AAAI 2026 论文列表

Control in high-dimensional action spaces remains a fundamental challenge in reinforcement learning (RL), primarily due to inefficient exploration of the action space. While recent methods attempt to guide exploration, they often fall short of achieving the agility and coordination exhibited in biological motor control. Inspired by how organisms exploit muscle synergies for efficient movement, we propose Explore to Learn (ETL), a two-stage framework that first discovers fundamental synergy patterns and then leverages them for task-specific policy learning. In the first stage, ETL discovers underlying synergy patterns by deploying a targeted exploration policy. These patterns are modeled as latent directions in a low-dimensional space, along which the agent is guided to collect diverse and structured muscle activation trajectories. A variational autoencoder (VAE) is then trained to encode high-dimensional actions into a latent space whose dimensions correspond to the synergy patterns. In the second stage, the policy is trained entirely in this synergy-aware latent space, producing synergy coefficients that the decoder maps back to full-dimensional muscle actions. This structured representation significantly reduces the complexity of learning, while the decoder is further fine-tuned to enhance expressiveness and generalization across downstream tasks. Extensive experiments across musculoskeletal environments and the DMControl suite demonstrate that ETL consistently outperforms prior methods in both exploration efficiency and control performance, achieving superior scalability and generalization in overactuated control tasks.

Reinforcement learning (RL) has achieved promising results in continuous control tasks, where efficient exploration of the state space is crucial for success. However, many recent RL approaches still struggle with sample inefficiency and insufficient exploration for long-horizon tasks, particularly in environments characterized by high-dimensional and complex state spaces. To address these challenges, we propose a novel exploration framework, Latent State Predictive Exploration (LSPE). The core idea behind LSPE is to endow the agent with a form of ``foresight" to enhance exploration in long-horizon settings. Specifically, LSPE employs a state encoder to learn compact latent representations from high-dimensional visual observations, effectively filtering out irrelevant or noisy information. To further enrich and stabilize these representations, we incorporate a diffusion-based self-predictive module that enforces temporal consistency by predicting future states, thereby improving both exploration and downstream predictive control. Additionally, we introduce an Exploration Reward Function (ERF) that explicitly encourages the agent to visit novel latent states. This reward signal promotes more efficient and scalable exploration in complex environments. We evaluate LSPE across a diverse set of challenging long-horizon navigation and manipulation tasks, spanning simulation environments such as Habitat and Robosuite, as well as deployment on a real robot in a **physical indoor environment**. Experimental results show that LSPE substantially enhances exploration efficiency and scales effectively to complex, high-dimensional tasks.

Multimodal sarcasm detection is a complex task that requires distinguishing subtle complementary signals across modalities while filtering out irrelevant information. Many advanced methods rely on learning shortcuts from datasets rather than extracting intended sarcasm-related features. However, our experiments show that shortcut learning impairs the model's generalization in real-world scenarios. Furthermore, we reveal the weaknesses of current modality fusion strategies for multimodal sarcasm detection through systematic experiments, highlighting the necessity of focusing on effective modality fusion for complex emotion recognition. To address these challenges, we construct MUStARD++R by removing shortcut signals from MUStARD++. Then, a Multimodal Conditional Information Bottleneck (MCIB) model is introduced to enable efficient multimodal fusion for sarcasm detection. Experimental results show that the MCIB achieves the best performance without relying on shortcut learning.

Multi-label learning is a practical machine learning paradigm dealing with instances associated with multiple labels simultaneously. Most existing multi-label learning studies are designed under the closed-world assumption, i.e. a fixed size of label space. However, it encounters significant difficulties in open-set scenarios, where test data may contain unknown labels absent from the training set to be recognized. Existing method typically tackles this challenging problem through sub-labeling approximations and prototype-based comparisons, which often overlooks the implicit information carried by unknown labels. To address this, we propose a novel framework CREM, i.e. Classifier-induced REciprocal point for Multi-label open-set recognition, which rethinks the above problem from the reciprocal point perspective. Specifically, reciprocal points are formulated by explicitly constraining the opposition feature space to a learnable bounded margin. Then reciprocal points can be induced through the classifier with the instance-wise bias eliminated. Subsequently, a unified optimization framework is introduced to jointly facilitate the classifier and reciprocal points induction. Extensive experiments demonstrate the effectiveness and superiority of the proposed CREM approach in the multi-label open-set recognition paradigm.

