AAAI 2026 Papers

Bayesian networks play a crucial role in various domains for unsupervised feature extraction and data interpretation. The Poisson gamma belief networks (PGBNs), as a type of Bayesian networks, have shown promise in analyzing high-dimensional count data. However, PGBNs encounter significant challenges when applied to sparse data, particularly in achieving accurate feature extraction and avoiding overfitting during missing value prediction. In this paper, we propose the sparse Poisson gamma belief networks (SPGBNs), a Bayesian network model designed to address these limitations. By incorporating sparse graph-structured priors over the weight matrices between adjacent layers, the proposed SPGBNs effectively capture the inherent sparsity and graph structures of latent features. Meanwhile, SPGBNs demonstrate superior generalization on missing data prediction and enable more stable extraction of meaningful latent features compared to existing approaches. Additionally, we develop an efficient Gibbs sampling algorithm that significantly improves the training stability and computational efficiency of SPGBNs. Extensive experiments on real-world datasets are conducted to validate the effectiveness of our approach.

Time series forecasting plays a critical role across a wide range of domains. Recently, an increasing number of Transformer-based forecasting models have emerged, achieving remarkably competitive performance. However, real-world time series data often exhibit complex multi-scale periodicities, which are not well-suited for modeling by the original Transformer architecture originally developed for NLP tasks. To address this limitation, we propose the Hierarchical Multi-scale Time Series Transformer (HMformer), employing a novel and sophisticated framework specifically designed for multi-scale time series forecasting. Specifically, HMformer incorporates a hierarchical cross-scale mixing mechanism that progressively aggregates temporal information from fine to coarse granularities, a scale-adaptive feature expansion design enhancing the extraction of high-level temporal semantics, and a multi-branch complementary prediction strategy for effectively integrating diverse temporal patterns. Collectively, these components enable HMformer to capture intricate, multi-scale temporal dynamics while retaining the Transformer’s inherent strength in modeling long-range dependencies. Extensive experiments conducted on multiple real-world benchmark datasets—encompassing both long-term and short-term forecasting tasks—demonstrate that HMformer achieves state-of-the-art performance.

Temporal graphs are essential for modeling complex real-world systems, such as social interactions, financial transactions, and recommendation systems, but the high computational cost and model complexity of dynamic graph neural networks (DGNNs) pose significant challenges for practical deployment. Although various pruning and sampling techniques have proven effective in accelerating static GNNs, they fall short in dynamic settings due to temporal dependencies in evolving graph structures. To address these challenges, we propose TrimDG, a general framework that accelerates DGNNs by eliminating both static and runtime redundancies. For static redundancy, we introduce a novel node influence metric, Temporal Personalized PageRank (TPP), to prune less informative nodes, and employ temporal binning to remove redundant events. For runtime redundancy during training, we develop an adaptive sampling strategy guided by graph information bottleneck and further reduce sampling frequency through temporal batch selector and sampling cache. Theoretical analysis supports our design, and experiments on real-world datasets show that TrimDG reduces runtime by an average of 83.49% across diverse DGNN backbones, while maintaining strong predictive performance, demonstrating both its efficiency and generalizability.

Bayesian optimization (BO) has been widely used to optimize expensive and gradient-free objective functions across various domains. However, existing BO methods have not addressed the objective where both inputs and outputs are functions, which increasingly arise in complex systems as advanced sensing technologies. To fill this gap, we propose a novel function-on-function Bayesian optimization (FFBO) framework. Specifically, we first introduce a function-on-function Gaussian process (FFGP) model with a separable operator-valued kernel to capture the correlations between function-valued inputs and outputs. Compared to existing Gaussian process models, FFGP is modeled directly in the function space. Based on FFGP, we define a scalar upper confidence bound (UCB) acquisition function using a weighted operator-based scalarization strategy. Then, a scalable functional gradient ascent algorithm (FGA) is developed to efficiently identify the optimal function-valued input. We further analyze the theoretical properties of the proposed method. Extensive experiments on synthetic and real-world data demonstrate the superior performance of FFBO over existing approaches.

