AAAI 2026 Papers

Expressive generative models have recently shown promise in offline reinforcement learning (RL) by capturing the complex, multimodal structure of dataset behavior. However, directly integrating these models into policy optimization introduces substantial computational and stability challenges due to the intricacies of their sampling processes. We introduce Flow Latent Policy (FLP), an offline RL framework that decouples expressivity from optimization by operating entirely in the latent space of a pre-trained, frozen flow-based behavior model. FLP learns a simple latent Gaussian policy whose samples are transformed through the flow to produce complex, behavior-aligned actions. This design enables closed-form behavior regularization via latent-space KL divergence and allows policy optimization without expensive backpropagation through the generative model. Experiments on the OGBench benchmark demonstrate that FLP achieves competitive or superior performance across diverse tasks, combining the benefits of expressive modeling and tractable optimization.

Visual reinforcement learning (RL) suffers from poor sample efficiency due to high-dimensional observations in complex tasks. While existing works have shown that vision-language models (VLMs) can assist RL, they often focus on knowledge distillation from the VLM to RL, overlooking the potential of RL-generated interaction data to enhance the VLM. To address this, we propose COVR, a collaborative optimization framework that enables the mutual enhancement of the VLM and RL policies. Specifically, COVR fine-tunes the VLM with RL-generated data to enhance the semantic reasoning ability consistent with the target task, and uses the enhanced VLM to further guide policy learning via action priors. To improve fine-tuning efficiency, we introduce two key modules: (1) an Exploration-Driven Dynamic Filter module that preserves valuable exploration samples using adaptive thresholds based on the degree of exploration, and (2) a Return-Aware Adaptive Loss Weight module that improves the stability of training by quantifying the inconsistency of sampling actions via return signals of RL. We further design a progressive fine-tuning strategy to reduce resource consumption. Extensive experiments show that COVR achieves strong performance across various challenging visual control tasks.

With the increasing scale and complexity of graph data, node attributes are also becoming richer and more complex, particularly in the form of informative text. Classic GNNs equipped with shallow attribute encoders are no longer sufficient to handle such data independently, making model collaboration across heterogeneous architectures an inevitable trend. Recently, the integration of Large Language Models (LLMs) and GNNs has attracted significant attention, yet the inherent disparity between these models remains a key challenge. Promising solutions have considered fine-tuning Small Language Models (SLMs) to bridge the gap between GNNs and frozen LLMs. However, this introduces another problem: these heterogeneous models bring complementary knowledge, but how to effectively integrate them and allow mutual refinement becomes a significant research gap. To address these challenges, we introduce COLA, a collaborative large–small model framework that enables seamless cooperation among semantic LLMs, task-specific fine-tuned SLMs, and structure-aware GNNs. COLA features a unique Consensus–Complement Coordination Mechanism (C3M), wherein its Mixture-of-Coordinators (MoC) architecturally aligns the LLM and SLM. Built upon this, a flexible graph-knowledge infusion strategy encourages the joint alignment and graph knowledge learning of textual representations. Extensive evaluations across nine diverse datasets show that COLA consistently achieves state-of-the-art performance, validating the effectiveness and generality of our collaborative paradigm.

Multi-view representation learning, which utilizes multiple channels to improve perceptual accuracy, is recognized for its effectiveness in the analysis of multi-view data. However, deploying these methods in real-world scenarios presents two primary challenges. 1) Lack of Variegation: Multi-view representation techniques commonly observe along a singular axis, i.e., the attribute axis; 2) Insufficient Relationship: Most multi-view models lack mechanisms for exploring potential relationships between attribute axis and channel axis. To mitigate these obstacles, we design a Dual Impulse Network framework for multi-view representation learning (DIN) to train a feature representation. In this framework, a strategy observed along the channel axis and attribute axis simultaneously is introduced, and two different representations are generated by two analogous impulse networks, which are capable of extracting information corresponding to different axes. Furthermore, we incorporate an integration network that analyzes the potential relationship between attribute axis and channel axis to generate two attention matrices. The final two feature representations derived from these attention matrices are aggregated to amplify the expression of internal information. Comprehensive experimental results support the efficacy and superiority of the proposed framework, demonstrating improvements in classification performance compared to state-of-the-art methods.