Tensor-based multi-view subspace clustering (MVSC) has achieved significant success by capturing high-order inter-view correlations. However, existing approaches face two principal limitations. First, most methods either exclusively emphasize the inter-view low‑rankness (R) prior while neglecting the intra-view local smoothness (S) prior, or treat R and S as two separate regularizers—complicating joint optimization. Second, conventional tensor‑based methods impose only low‑rank constraints on the representation tensor, which limits their ability to simultaneously model consistency and complementary information. To address these issues, we propose a Unified View Extraction with Low‑Rankness and Smoothness Fusion (UVELRS) method. Our framework first extracts a consistent cross‑view representation and then constructs a tensor by stacking these representations. We introduce a novel tensor total variation Schatten-p norm that simultaneously encodes both R and S priors while offering flexible singular‑value control. This unified formulation effectively captures both high-order inter-view correlations and intra-view local smoothness. Extensive experiments on real‑world datasets demonstrate UVELRS's superior performance and robustness.

Knowledge distillation based on large vision-language models (VLMs) has recently emerged as a significant solution to transfer knowledge from the source domain to the target domain in unsupervised domain adaptation (UDA) tasks. However, existing methods employ a two-stage training pipeline, which not only complicates the training procedure but also lacks interactions between the source and target domains, severely hindering real-time cross-domain knowledge transfer. To address these challenges, we propose End-to-End Knowledge Distillation for UDA with large VLMs (termed as EKDA). (1) EKDA employs a lightweight prompt learning mechanism to first embed the knowledge from the source domain into VLMs, and then simultaneously utilize the image encoder and text encoder of VLMs to perform knowledge distillation on the target domain, significantly reducing the domain gap. (2) EKDA designs a teacher-student alternating training strategy to implement real-time collaborative interactions across domains, enabling an end-to-end paradigm to provide accurate source domain-aware supervision for the target domain. We conduct extensive experiments on 4 widely recognized benchmark datasets including Office-31, Office-Home, VisDA-2017, and Mini-DomainNet. Experimental results demonstrate that EKDA achieves significant performance improvement over the state-of-the-art UDA approaches, while maintaining a much lower model complexity. Take Office-Home for example, EKDA has gained at least 2.7% performance improvement while reducing the learnable parameters by over 80% compared with the state-of-the-art UDA baselines.

Early warning of intraoperative adverse events plays a vital role in reducing surgical risk and improving patient safety. While deep learning has shown promise in predicting the single adverse event, several key challenges remain: overlooking adverse event dependencies, underutilizing heterogeneous clinical data, and suffering from the class imbalance inherent in medical datasets. To address these issues, we construct the first Multi-label Adverse Events dataset (MuAE) for intraoperative adverse events prediction, covering six critical events. Next, we propose a novel Transformer-based multi-label learning framework (IAENet) that combines an improved Time-Aware Feature-wise Linear Modulation (TAFiLM) module for static covariates and dynamic variables robust fusion and complex temporal dependencies modeling. Furthermore, we introduce a Label-Constrained Reweighting Loss (LCRLoss) with co-occurrence regularization to effectively mitigate intra-event imbalance and enforce structured consistency among frequently co-occurring events. Extensive experiments demonstrate that IAENet consistently outperforms strong baselines on 5, 10, and 15-minute early warning tasks, achieving improvements of +5.05%, +2.82%, and +7.57% on average F1 score. These results highlight the potential of IAENet for supporting intelligent intraoperative decision-making in clinical practice.