Graph neural networks (GNNs) face dual challenges of limited structural expressiveness and opaque decision-making processes. Recent research on Subgraph Neural Networks (SGNNs) enhance model expressiveness through subgraph ensembles. However, their reliance on predefined sampling strategies leads to poor interpretability and computational inefficiency. Meanwhile, post-hoc GNN explainers enhance model interpretability but still struggle to translate their explanations into model improvements. This paper presents a novel framework that fundamentally bridges this gap by developing SGNNs with intrinsic interpretability. Our key innovation lies in constructing a self-interpretable architecture where the explanation generation mechanism is organically integrated with the prediction process. Our proposed Self-Interpretable SGNN introduces a reinforcement walk exploration (RWE-SGNN) as its data-driven sampling strategy, which can dynamically extract discriminative substructures during model training. This reinforcement walk exploration module not only provides inherent interpretability, but also enables: (1) efficient substructure extraction with less candidate number and simper embedding than traditional subgraph generation methods; and (2) provable equivalence in node coverage to traditional subgraph generation methods for connected subgraphs. Experiments on graph classification tasks show accuracy improvements over state-of-the-art GNNs, with case studies validating that the automatically identified subgraphs align with domain-specific knowledge.

With the rapid growth of the Internet of Things (IoT), integrating artificial intelligence (AI) on extremely weak embedded devices has garnered significant attention, enabling improved real-time performance and enhanced data privacy. However, the resource limitations of such devices and unreliable network conditions necessitate error-resilient device-edge collaboration systems. Traditional approaches focus on bit-level transmission correctness, which can be inefficient under dynamic channel conditions. In contrast, we propose SemanticNN, a semantic codec that tolerates bit-level errors in pursuit of semantic-level correctness, enabling compressive and resilient collaborative inference offloading under strict computational and communication constraints. It incorporates a Bit Error Rate (BER)-aware decoder that adapts to dynamic channel conditions and a Soft Quantization (SQ)-based encoder to learn compact representations. Building on this architecture, we introduce Feature-augmentation Learning, a novel training strategy that enhances offloading efficiency. To address encoder-decoder capability mismatches from asymmetric resources, we propose XAI-based Asymmetry Compensation to enhance decoding semantic fidelity. We conduct extensive experiments on STM32 using three models and six datasets across image classification and object detection tasks. Experimental results demonstrate that, under varying transmission error rates, SemanticNN significantly reduces feature transmission volume by 56.82–344.83× while maintaining superior inference accuracy.

Neural algorithmic reasoning has recently emerged as a popular research direction. It aims to train neural networks to mimic the step-by-step behavior of classical rule-based algorithms. More specifically, the execution of such algorithms can be abstracted as a sequence of states, where each state represents the intermediate outcome after an execution step. The training objective is to generate state sequences that replicate the underlying algorithmic process. A common framework for this task adopts an ``encoder-processor-decoder'' architecture, where the encoder learns representations of states, the processor simulates algorithmic steps, and the decoder reconstructs output states. While prior work has primarily focused on improving the processor, the role of the encoder in representation learning has received little attention. Most existing methods rely on simple MLP encoders, raising the question of whether such representations are sufficiently informative for supporting algorithmic reasoning. This paper investigates how to improve encoder representations for neural algorithmic reasoning. We propose a reconstruction module that aims to recover the input state from its encoded representation. This auxiliary reconstruction task encourages the encoder to retain critical information about the input. We demonstrate that incorporating this task during training improves the performance of existing neural architectures on standard benchmarks. Furthermore, we observe that current encoders often underutilize the correlations among features within a state. To address this, we draw inspiration from self-supervised learning and design an enhanced variant of the auxiliary task that encourages the encoder to capture intra-state feature dependencies. Experimental results show that our method enables the encoder to learn richer representations, thereby enhancing the performance of existing processors on algorithmic reasoning tasks.