Despite the remarkable performance of deep models in medical imaging, they still require source data for training, which limits their potential in light of privacy concerns. Federated learning (FL), as a decentralized learning framework that trains a shared model with multiple hospitals (a.k.a., FL clients), provides a feasible solution. However, data heterogeneity and resource costs hinder the deployment of FL models, especially when using vision language models (VLM). To address these challenges, we propose a novel contrastive language-image pre-training (CLIP) based FL approach for medical image classification. Specifically, we introduce a masked feature adaptation module (FAM) as a communication module to reduce the communication load while freezing the CLIP encoders to reduce the computational overhead. Furthermore, we propose a masked multi-layer perceptron (MLP) as a private local classifier to adapt to the client tasks. Moreover, we design an adaptive Kullback-Leibler (KL) divergence-based distillation regularization method to enable mutual learning between FAM and MLP. Finally, we incorporate model compression to transmit the FAM parameters while using ensemble predictions for classification. Extensive experiments on four publicly available medical datasets demonstrate that our model provides feasible performance (e.g., 8% higher compared to second best baseline on ISIC2019) with reasonable resource cost (e.g., 120 times faster than FedAVG).

Data-protection regulations such as the GDPR grant every participant in a federated system a right to be forgotten. Federated unlearning has therefore emerged as a research frontier, aiming to remove a specific party's contribution from the learned model while preserving the utility of the remaining parties. However, most unlearning techniques focus on Horizontal Federated Learning (HFL), where data are partitioned by samples. In contrast, Vertical Federated Learning (VFL) allows organizations that possess complementary feature spaces to train a joint model without sharing raw data. The resulting feature-partitioned architecture renders HFL-oriented unlearning methods ineffective. In this paper, we propose ReMisVFU, a plug-and-play representation-misdirection framework that enables fast, client-level unlearning in splitVFL systems. When a deletion request arrives, the forgetting party collapses its encoder output to a randomly sampled anchor on the unit sphere, severing the statistical link between its features and the global model. To maintain utility for the remaining parties, the server jointly optimizes a retention loss and a forgetting loss, aligning their gradients via orthogonal projection to eliminate destructive interference. Evaluations on public benchmarks show that ReMisVFU suppresses back-door attack success to the natural class-prior level and sacrifices only about 2.5% points of clean accuracy, outperforming state-of-the-art baselines.

Contrastive vision-language models like CLIP have achieved impressive results in image-text retrieval by aligning image and text representations in a shared embedding space. However, these models often treat text as flat sequences, limiting their ability to handle complex, compositional, and long-form descriptions. In particular, they fail to capture two essential properties of language: semantic hierarchy, which reflects the multi-level compositional structure of text, and semantic monotonicity, where richer descriptions should result in stronger alignment with visual content. To address these limitations, we propose HiMo-CLIP, a representation-level framework that enhances CLIP-style models without modifying the encoder architecture. HiMo-CLIP introduces two key components: a hierarchical decomposition (HiDe) module that extracts latent semantic components from long-form text via in-batch PCA, enabling flexible, batch-aware alignment across different semantic granularities, and a monotonicity-aware contrastive loss (MoLo) that jointly aligns global and component-level representations, encouraging the model to internalize semantic ordering and alignment strength as a function of textual completeness. These components work together to produce structured, cognitively aligned cross-modal representations. Experiments on multiple image-text retrieval benchmarks show that HiMo-CLIP consistently outperforms strong baselines, particularly under long or compositional descriptions.

Recently, Few-shot Learning (FSL) has become a popular task that aims to recognize new classes from only a few labeled examples and has been widely applied in fields such as natural science, remote sensing, and medical images. However, most existing methods focus only on the visual modality and compute prototypes directly from raw support images, which lack comprehensive and rich multimodal information. To address these limitations, we propose a novel Multimodal Prototype Augmentation FSL framework called MPA, including LLM-based Multi-Variant Semantic Enhancement (LMSE), Hierarchical Multi-View Augmentation (HMA), and an Adaptive Uncertain Class Absorber (AUCA). LMSE leverages large language models to generate diverse paraphrased category descriptions, enriching the support set with additional semantic cues. HMA exploits both natural and multi-view augmentations to enhance feature diversity (e.g., changes in viewing distance, camera angles, and lighting conditions). AUCA models uncertainty by introducing uncertain classes via interpolation and Gaussian sampling, effectively absorbing uncertain samples. Extensive experiments on four single-domain and six cross-domain FSL benchmarks demonstrate that MPA achieves superior performance compared to existing state-of-the-art methods across most settings. Notably, MPA surpasses the second-best method by 12.29% and 24.56% in the single-domain and cross-domain setting, respectively, in the 5-way 1-shot setting.