Graph data often exhibits complex geometric heterogeneity, where structures with varying local curvature, such as tree-like hierarchies and dense communities, coexist within a single network. Existing geometric GNNs, which embed graphs into single fixed-curvature manifolds or discrete product spaces, struggle to capture this diversity. We introduce Adaptive Riemannian Graph Neural Networks (ARGNN), a novel framework that learns a continuous and anisotropic Riemannian metric tensor field over the graph. It allows each node to determine its optimal local geometry, enabling the model to fluidly adapt to the graph's structural landscape. Our core innovation is an efficient parameterization of the node-wise metric tensor, specializing to a learnable diagonal form that captures directional geometric information while maintaining computational tractability. To ensure geometric regularity and stable training, we integrate a Ricci flow-inspired regularization that smooths the learned manifold. Theoretically, we establish the rigorous geometric evolution convergence guarantee for ARGNN and provide a continuous generalization that unifies prior fixed or mixed-curvature GNNs. Empirically, our method demonstrates superior performance on both homophilic and heterophilic benchmark datasets with the ability to capture diverse structures adaptively. Moreover, the learned geometries both offer interpretable insights into the underlying graph structure and empirically corroborate our theoretical analysis.

Internet memes serve as widely distributed multimodal social content that conveys complex ideas through metaphorical expressions, often containing harmful implications that make accurate harmful meme detection an important problem. Reasoning knowledge extracted from large language models plays a crucial role in recent advances in harmful meme detection. However, these methods only perform reasoning analysis on memes from a single opinion, ignoring that memes are essentially products of group consensus, where their true meaning interpretation highly depends on the collision and aggregation process of diverse user viewpoints. To address this problem, we propose a Social Graph of Thought Reasoning Enhancement (SGoTRE) framework for harmful meme detection. The SGoTRE contains three key steps: First, through multi-agent simulation technology, we obtain diverse chains of thought that represent the parsing logic of users from different backgrounds toward memes, authentically restoring the diversity characteristics of group cognition. Second, we construct a Social Graph of Thought (SGoT) that effectively integrates multi-chain reasoning processes and structurally expresses the consensus and diversity of viewpoints among users. Finally, we utilize the SGoT for cognitive distillation, internalizing multi-opinion reasoning logic into a single multimodal large model SGoT-R1 to achieve efficient and interpretable harmful meme detection. Experimental results show that SGoT-R1 significantly improves detection performance on mainstream datasets. Particularly on the most challenging FHM dataset, SGoT-R1 achieves an 8.9% improvement over state-of-the-art models.

Typical deep clustering methods, while achieving notable progress, can only provide one clustering result per dataset. This limitation arises from their assumption of a fixed underlying data distribution, which may fail to meet user needs and provide unsatisfactory clustering outcomes. Our work investigates how multi-modal large language models (MLLMs) can be leveraged to achieve user-driven clustering, emphasizing their adaptability to user-specified semantic requirements. However, directly using MLLM output for clustering has risks for producing unstructured and generic image descriptions instead of feature-specific and concrete ones. To address these issues, our method first discovers that MLLMs' hidden states of text tokens are strongly related to the corresponding features, and leverages these embeddings to perform clusterings from any user-defined criteria. We also employ a lightweight clustering head augmented with pseudo-label learning, significantly enhancing clustering accuracy. Extensive experiments demonstrate its competitive performance on diverse datasets and metrics.

Decision-focused learning (DFL) has emerged as a powerful end-to-end alternative to conventional predict-then-optimize (PTO) pipelines by directly optimizing predictive models through downstream decision losses. Existing DFL frameworks are limited by their strictly sequential structure, referred to as sequential DFL (S-DFL). However, S-DFL fails to capture the bidirectional feedback between prediction and optimization in complex interaction scenarios. In view of this, we first time propose recursive decision-focused learning (R-DFL), a novel framework that introduces bidirectional feedback between downstream optimization and upstream prediction. We further extend two distinct differentiation methods: explicit unrolling via automatic differentiation and implicit differentiation based on fixed-point methods, to facilitate efficient gradient propagation in R-DFL. We rigorously prove that both methods achieve comparable gradient accuracy, with the implicit method offering superior computational efficiency. Extensive experiments on both synthetic and real-world datasets, including the newsvendor problem and the bipartite matching problem, demonstrate that R-DFL not only substantially enhances the final decision quality over sequential baselines but also exhibits robust adaptability across diverse scenarios in closed-loop decision-making problems.