Multimodal sentiment analysis (MSA) seeks to decode human emotions by integrating heterogeneous modalities. However, real-world scenarios often involve missing or misaligned data due to sensor failures or transmission errors, leading to disrupted temporal dynamics and degraded cross-modal correlations. To address these challenges, we propose RECAP (REcovery of Coherent Affective Patterns), a robust two-stage framework to restore temporal and structural emotional integrity under modality incompleteness. The first stage employs a causality-aware adversarial generator for multi-granularity temporal reconstruction, complemented by a contrastive mutual information factorization module that disentangles shared and modality-specific semantics. The second stage introduces a mutual information-guided attention fusion mechanism with a ranking-based objective, enabling adaptive integration of complementary signals for refined prediction. Extensive experiments on MOSI, MOSEI, and SIMS under various missing-modality conditions demonstrate that RECAP consistently outperforms state-of-the-art methods. Notably, it improves ACC-7 on MOSI by 2.71 percentage points and F1 on SIMS by 6.38 percentage points. These results verify the performance of RECAP in terms of capturing fine-grained emotional cues and robustness.

Large language models have enabled sophisticated dialogue planning policy, but their reliance on LLM-generated simulation and feedback for policy optimization may introduce systematic preference bias. We present the first comprehensive analysis of preference bias in LLM-based dialogue planners, evaluating four state-of-the-art planning policies across three dialogue domains using multiple LLM families at varying scales. Our investigation reveals that all tested planners exhibit significant preference bias, systematically favoring narrow strategy sets rather than maintaining balanced distributions. User simulation emerges as the primary bias driver, while diverse persona simulation fails as an effective mitigation strategy. Most concerning, preference bias drives planners toward ethically problematic strategies that achieve short-term success while undermining real-world effectiveness and ethical standards. Our findings establish fundamental challenges for responsible deployment of LLM-based dialogue systems and provide crucial insights for developing more reliable and ethically-aligned planning approaches.

Human cognition excels at transcending sensory input and forming latent representations that structure our understanding of the world. While Large Language Model (LLM) agents demonstrate emergent reasoning and decision-making abilities, they lack a principled framework for capturing latent structures and modeling uncertainty. In this work, we explore for the first time how to bridge LLM agents with probabilistic graphical models (PGMs) to address agentic reasoning under uncertainty. To this end, we introduce Verbalized Probabilistic Graphical Modeling (vPGM), a Bayesian agentic framework that (i) guides LLM agents in following key principles of PGMs through natural language and (ii) refines the resulting posterior distributions via numerical Bayesian inference. Unlike many traditional probabilistic methods requiring substantial domain expertise, vPGM bypasses expert‐driven model design, making it well‐suited for scenarios with limited assumptions. We evaluated our model on several agentic reasoning tasks, both close-ended and open-ended. Our results indicate that the model effectively enhances confidence calibration and text generation quality.

Achieving a balance between low parameter count, reduced FLOPs, and high accuracy and throughput remains a central challenge in neural network design. To address this, we propose the partial channel mechanism (PCM), which leverages the inherent redundancy in feature map channels. PCM divides feature map channels into multiple groups, each processed by distinct operations such as convolution, attention, pooling, or identity mapping. Building on this, we introduce partial attention convolution (PATConv), a novel module that efficiently fuses convolution and visual attention within a unified framework. Our results demonstrate that PATConv can fully replace both standard convolution and visual attention modules, leading to significant reductions in parameters and FLOPs. Furthermore, PATConv enables three efficient visual attention variants: Partial Channel Attention, Partial Spatial Attention, and Partial Self-Attention. To further optimize the allocation of channel splits, we propose dynamic {partial convolution (DPConv), which adaptively learns the optimal split ratio for each layer, achieving a better trade-off between speed and accuracy. By integrating PATConv and DPConv, we develop a new hybrid network family, PartialNet, which achieves superior top-1 accuracy and inference speed on ImageNet-1K, and demonstrates strong performance on COCO detection and segmentation tasks.

Model quantization is widely applied for compressing and accelerating deep neural networks (DNNs). However, conventional Quantization-Aware Training (QAT) focuses on training DNNs with uniform bit-width. The bit-width settings vary across different hardware and transmission demands, which induces considerable training and storage costs. Hence, the scheme of one-shot joint training multiple precisions is proposed to address this issue. Previous works either store a larger FP32 model to switch between different precision models for higher accuracy or store a smaller INT8 model but compromise accuracy due to using shared quantization parameters. In this paper, we introduce the Double Rounding quantization method, which fully utilizes the quantized representation range to accomplish nearly lossless bit-switching while reducing storage by using the highest integer precision instead of full precision. Furthermore, we observe a competitive interference among different precisions during one-shot joint training, primarily due to inconsistent gradients of quantization scales during backward propagation. To tackle this problem, we propose an Adaptive Learning Rate Scaling (AdaScale) technique that dynamically adapts learning rates for various precisions to optimize the training process. Additionally, we extend our Double Rounding to one-shot mixed precision training and develop a Hessian-Aware Stochastic Bit-switching (HessBit) strategy. Experimental results on the ImageNet-1K classification demonstrate that our methods have enough advantages to state-of-the-art one-shot joint QAT in both multi-precision and mixed-precision. We validate the feasibility of our method on detection and segmentation tasks, as well as on LLMs task.