Real-world systems often exhibit complex behaviors and are influenced by various external factors, making the integration of exogenous variables essential for accurate and robust time series forecasting. However, modeling time series with exogenous variables remains challenging due to dynamic cross-variable dependencies and the semantic gap between numerical time series data and external contextual knowledge. Large language models (LLMs) have demonstrated powerful language understanding and knowledge representation capabilities in real-world systems, offering a promising solution to bridge this gap. Motivated by this, we propose ExoTimer, a framework that deeply integrates LLMs for time series modeling with exogenous variables. We begin by introducing an Exo-Aware Endogenous Encoder to dynamically incorporate important exogenous variable information and generate patch-level representations for endogenous variables. To leverage the rich knowledge in LLMs, a Multi-Attribute Prompt Embedding module is elaborately designed to convert heterogeneous temporal features, contextual information and task specifications into LLM-interpretable textual prompts. Additionally, we propose Bi-Hash Alignment, a lightweight cross-modal alignment mechanism that bridges textual and temporal modalities in a shared hash space. Finally, a Dual-Branch Predictor with a learnable coefficient is employed to obtain the final time series prediction by integrating temporal-text and text-temporal representations. Extensive experiments on twelve real-world datasets demonstrate that ExoTimer achieves state-of-the-art performance and exhibits generalizability and scalability in both few-shot and zero-shot scenarios.

Positional bias—where models overemphasize certain positions regardless of content—has been shown to negatively impact model performance across various tasks. While recent research has extensively examined positional bias in text generation models, its presence and effects in representation models remain underexplored. Even less is known about such biases in multimodal models. In this work, we investigate positional bias in multimodal representation models, specifically in the context of image-text retrieval. We begin by distinguishing between context importance and positional bias, and then assess the presence and extent of positional bias across different models and datasets. Our experiments demonstrate that positional bias is prevalent in multimodal models, but manifests differently across modalities: text encoders tend to exhibit bias toward the beginning of the input, whereas image encoders show bias at both the beginning and end. Furthermore, we find that this bias arises from, or is amplified by, a combination of factors, including the positional encoding scheme, training loss, context importance, and the nature of using image-text pairs in multimodal training

Facial Emotion Analysis (FEA) extends traditional facial emotion recognition by incorporating explainable, fine-grained reasoning. The task integrates three subtasks—emotion recognition, facial Action Unit (AU) recognition, and AU-based emotion reasoning—to jointly model affective states. While recent approaches leverage Vision-Language Models (VLMs) and achieve promising results, they face two critical limitations: (1) hallucinated reasoning, where VLMs generate plausible but inaccurate explanations due to insufficient emotion-specific knowledge; and (2) misalignment between emotion reasoning and recognition, caused by fragmented connections between observed facial features and final labels. We propose Facial-R1, a three-stage alignment framework that effectively addresses both challenges with minimal supervision. First, we employ instruction fine-tuning to establish basic emotional reasoning capability for reducing hallucinations. Second, we introduce reinforcement training guided by emotion and AU labels as reward signals, which explicitly aligns the generated reasoning process with the predicted emotion. Third, we design a data synthesis pipeline that iteratively leverages the prior stages to expand the training dataset, enabling scalable self-improvement of the model. Built upon this framework, we introduce FEA-20K, a benchmark dataset comprising 17,737 training and 1,688 test samples with fine-grained emotion analysis annotations. Extensive experiments across eight standard benchmarks demonstrate that Facial-R1 achieves state-of-the-art performance in FEA, with strong generalization and robust interpretability.