Graph-based incomplete multi-view clustering algorithms have gathered much attention due to their impressive clustering performance. However, existing methods primarily leverage intra-view correlation from observed views, while ignoring the exploration of explicit compensation relationships between different views. Moreover, these methods need post-processing to get labels, and the separate steps lack negotiation, which may lead to sub-optimal solutions. To address these issues, we propose a Cross-view Anchor Graph Learning and Factorization (AGLF) method. AGLF develops an Anchor Graph Completion (AGC) framework that explicitly learn the missing subgraph structures. Instead of requiring post-processing, AGC directly produces soft labels. By establishing a third-order tensor of soft labels, it employs the tensor Schatten p-norm to enhance anchor graph learning and factorization. To significantly improve the quality of subgraph learning, AGLF incorporates compensation subgraphs from supplementary views into the AGC framework, enabling the construction of a better anchor graph for label learning. An optimization algorithm is devised to solve the objective function. Experimental results across various datasets demonstrate the effectiveness of our method.

The Dragonfly network, with its high-radix and low-diameter structure, is a leading interconnect in high-performance computing. A major challenge is workload interference on shared network links. Parallel discrete event simulation (PDES) is commonly used to analyze workload interference. However, high-fidelity PDES is computationally expensive, making it impractical for large-scale or real-time scenarios. Hybrid simulation that incorporates data-driven surrogate models offers a promising alternative, especially for forecasting application runtime, a task complicated by the dynamic behavior of network traffic. We present SMART, a surrogate model that combines graph neural networks (GNNs) and large language models (LLMs) to capture both spatial and temporal patterns from port level router data. SMART outperforms existing statistical and machine learning baselines, enabling accurate runtime prediction and supporting efficient hybrid simulation of Dragonfly networks.

The rapid expansion of materials databases offers unprecedented opportunities for accelerating materials discovery via machine learning. However, the widespread assumption that larger datasets inherently produce better models does not hold in practice. We propose FUSION (Fusing Uncertainty with Structural Information for Optimal Neural training), an offline dataset pruning strategy that synergistically combines uncertainty quantification with crystallographic structure analysis via geometric fingerprinting, framing dataset pruning as a discrete optimization problem. Through evaluation across 3 benchmark datasets, FUSION consistently outperforms baselines, including random pruning, uncertainty sampling, weighting factor pruning, diversity sampling, and active learning. It demonstrates robust transferability across 11 diverse architectures, outperforming random pruning by 1.91–13.65% across different datasets, with an average improvement of 6.36%. Moreover, our analysis suggests that different models exhibit varying robustness characteristics when faced with pruned training data, highlighting the importance of model selection tailored to dataset composition. We identify optimal pruning points where removing just 0–8% of training data improves model performance, yielding gains up to 12.67% in specific model–dataset combinations. These results establish a new paradigm for materials informatics that prioritizes data quality over quantity, offering a pathway toward more efficient and sustainable machine learning workflows in computational materials science.

Tensor Compressive Sensing (TCS) has gained significant attention recently due to its strong ability to preserve the multidimensional structure of data. However, existing TCS methods face three critical challenges: 1) Biased approximation of tensor rank imposed by the convex surrogate Tensor Nuclear Norm (TNN) may interfere with the original low-rank structure of tensor data. 2) Vulnerability to non-Gaussian noise and outliers makes TCS methods highly susceptible to complex noise environments ubiquitous in real-world applications. 3) Most of them are confined to third-order tensors and cannot handle high-order tensor data effectively. Being aware of these, we propose Robust Tensor Compressive Sensing (RTCS) based on M-estimators with three key innovations: 1) We design a novel M-estimator-based low-rank regularizer for high-order tensors, which provides a superior approximation of tensor rank and better preserves the original data structure. 2) RTCS incorporates a robust Welsch estimator that adaptively mitigates the influence of complex noises and outliers in tensor recovery. 3) RTCS is developed to handle high-order tensors, thereby allowing for broader applicability beyond conventional third-order tensors. We further design an efficient algorithm based on the Alternating Direction Method of Multipliers (ADMM) to handle the complex optimization problem. Experiments show that RTCS consistently outperforms existing approaches across various noises.