Decentralized cooperative multi-agent multi-armed bandits (DeCMA2B) considers how multiple agents collaborate in a decentralized multi-armed bandit setting. Though this problem has been extensively studied in previous work, most existing methods remain susceptible to various adversarial attacks. In this paper, we first study DeCMA2B with adversarial corruption, where an adversary can corrupt reward observations of all agents with a limited corruption budget. We propose a robust algorithm, called DeMABAR, which ensures that each agent’s individual regret suffers only an additive term proportional to the corruption budget. Then we consider a more realistic scenario where the adversary can only attack a small number of agents. Our theoretical analysis shows that the DeMABAR algorithm can also almost completely eliminate the influence of adversarial attacks and is inherently robust in the Byzantine setting, where an unknown fraction of the agents can be Byzantine, i.e., may arbitrarily select arms and communicate wrong information. We also conduct numerical experiments to illustrate the robustness and effectiveness of the proposed method.

Despite the remarkable success of Large Language Models (LLMs), evaluating their outputs' quality regarding preference remains a critical challenge. While existing works usually leverage a strong LLM as the judge for comparing LLMs' response pairwisely, such a single-evaluator approach is vulnerable to cyclic preference, i.e., output A is better than B, B than C, but C is better than A, causing contradictory evaluation results. To address this, we introduce PGED (Preference Graph Ensemble and Denoise), a novel approach that leverages multiple model-based evaluators to construct preference graphs, and then ensembles and denoises these graphs for acyclic, non-contradictory evaluation results. We provide theoretical guarantees for our framework, demonstrating its efficacy in recovering the ground truth preference structure. Extensive experiments on ten benchmarks demonstrate PGED 's superiority in three applications: 1) model ranking for evaluation, 2) response selection for test-time scaling, and 3) data selection for model fine-tuning. Notably, PGED combines small LLM evaluators (e.g., Llama3-8B, Mistral-7B, Qwen2-7B) to outperform strong ones (e.g., Qwen2-72B), showcasing its effectiveness in enhancing evaluation reliability and improving model performance.

Offline reinforcement learning (RL) can learn policies from pre-collected offline datasets without interacting with the environment, but it suffers from the issue of out-of-distribution (OOD). Recent methods use the generative adversarial paradigm to learn policies, but easily fail to handle the conflict of fooling the discriminator and maximizing expected returns. In this paper, we propose a novel offline RL method named Distribution-Matching Generator-based Diffusion Policies (DMGDP). A distribution matching-based policy learning method is first developed, where the diffusion serves as the policy generator, to handle the conflict of fooling the discriminator and maximizing expected returns. Furthermore, a policy confidence mechanism based on discriminator regularization is designed to prevent the agent from taking OOD actions, with the aim of robust generative adversarial learning. We conducted extensive experiments on the D4RL benchmarks, and the results demonstrate that DMGDP outperforms state-of-the-art methods.

