AAAI 2026 Papers

Exploration is critical for cooperative multi agent reinforcement learning (MARL) to improve sample efficiency. However, existing intrinsic motivation based exploration strategies in MARL overlook the causal relationships among agents, global states, and rewards, suffering from interference by irrelevant factors and resulting in sample inefficiency. To address this issue, we propose Causality aware Efficient Exploration (CEE), a novel framework that enhances sample efficiency by inferring causal relationships between agents, global states with respect to rewards, thereby enabling causality guided exploration. Specifically, CEE operates through two components. First, CEE identifies causal relationships between global states and rewards, filtering out causally irrelevant state features that do not have a high impact on rewards to keep decision critical state information. Second, CEE discovers causal relationships between agents' behaviors and rewards to quantify each agent's contribution to collective performance. To achieve this, we introduce a causal entropy objective that promotes exploration aligned with decision critical aspects of the underlying causal structure. We provide comprehensive validation through experiments on 21 challenging tasks spanning SMAC, SMAC v2, and Google Research Football (GRF) environments. Our results demonstrate that CEE achieves superior performance in terms of sample efficiency and asymptotic performance compared to existing MARL methods.

Estimating Individual Treatment Effects (ITE) from observational data is challenging due to confounding bias. Most studies tackle this bias by balancing distributions globally, but ignore individual heterogeneity and fail to capture the local structure that represents the natural clustering among individuals, which ultimately compromises ITE estimation. While instance-level alignment methods consider heterogeneity, they similarly overlook the local structure information. To address these issues, we propose an end-to-end Multi-Prototype alignment method for ITE estimation (PITE). PITE effectively captures local structure within groups and enforces cross-group alignment, thereby achieving robust ITE estimation. Specifically, we first define prototypes as cluster centroids based on similar individuals under the same treatment. To identify local similarity and the distribution consistency, we perform instance-to-prototype matching to assign individuals to the nearest prototype within groups, and design a multi-prototype alignment strategy to encourage the matched prototypes to be close across treatment arms in the latent space. PITE not only reduces distribution shift through fine-grained, prototype-level alignment, but also preserves the local structures of treated and control groups, which provides meaningful constraints for ITE estimation. Extensive evaluations on benchmark datasets demonstrate that PITE outperforms 13 state-of-the-art methods, achieving more accurate and robust ITE estimation.

Diffusion Transformers (DiTs) have achieved state-of-the-art performance in generative modeling, yet their high computational cost hinders real-time deployment. While feature caching offers a promising training-free acceleration solution by exploiting temporal redundancy, existing methods suffer from two key limitations: (1) uniform caching intervals fail to align with the non-uniform temporal dynamics of DiT, and (2) naive feature reuse with excessively large caching intervals can lead to severe error accumulation. In this work, we analyze the evolution of DiT features during denoising and reveal that both feature changes and error propagation are highly time- and depth-varying. Motivated by this, we propose ProCache, a training-free dynamic feature caching framework that addresses these issues via two core components: (i) a constraint-aware caching pattern search module that generates non-uniform activation schedules through offline constrained sampling, tailored to the model’s temporal characteristics; and (ii) a selective computation module that selectively compute within deep blocks and high-importance tokens for cached segments to mitigate error accumulation with minimal overhead. Extensive experiments on PixArt-alpha and DiT demonstrate that ProCache achieves up to 1.96 times and 2.90 times acceleration with negligible quality degradation, significantly outperforming prior caching-based methods.

Real-world graphs or networks are usually heterogeneous, involving multiple types of nodes and relationships. Heterogeneous graph neural networks (HGNNs) can effectively handle these diverse nodes and edges, capturing heterogeneous information within the graph, thus exhibiting outstanding performance. However, most methods of HGNNs usually involve complex structural designs, leading to problems such as high memory usage, long inference time, and extensive consumption of computing resources. These limitations pose certain challenges for the practical application of HGNNs, especially for resource-constrained devices. To mitigate this issue, we propose the Spiking Heterogeneous Graph Attention Networks (SpikingHAN), which incorporates the brain-inspired and energy-saving properties of Spiking Neural Networks (SNNs) into heterogeneous graph learning to reduce the computing cost without compromising the performance. Specifically, SpikingHAN aggregates metapath-based neighbor information using a single-layer graph convolution with shared parameters. It then employs a semantic-level attention mechanism to capture the importance of different meta-paths and performs semantic aggregation. Finally, it encodes the heterogeneous information into a spike sequence through SNNs, simulating bioinformatic processing to derive a binarized 1-bit representation of the heterogeneous graph. Comprehensive experimental results from three real-world heterogeneous graph datasets show that SpikingHAN delivers competitive node classification performance. It achieves this with fewer parameters, quicker inference, reduced memory usage, and lower energy consumption.