Existing stereotype auditing methods for Large Language Models (LLMs) typically rely on isolated rating schemes or task-specific probes, lacking theoretical grounding and failing to reveal internal organization beyond surface-level output patterns. In this paper, we introduce SCoUT (Stereotype Content-oriented Utility structure via Thurstonian modeling), a closed-loop framework that structurally models, explicitly probes, and functionally steers stereotype dimensions (warmth and competence) in LLMs. SCoUT first reconstructs a global stereotype utility structure aligned with Stereotype Content Model theory via Thurstonian comparative judgments. Across multiple open-source LLMs, this modeling achieves high pairwise-preference prediction accuracy (≥ 0.90 on larger-scale models) and exhibits strong cross-model consistency. Probing internal attention mechanisms localizes this structure to specific heads (Spearman’s ρ up to 0.83 for warmth and 0.90 for competence) and surfaces a salient asymmetry between warmth and competence. Further, targeted inference-time activation modifications on these dimension-sensitive heads consistently steer model outputs along the intended axes. By bridging behavioral measurement with internal representation and controllable steering, SCoUT offers an end-to-end framework that uncovers and interprets the latent structure of stereotypes, advancing stereotype auditing from surface detection to structural analysis.

The discovery rate of optical transients will explode to 10 million public alerts per night once the Vera C. Rubin Observatory’s Legacy Survey of Space and Time comes online, overwhelming the traditional physics-based inference pipelines. A continuous-time forecasting AI model is of interest because it can deliver millisecond-scale inference for thousands of objects per day, whereas legacy MCMC codes need hours per object. In this paper, we propose SELDON, a new continuous-time variational autoencoder for panels of sparse and irregularly time-sampled (gappy) astrophysical light curves that are nonstationary, heteroscedastic, and inherently dependent. SELDON combines a masked GRU-ODE encoder with a latent neural ODE propagator and an interpretable Gaussian-basis decoder. The encoder learns to summarize panels of imbalanced and correlated data even when only a handful of points are observed. The neural ODE then integrates this hidden state forward in continuous time, extrapolating to future unseen epochs. This extrapolated time series is further encoded by deep sets to a latent distribution that is decoded to a weighted sum of Gaussian basis functions, the parameters of which are physically meaningful. Such parameters (e.g., rise time, decay rate, peak flux) directly drive downstream prioritization of spectroscopic follow-up for astrophysical surveys. Beyond astronomy, the architecture of SELDON offers a generic recipe for interpretable and continuous-time sequence modeling in any time domain where data are multivariate, sparse, heteroscedastic, and irregularly spaced.

Accurate identification of protein active sites at the residue level is crucial for understanding protein function and advancing drug discovery. However, current methods face two critical challenges: vulnerability in single-instance prediction due to sparse training data, and inadequate modality reliability estimation that leads to performance degradation when unreliable modalities dominate fusion processes. To address these challenges, we introduce Multimodal Mixtureof-Experts with Retrieval Augmentation (MERA), the first retrieval-augmented framework for protein active site identification. MERA employs hierarchical multi-expert retrieval that dynamically aggregates contextual information from chain, sequence, and active-site perspectives through residuelevel mixture-of-experts gating. To prevent modality degradation, we propose a reliability-aware fusion strategy based on Dempster–Shafer evidence theory that quantifies modality trustworthiness through belief mass functions and learnable discounting coefficients, enabling principled multimodal integration. Extensive experiments on ProTAD-Gen and TS125 datasets demonstrate that MERA achieves state-of-the-art performance, with 90% AUPRC on active site prediction and significant gains on peptide-binding site identification, validating the effectiveness of retrieval-augmented multi-expert modeling and reliability-guided fusion

Model merging combines expert models for multitask performance but faces challenges from parameter interference. This has sparked recent interest in controllable model merging, giving users the ability to explicitly balance performance trade-offs. Existing approaches employ a compile-then-query paradigm, performing a costly offline multi-objective optimization to enable fast, preference-aware model generation. This offline stage typically involves iterative search or dedicated training, with complexity that grows exponentially with the number of tasks. To overcome these limitations, we shift the perspective from parameter-space optimization to a direct correction of the model's final representation. Our approach models this correction as an optimal linear transformation, yielding a closed-form solution that replaces the entire offline optimization process with a single-step, architecture-agnostic computation. This solution directly incorporates user preferences, allowing a Pareto-optimal model to be generated on-the-fly with complexity that scales linearly with the number of tasks. Experimental results show our method generates a superior Pareto front with more precise preference alignment and drastically reduced computational cost.

