-based policy that integrates critical predictive information about teammates into the denoising process. Extensive experiments across three environments demonstrate that PADiff outperforms existing AHT methods significantly.

Clustering is a fundamental task in machine learning and data analysis, but it frequently fails to provide fair representation for various marginalized communities defined by multiple protected attributes -- a shortcoming often caused by biases in the training data. As a result, there is a growing need to enhance the fairness of clustering outcomes, ideally by making minimal modifications, possibly as a post-processing step after conventional clustering. A recent work initiated the study of closest fair clustering, though in a restricted scenario where data points belong to only two groups. In practice, however, data points are typically characterized by many groups, reflecting diverse protected attributes such as age, ethnicity, gender, etc. In this work, we generalize the study of the closest fair clustering problem to settings with an arbitrary number (more than two) of groups. We begin by showing that the problem is NP-hard even when all groups are of equal size -- a stark contrast with the two-group case, for which an exact algorithm exists. Next, we propose near-linear time approximation algorithms that efficiently handle arbitrary-sized multiple groups. Leveraging our closest fair clustering algorithms, we further achieve improved approximation guarantees for the fair correlation clustering problem, advancing the state-of-the-art results. Additionally, we are the first to provide approximation algorithms for the fair consensus clustering problem involving multiple (more than two) groups.

Machine learning and data processing techniques relying on covariance information are widespread as they identify meaningful patterns in unsupervised and unlabeled settings. As a prominent example, Principal Component Analysis (PCA) projects data points onto the eigenvectors of their covariance matrix, capturing the directions of maximum variance. This mapping, however, falls short in two directions: it fails to capture information in low-variance directions, relevant when, e.g., the data contains high-variance noise; and it provides unstable results in low-sample regimes, especially when covariance eigenvalues are close. CoVariance Neural Networks (VNNs), i.e., graph neural networks using the covariance matrix as a graph, show improved stability to estimation errors and learn more expressive functions in the covariance spectrum than PCA, but require training and operate in a labeled setup. To get the benefits of both worlds, we propose Covariance Scattering Transforms (CSTs), deep untrained networks that sequentially apply filters localized in the covariance spectrum to the input data and produce expressive hierarchical representations via nonlinearities. We define the filters as covariance wavelets that capture specific and detailed covariance spectral patterns. We improve CSTs' computational and memory efficiency via a pruning mechanism, and we prove that their error due to finite-sample covariance estimations is less sensitive to close covariance eigenvalues compared to PCA, improving their stability. Our experiments on age prediction from cortical thickness measurements on 4 datasets collecting patients with neurodegenerative diseases show that CSTs produce stable representations in low-data settings, as VNNs but without any training, and lead to comparable or better predictions w.r.t. more complex learning models.

Ensuring the safety of AI-enabled systems, particularly in high-stakes domains such as autonomous driving and healthcare, has become increasingly critical. Traditional formal verification tools fall short when faced with systems that embed both opaque, black-box AI components and complex stochastic dynamics. To address these challenges, we introduce LUCID (Learning-enabled Uncertainty-aware Certification of stochastIc Dynamical systems), a verification engine for certifying safety of black-box stochastic dynamical systems from a finite dataset of random state transitions. As such, LUCID is the first known tool capable of establishing quantified safety guarantees for such systems. Thanks to its modular architecture and extensive documentation, LUCID is designed for easy extensibility. LUCID employs a data-driven methodology rooted in control barrier certificates, which are learned directly from system transition data, to ensure formal safety guarantees. We use conditional mean embeddings to embed data into a Reproducing Kernel Hilbert Space (RKHS), where an RKHS ambiguity set is constructed that can be inflated to robustify the result to out-of-distribution behavior. A key innovation within LUCID is its use of a finite Fourier kernel expansion to reformulate a semi-infinite non-convex optimization problem into a tractable linear program. The resulting spectral barrier allows us to leverage the fast Fourier transform to generate the relaxed problem efficiently, offering a scalable yet distributionally robust framework for verifying safety. LUCID thus offers a robust and efficient verification framework, able to handle the complexities of modern black-box systems while providing formal guarantees of safety. These unique capabilities are demonstrated on challenging benchmarks.

Sharpness-Aware Minimization (SAM) has been proven to be an effective optimization technique for improving generalization in overparameterized models. While prior works have explored the implicit regularization of SAM in simple two-core scale-invariant settings, its behavior in more general tensorized or scale-invariant models remains underexplored. In this work, we leverage scale-invariance to analyze the norm dynamics of SAM in general tensorized models. We introduce the notion of Norm Deviation as a global measure of core norm imbalance, and derive its evolution under SAM using gradient flow analysis. We show that SAM's implicit control of Norm Deviation is governed by the covariance between core norms and their gradient magnitudes. Motivated by these findings, we propose a simple yet effective method, Deviation-Aware Scaling (DAS), which explicitly mimics this regularization behavior by scaling core norms in a data-adaptive manner. Our experiments across tensor completion, noisy training, model compression, and parameter-efficient fine-tuning confirm that DAS achieves competitive or improved performance over SAM, while offering reduced computational overhead.

Personalized Federated Learning (PFL) customizes models for each client to mitigate challenges from non-IID data, wherein a dominant strategy is model decoupling that partitions models into shared and personalized parts based on architectural priors (e.g., backbone vs. head). However, we reveal a critical flaw in this strategy: it induces "intrinsic drift," a performance degradation often more severe than the well-known client drift, which limits final accuracy. We trace this drift to a steep cliff of high loss emerging from the naive stitching of shared and personalized parts. To address this, we shift from architectural partitioning to a parameter behavior-driven paradigm. We introduce PPFL, an approach that employs a novel soft-fusion strategy guided by parameter-wise behavioral perception. PPFL dynamically infers each parameter's functional role—whether it behaves more like a 'personalist' or a 'generalist' in the current context—by synthesizing its multifaceted behavior observed during local training. Extensive experiments on image, text, and multimodal classification benchmarks show that PPFL outperforms eight state-of-the-art baselines by up to 5.3%. Moreover, it can function as a plug-in module, boosting the accuracy of vanilla FedAvg with a 16.82% absolute gain.