As synthetic data proliferates across the Internet, it is often reused to train successive generations of generative models. This creates a "self-consuming loop" that can lead to training instability or *model collapse*. Common strategies to address the issue---such as accumulating historical training data or injecting fresh real data---either increase computational cost or require expensive human annotation. In this paper, we empirically analyze the latent space dynamics of self-consuming diffusion models and observe that the low-dimensional structure of latent representations extracted from synthetic data degrade over generations. Based on this insight, we propose *Latent Space Filtering* (LSF), a novel approach that mitigates model collapse by filtering out less realistic synthetic data from mixed datasets. Theoretically, we present a framework that connects latent space degradation to empirical observations. Experimentally, we show that LSF consistently outperforms existing baselines across multiple real-world datasets, effectively mitigating model collapse without increasing training cost or relying on human annotation.

Video Anomaly Detection (VAD) aims to locate events that deviate from normal patterns in videos. Traditional approaches often rely on extensive labeled data and incur high computational costs. Recent tuning-free methods based on Multimodal Large Language Models (MLLMs) offer a promising alternative by leveraging their rich world knowledge. However, these methods typically rely on textual outputs, which introduces information loss, exhibits normalcy bias, and suffers from prompt sensitivity, making them insufficient for capturing subtle anomalous cues. To address these constraints, we propose HeadHunt-VAD, a novel tuning-free VAD paradigm that bypasses textual generation by directly hunting robust anomaly-sensitive internal attention heads within the frozen MLLM. Central to our method is a Robust Head Identification module that systematically evaluates all attention heads using a multi-criteria analysis of saliency and stability, identifying a sparse subset of heads that are consistently discriminative across diverse prompts. Features from these expert heads are then fed into a lightweight anomaly scorer and a temporal locator, enabling efficient and accurate anomaly detection with interpretable outputs. Extensive experiments show that HeadHunt-VAD achieves state-of-the-art performance among tuning-free methods on two major VAD benchmarks while maintaining high efficiency, validating head-level probing in MLLMs as a powerful and practical solution for real-world anomaly detection.

Feature attribution has gained prominence as a tool for explaining model decisions, yet evaluating explanation quality remains challenging due to the absence of ground-truth explanations. To circumvent this, explanation-guided input manipulation has emerged as an indirect evaluation strategy, measuring explanation effectiveness through the impact of input modifications on model outcomes during inference. Despite the widespread use, a major concern with inference-based schemes is the distribution shift caused by such manipulations, which undermines the reliability of their assessments. The retraining-based scheme ROAR overcomes this issue by adapting the model to the altered data distribution. However, its evaluation results often contradict the theoretical foundations of widely accepted explainers. This work investigates this misalignment between empirical observations and theoretical expectations. In particular, we identify the Sign issue as a key factor responsible for residual information that ultimately distorts retraining-based evaluation. Based on the analysis, we show that a straightforward reframing of the evaluation process can effectively resolve the identified issue. Building on the existing framework, we further propose novel variants that jointly structure a comprehensive perspective on explanation evaluation. These variants largely improve evaluation efficiency over the standard retraining protocol, thereby enhancing practical applicability for explainer selection and benchmarking. Following our proposed schemes, empirical results across various data scales provide deeper insights into the performance of carefully selected explainers, revealing open challenges and future directions in explainability research.

Out-of-distribution (OOD) detection plays a critical role in ensuring the robustness of machine learning models in open-world settings. While extensive efforts have been made in vision, language, and graph domains, the challenge of OOD detection in hypergraph-structured data remains unexplored. In this work, we formalize the problem of hypergraph out-of-distribution (HOOD) detection, which aims to identify nodes or hyperedges whose high-order relational contexts differ significantly from those seen during training. We propose HyperGOOD, a unified energy-based detection framework that integrates multi-scale spectral decomposition with structure-aware uncertainty propagation. By preserving both low- and high-frequency signals and diffusing uncertainty across the hypergraph, HyperGOOD effectively captures subtle and relationally entangled anomalies. Experimental results on nine hypergraph datasets demonstrate the effectiveness of our approach, establishing a new foundation for robust hypergraph learning under distributional shifts.