Training large language models (LLMs) is fundamentally constrained by limited device memory and costly inter-device communication. Although pipeline parallelism alleviates memory pressure by partitioning models across devices, it incurs activation communication overhead that scales linearly with sequence length, limiting efficiency in long-context training. Recent weight-passing approaches (e.g., WeiPipe) mitigate this by transmitting model weights instead of activations, but suffer from redundant peer-to-peer (P2P) transfers and underutilized intra-node bandwidth. We propose TawPipe—topology-aware weight pipeline parallelism, which exploits hierarchical bandwidth in distributed clusters for improved communication efficiency. TawPipe: (i) groups devices based on topology to optimize intra-node collective and inter-node P2P communication; (ii) assigns each device a fixed shard of model weights and gradients, avoiding redundant transfers; and (iii) overlaps communication with computation to hide latency. Unlike global collective operations used in fully sharded data parallelism (FSDP), TawPipe confines most communication within node boundaries, significantly reducing cross-node traffic. Extensive experiments on up to 24 GPUs with LLaMA‑style models show that TawPipe achieves superior throughput and scalability compared to state-of-the-art baselines.

Training physics-informed neural networks (PINNs) can be viewed as a multi-task optimization problem, where data-driven and physics-driven loss functions must be simultaneously minimized, despite the potential competition between them. Manually tuning the weight coefficients for various loss terms in PINNs is often time-consuming and lacks a systematic approach. To address this challenge, this work proposes an adaptive loss balancing framework for PINNs, using multi-objective optimization (MOO) algorithms to dynamically balance competing loss terms during training. Specifically, the Non-dominated Sorting Genetic Algorithm II (NSGA-II) is integrated into the PINN training process to explore the Pareto front of the multiple objectives. A novel variance-aware relative improvement (VARI) weighting method is proposed to translate Pareto-optimal information into adaptive loss weights. The proposed MOO-VARI method is validated through several examples, where the results show that the MOO-VARI PINN consistently outperforms standard PINN and other state-of-the-art adaptive weighting strategies in terms of convergence speed, predictive accuracy, and parameter estimation performance.

Mixture-of-Experts (MoE) architecture with experts parallelism scales LLMs efficiently by activating only a subset of experts per input, avoiding proportional training costs. However, the intensive and heterogeneous communication substantially hinders the efficiency and scalability of MoE training in the resource-constrained scenario. Existing communication compression techniques fall short in MoE training due to: (i) Intensive training amplifies compression overhead, compromising training efficiency; (ii) Accumulated compression errors propagate through the network, degrading training quality. In this paper, we propose RCMoE, a communication-efficient Random Compression framework for MoE training with two core modules: (1) Local-Stochastic Quantization compresses the all-to-all communication by stochastically quantizing each row of the expert's intermediate computing results in parallel, effectively improving the compression efficiency and reducing compression error; (2) Probabilistic Thresholding Sparsification compresses the all-reduce communication by probabilistically sampling large gradients at high probability, thereby reducing the computational complexity and maintaining the convergence efficiency. Experiments on four typical MoE training tasks prove that RCMoE achieves higher 5.9x-8.1x total communication compression ratios and 1.3x-10.1x training speedup compared with the state-of-the-art compression techniques while maintaining the MoE training accuracy.

Hyperparameter optimization (HPO) is a crucial step in achieving strong predictive performance. Yet, the impact of individual hyperparameters on model generalization is highly context-dependent, prohibiting a one-size-fits-all solution and requiring opaque HPO methods to find optimal configurations. However, the black-box nature of most HPO methods undermines user trust and discourages adoption. To address this, we propose a game-theoretic explainability framework for HPO based on Shapley values and interactions. Our approach provides an additive decomposition of a performance measure across hyperparameters, enabling local and global explanations of hyperparameters' contributions and their interactions. The framework, named HyperSHAP, offers insights into ablation studies, the tunability of learning algorithms, and optimizer behavior across different hyperparameter spaces. We demonstrate HyperSHAP's capabilities on various HPO benchmarks to analyze the interaction structure of the corresponding HPO problems, demonstrating its broad applicability and actionable insights for improving HPO.