Sparse Mixture-of-Experts (SMoE) architectures have enabled a new frontier in scaling Large Language Models (LLMs), offering superior performance by activating only a fraction of their total parameters during inference. However, their practical deployment is severely hampered by substantial static memory overhead, as all experts must be loaded into memory. Existing post-training pruning methods, while reducing model size, often derive their pruning criteria from a single, general-purpose corpus. This leads to a critical limitation: a catastrophic performance degradation when the pruned model is applied to other domains, necessitating a costly re-pruning for each new domain. To address this generalization gap, we introduce Mosaic Pruning (MoP). The core idea of MoP is to construct a functionally comprehensive set of experts through a structured ``cluster-then-select" process. This process leverages a similarity metric that captures expert performance across different task domains to functionally cluster the experts, and subsequently selects the most representative expert from each cluster based on our proposed Activation Variability Score. Unlike methods that optimize for a single corpus, our proposed Mosaic Pruning ensures that the pruned model retains a functionally complementary set of experts, much like the tiles of a mosaic that together form a complete picture of the original model's capabilities, enabling it to handle diverse downstream tasks.Extensive experiments on various MoE models demonstrate the superiority of our approach. MoP significantly outperforms prior work, achieving a 7.24\% gain on general tasks and 8.92\% on specialized tasks like math reasoning and code generation.

Value decomposition is a central approach in multi-agent reinforcement learning (MARL), enabling centralized training with decentralized execution by factorizing the global value function into local values. To ensure individual-global-max (IGM) consistency, existing methods either enforce monotonicity constraints, which limit expressive power, or adopt softer surrogates at the cost of algorithmic complexity. In this work, we present a dynamical systems analysis of non-monotonic value decomposition, modeling learning dynamics as continuous-time gradient flow. We prove that, under approximately greedy exploration, all zero-loss equilibria violating IGM consistency are unstable saddle points, while only IGM-consistent solutions are stable attractors of the learning dynamics. Extensive experiments on both synthetic matrix games and challenging MARL benchmarks demonstrate that unconstrained, non-monotonic factorization reliably recovers IGM-optimal solutions and consistently outperforms monotonic baselines. Additionally, we investigate the influence of temporal-difference targets and exploration strategies, providing actionable insights for the design of future value-based MARL algorithms.

Domain adaptive retrieval aims to transfer knowledge from a labeled source domain to an unlabeled target domain, enabling effective retrieval while mitigating domain discrepancies. However, existing methods encounter several fundamental limitations: 1) neglecting class-level semantic alignment and excessively pursuing pair-wise sample alignment; 2) lacking either pseudo-label reliability consideration or geometric guidance for assessing label correctness; 3) directly quantizing original features affected by domain shift, undermining the quality of learned hash codes. In view of these limitations, we propose Prototype-based Semantic Consistency Alignment (PSCA), a two-stage framework for effective domain adaptive retrieval. In the first stage, a set of orthogonal prototypes directly establishes class-level semantic connections, maximizing inter-class separability while gathering intra-class samples. During the prototype learning, geometric proximity provides a reliability indicator for semantic consistency alignment through adaptive weighting of pseudo-label confidences. The resulting membership matrix and prototypes facilitate feature reconstruction, ensuring quantization on reconstructed rather than original features, thereby improving subsequent hash coding quality and seamlessly connecting both stages. In the second stage, domain-specific quantization functions process the reconstructed features under mutual approximation constraints, generating unified binary hash codes across domains. Extensive experiments validate PSCA's superior performance across multiple datasets.

Federated learning has emerged as a promising paradigm for collaborative model training while preserving data privacy. However, many existing FL methods implicitly assume that clients have sufficient computational and storage resources, making them less applicable in real-world scenarios with severe system heterogeneity. To address this, submodel extraction has recently gained attention as a promising strategy to tailor the global model to resource-constrained clients. Despite this progress, existing methods often suffer from noticeable performance gaps across clients and structural inconsistency in the extracted models, leading to degraded global performance and increased communication overhead. In this work, we propose FedLAGC, a novel federated framework that jointly tackles performance imbalance and communication inefficiency through Layer-Adaptive submodel extraction and Gradient Correction. Specifically, FedLAGC constructs client-specific submodels by selecting structurally important parameters according to layer-wise importance scores, ensuring both resource adaptiveness and architectural consistency. Additionally, we propose a lightweight correction mechanism that captures historical optimization drift, helping to align local updates with the global direction and reduce redundant communication. The rigorous convergence analysis of FedLAGC for system-heterogeneous federated learning under non-convex objectives is given. Extensive experiments on CIFAR-10 and CIFAR-100 with ResNet-18 and ResNet-34 under various system and data heterogeneity settings demonstrate the significant superiority of FedLAGC (up to 24% accuracy improvement and 3.66× communication efficiency) over state-of-the-art methods.