Best arm identification (BAI) aims to identify the highest- performance arm among a set of K arms by collecting stochastic samples from each arm. In real-world problems, the best arm needs to satisfy additional feasibility constraints. While there is limited prior work on BAI with feasibility constraints, they typically assume the performance and con- straints are observed simultaneously on each pull of an arm. However, this assumption does not reflect most practical use cases, e.g., in drug discovery, we wish to find the most potent drug whose toxicity and solubility are below certain safety thresholds. These safety experiments can be conducted separately from the potency measurement. Thus, this requires de- signing BAI algorithms that not only decide which arm to pull but also decide whether to test for the arm’s performance or feasibility. In this work, we study feasible BAI which allows a decision-maker to choose a tuple (i, ℓ), where i ∈ [K] de- notes an arm and ℓ denotes whether she wishes to test for its performance (ℓ = 0) or any of its N feasibility constraints (ℓ ∈ [N ]). We focus on the fixed confidence setting, which is to identify the feasible arm with the highest performance, with a probability of at least 1 − δ. We propose an efficient algorithm and upper-bound its sample complexity, showing our algorithm can naturally adapt to the problem’s difficulty and eliminate arms by worse performance or infeasibility, whichever is easier. We complement this upper bound with a lower bound showing that our algorithm is asymptotically (δ → 0) optimal. Finally, we empirically show that our algorithm outperforms other state-of-the-art BAI algorithms in both synthetic and real-world datasets.

We make three novel contributions to parameter learning and inference in probabilistic sentential decision diagrams (PSDDs). First, rather than traversing the entire PSDD during parameter learning for each dataset example, we pioneer the use of determinism to focus only on the activated partition. Second, we demonstrate how to prune deterministic computation in inference, thereby eliminating the need to propagate probability over every node in the network for each query. Third, we introduce a technique that parallelizes a single circuit evaluation, rather than parallelizing individual multiplications or layer-wise inference. For both learning and inference, experimental results on benchmark PSDDs from various application domains demonstrate state-of-the-art performance.

In this work, we study the problem of offline safe imitation learning (IL). In many real-world settings, online interactions can be risky, and accurately specifying the reward and the safety cost information at each timestep can be difficult. However, it is often feasible to collect trajectories reflecting undesirable or risky behavior, implicitly conveying the behavior the agent should avoid. We refer to these trajectories as non-preferred trajectories. Unlike standard IL, which aims to mimic demonstrations, our agent must also learn to avoid risky behavior using non-preferred trajectories. In this paper, we propose a novel approach, SafeMIL, to learn a parameterized cost that predicts if the state-action pair is risky via Multiple Instance Learning. The learned cost is then used to avoid non-preferred behaviors, resulting in a policy that prioritizes safety. We empirically demonstrate that our approach can learn a safer policy that satisfies cost constraints without degrading the reward performance, thereby outperforming several baselines.

Cancer survival analysis with multimodal data is crucial for precise treatments and patient benefits. However, the following challenges prohibit integrating histopathology and genomics: (i) multimodal data is not always complete, especially for the more costly genomics data; (ii) intricate interactions between different modalities are difficult to capture and understand. To response, we propose an end-to-end framework (CIMA) that coordinates Cyclic modality generation and Multi-grained multimodal Alignment. Specifically, CIMA designs a cyclic modality reconstruction module to reciprocally impute missing modalities and infer the interactions between them. Next, it introduces the multi-grained alignment module over the imputed data and interactions to mine fine-grained alignments between histopathology (slide patches) and genomics (biological pathways). CIMA then constructs the adaptive fusion module to leverage multimodal data and alignments for survival prediction. Extensive experiments on cancer benchmark datasets demonstrate that CIMA outperforms existing methods and exhibits good interpretability, providing valuable insights into intricate relationships between pathological phenotypes and biological pathways.Our code is released in the supplementary materials.

Designing mechanical mechanisms to trace specific paths is a classic yet notoriously difficult engineering problem, characterized by a vast and complex search space of discrete topologies and continuous parameters. We introduce MechaFormer, a Transformer-based model that tackles this challenge by treating mechanism design as a conditional sequence generation task. Our model learns to translate a target curve into a domain-specific language (DSL) string, simultaneously determining the mechanism's topology and geometric parameters in a single, unified process. MechaFormer significantly outperforms existing baselines, achieving state-of-the-art path-matching accuracy and generating a wide diversity of novel and valid designs. We demonstrate a suite of sampling strategies that can dramatically improve solution quality and offer designers valuable flexibility. Furthermore, we show that the high-quality outputs from MechaFormer serve as excellent starting points for traditional optimizers, creating a hybrid approach that finds superior solutions with remarkable efficiency.

Stochastic approximation is a powerful class of algorithms with celebrated success. However, a large body of previous analysis focuses on stochastic approximations driven by contractive operators, which is not applicable in some important reinforcement learning settings like the average reward setting. This work instead investigates stochastic approximations with merely nonexpansive operators. In particular, we study nonexpansive stochastic approximations with Markovian noise, providing both asymptotic and finite sample analysis. Key to our analysis are novel bounds of noise terms resulting from the Poisson equation. As an application, we prove for the first time that classical tabular average reward temporal difference learning converges to a sample-path dependent fixed point.