Weight Averaging (WA) has emerged as a powerful technique for enhancing generalization by promoting convergence to a flat loss landscape, which correlates with stronger out-of-distribution performance. However, applying WA directly to multi-modal domain generalization (MMDG) is challenging: differences in optimization speed across modalities lead WA to overfit to faster-converging ones in early stages, suppressing the contribution of slower yet complementary modalities, thereby hindering effective modality fusion and skewing the loss surface toward sharper, less generalizable minima. To address this issue, we propose MBCD, a unified collaborative distillation framework that retains WA's flatness-inducing advantages while overcoming its shortcomings in multi-modal contexts. MBCD begins with adaptive modality dropout in the student model to curb early-stage bias toward dominant modalities. A gradient consistency constraint then aligns learning signals between uni-modal branches and the fused representation, encouraging coordinated and smoother optimization. Finally, a WA-based teacher conducts cross-modal distillation by transferring fused knowledge to each uni-modal branch, which strengthens cross-modal interactions and steer convergence toward flatter solutions. Extensive experiments on MMDG benchmarks show that MBCD consistently outperforms existing methods, achieving superior accuracy and robustness across diverse unseen domains.

Driving world models are used to simulate futures by video generation based on the condition of the current state and actions. However, current models often suffer serious error accumulations when predicting the long-term future, which limits practical applications. Recent studies utilize the Diffusion Transformer (DiT) as the backbone of driving world models to improve learning flexibility. However, these models are always trained on short video clips, and multiple roll-out generations struggle to produce consistent and reasonable long videos due to the training-inference gap. To this end, we propose several solutions to build a simple yet effective long-term driving world model. First, we hierarchically decouple world model learning into large motion learning and bidirectional continuous motion learning. Then, considering the continuity of driving scenes, we propose a simple distillation method where fine-grained video flows are self-supervised signals for coarse-grained flows. The distillation is designed to improve the coherence of infinite video generation. The coarse-grained and fine-grained modules are coordinated to generate long-term and temporally coherent videos. On NuScenes, compared with the state-of-the-art front-view models, our model improves FVD by 27% and reduces inference time by 85% for the video task of generating 110+ frames.

The development of multimodal large language models (MLLMs) has advanced general video understanding. However, existing video evaluation benchmarks primarily focus on non-interactive videos, such as movies and recordings. To fill this gap, this paper proposes the first omnimodal benchmark for interactive livestream videos, LiViBench. It features a diverse set of 24 tasks, highlighting the perceptual, reasoning, and livestream-specific challenges. To efficiently construct the dataset, we design a standardized semi-automatic annotation workflow that incorporates the human-in-the-loop at multiple stages. The workflow leverages multiple MLLMs to form a multi-agent system for comprehensive video description and uses a seed-question-driven method to construct high-quality annotations. All interactive videos in the benchmark include audio, speech, and real-time comments modalities. To enhance models' understanding of interactive videos, we design tailored two-stage instruction-tuning and propose a Video-to-Comment Retrieval (VCR) module to improve the model's ability to utilize real-time comments. Based on these advancements, we develop LiVi-LLM-7B, an MLLM with enhanced knowledge of interactive livestreams. Experiments show that our model outperforms larger open-source models with up to 72B parameters, narrows the gap with leading proprietary models on LiViBench, and achieves enhanced performance on general video benchmarks, including VideoMME, LongVideoBench, MLVU, and VideoEval-Pro.

Zero-shot classifier expansion aims to adapt existing model to new, unseen classes. It utilizes class attributes or textual descriptions to learn a mapping from the semantic space to the classifier's weight space, without requiring new visual training data. However, the learning process for this mapping relies solely on correlating semantic patterns with their corresponding classifier weights and lacks explicit modeling of inter-class differences. This makes it difficult for the model to capture the critical discriminative features required to define classification boundaries. To overcome this limitation, we reframe the problem from a causal perspective and introduce a novel framework driven by counterfactuals. Our method first generates factual descriptions alongside corresponding inter-class counterfactuals to pinpoint the causal attributes essential for classification, then refines these representations via a mutual purification process, and finally leverages a novel separation loss to explicitly push the factual and counterfactual classifier weights apart. This strategy forces the model to forge clearer and more discriminative classification boundaries, achieving more accurate and robust classification. Extensive experiments demonstrate that our approach significantly outperforms existing state-of-the-art methods.