Condensed datasets offer a compact representation of larger datasets, but training models directly on them or using them to enhance model performance through knowledge distillation (KD) can result in suboptimal outcomes due to limited information. To address this, we propose a method that expands condensed datasets using model inversion, a technique for generating synthetic data based on the impressions of a pre-trained model on its training data. This approach is particularly well-suited for KD scenarios, as the teacher model is already pre-trained and retains knowledge of the original training data. By creating synthetic data that complements the condensed samples, we enrich the training set and better approximate the underlying data distribution, leading to improvements in student model accuracy during knowledge distillation. Our method demonstrates significant gains in KD accuracy compared to using condensed datasets alone and outperforms standard model inversion-based KD methods by up to 11.4% across various datasets and model architectures. Importantly, it remains effective even when using as few as one condensed sample per class, and can also enhance performance in few-shot scenarios where only limited real data samples are available.

Model discovery aims to uncover governing differential equations of dynamical systems directly from experimental data. Benchmarking such methods is essential for tracking progress and understanding trade-offs in the field. While prior efforts have focused mostly on identifying single equations, typically framed as symbolic regression, there remains a lack of comprehensive benchmarks for discovering dynamical models. To address this, we introduce MDBench, an open-source benchmarking framework for evaluating model discovery methods on dynamical systems. MDBench assesses 12 algorithms on 14 partial differential equations (PDEs) and 63 ordinary differential equations (ODEs) under varying levels of noise. Evaluation metrics include derivative prediction accuracy, model complexity, and equation fidelity. We also introduce seven challenging PDE systems from fluid dynamics and thermodynamics, revealing key limitations in current methods. Our findings illustrate that linear methods and genetic programming methods achieve the lowest prediction error for PDEs and ODEs, respectively. Moreover, linear models are in general more robust against noise. MDBench accelerates the advancement of model discovery methods by offering a rigorous, extensible benchmarking framework and a rich, diverse collection of dynamical system datasets, enabling systematic evaluation, comparison, and improvement of equation accuracy and robustness.

Real-world control systems require policies that are not only high-performing but also interpretable and robust. A promising direction toward this goal is model-based control, which learns system dynamics and cost functions from historical data and then uses these models to inform decision-making. Building on this paradigm, we introduce DiffOP, a novel framework for learning optimization-based control policies defined implicitly through optimization control problems. Without relying on value function approximation, DiffOP jointly learns the cost and dynamics models and directly optimizes the actual control costs using policy gradients. To enable this, we derive analytical policy gradients by applying implicit differentiation to the underlying optimization problem and integrating it with the standard policy gradient framework. Under standard regularity conditions, we establish that DiffOP converges to an epsilon-stationary point within O(1/epsilon) iterations. We demonstrate the effectiveness of DiffOP through experiments on nonlinear control tasks and power system voltage control with constraints.

Fine-tuning large language models (LLMs) in federated settings enables privacy-preserving adaptation but suffers from cross-client interference due to model aggregation. Existing federated LoRA fine-tuning methods, primarily based on FedAvg, struggle with data heterogeneity, leading to harmful cross-client interference and suboptimal personalization. In this work, we propose FedALT, a novel personalized federated LoRA fine-tuning algorithm that fundamentally departs from FedAvg. Instead of using an aggregated model to initialize local training, each client continues training its individual LoRA while incorporating shared knowledge through a separate Rest-of-World (RoW) LoRA component. To effectively balance local adaptation and global information, FedALT introduces an adaptive mixer that dynamically learns input-specific weightings between the individual and RoW LoRA components, drawing conceptual foundations from the Mixture-of-Experts (MoE) paradigm. Through extensive experiments on NLP benchmarks, we demonstrate that FedALT significantly outperforms state-of-the-art personalized federated LoRA fine-tuning methods, achieving superior local adaptation without sacrificing computational efficiency.

At present, spectral clustering is an important branch of unsupervised learning, and its application in deep learning has been widely concerned. However, for high-dimensional sparse datasets, the complexity of network scale leads to parameter explosion, and static Gaussian kernel often has wrong preset data structure. To overcome these challenges, we propose a novel deep clustering model, Deep Clustering Based on Sparse Kolmogorov-Arnold Network (KAN) and Spectral Constraint. It contains a deep sparse clustering framework, in which sparse KAN and the orthogonal layer are designed to enhance the sparsity of the activation function matrix, reduce the number of parameters and improve the stability of model convergence. Additionally, we add an adaptive optimized affinity matrix based on spectral constraint, which overcomes the limitations of static Gaussian kernels, and improves the performance and stability of spectral constraint. Experimental results on both synthetic and real datasets demonstrate that our model outperforms existing methods in clustering performance, computational efficiency, and stability.