AAAI 2026 Papers

Semi-supervised learning (SSL) has demonstrated high performance in image classification tasks by effectively utilizing both labeled and unlabeled data. However, existing SSL methods often suffer from poor calibration, with models yielding overconfident predictions that misrepresent actual prediction likelihoods. Recently, neural networks trained with mixup that linearly interpolates random examples from the training set have shown better calibration in supervised settings. However, calibration of neural models remains under-explored in SSL settings. Although effective in supervised model calibration, random mixup of pseudolabels in SSL presents challenges due to the overconfidence and unreliability of pseudolabels. In this work, we introduce CalibrateMix, a targeted mixup-based approach that aims to improve the calibration of SSL models while maintaining or even improving their classification accuracy. Our method leverages training dynamics of labeled and unlabeled samples to identify ''easy-to-learn'' and ''hard-to-learn'' samples, which in turn are utilized in a targeted mixup of easy and hard samples. Experimental results across several benchmark datasets show that our method achieves lower expected calibration error (ECE) and superior accuracy compared to existing SSL approaches.

Robust medical image classification under input corruption and bag-level annotation remains a critical challenge in clinical AI applications. We propose QAPNet, a Quantum- Attentive Patchwise Network that integrates quantum neural encoding, additive attention-based instance reweighting, and prototype-contrastive regularization for reliable diagnosis from degraded inputs. Our framework uses a sliding-window strategy to divide each MRI medical Image into overlapping patches, where each is encoded via an 8-qubit quantum circuit using RY -based noise-sensitive layers for yielding expressive low-dimensional representations without relying on classical CNNs. A lightweight additive attention mechanism computes instance-wise importance weights that enable interpretable and noise-aware bag-level aggregation. To enhance robustness, we apply a contrastive loss that aligns clean and noisy embeddings and enforce prototype-guided clustering via class-wise centroids. We evaluate QAPNet across seven benchmark medical imaging datasets under three levels of additive Gaussian noise (σ ∈ {5%, 10%, 30%}). QAPNet consistently outperforms eight strong baselines and achieves up to +20.8% higher accuracy in OASIS (with 30% noise), +17.7% in PathMNIST, and maintains stable performance (< 4% degradation) in all settings. Ablation studies confirm the critical role of quantum encoding, attention-based aggregation, and prototype contrastive learning. These results suggest that QAPNet offers a scalable and interpretable architecture for noisy medical imaging tasks in the real world to bridge the quantum representation learning with robust clinical prediction.

Exemplar-Free Continual Learning (EFCL) restricts the storage of previous task data and is highly susceptible to catastrophic forgetting. While pre-trained models (PTMs) are increasingly leveraged for EFCL, existing methods often overlook the inherent imbalance of real-world data distributions. We discovered that real-world data streams commonly exhibit dual-level imbalances, dataset-level distributions combined with extreme or reversed skews within individual tasks, creating both intra-task and inter-task disparities that hinder effective learning and generalization. To address these challenges, we propose PANDA, a Patch-and-Distribution-Aware Augmentation framework that integrates seamlessly with existing PTM-based EFCL methods. PANDA amplifies low-frequency classes by using a CLIP encoder to identify representative regions and transplanting those into frequent-class samples within each task. Furthermore, PANDA incorporates an adaptive balancing strategy that leverages prior task distributions to smooth inter-task imbalances, reducing the overall gap between average samples across tasks and enabling fairer learning with frozen PTMs. Extensive experiments and ablation studies demonstrate PANDA's capability to work with existing PTM-based CL methods, improving accuracy and reducing catastrophic forgetting.

Domain Generalization (DG) requires models to generalize across unseen data distributions. Kernel-based theory reveals a No-Free-Lunch problem: any model with a fixed representation is fundamentally sub-optimal for all possible shifts. While large ensembles mitigate this, they are computationally expensive and remain static once trained, inheriting the same theoretical limitation. We introduce MoE² (Mixture-of-Mixtures of Experts), a framework that uses a single frozen backbone to dynamically synthesize a bespoke adapter for each input, allowing it to continuously adapt its effective kernel. We provide a theoretical grounding for this process, proving our routing mechanism is a principled non-parametric estimator for the optimal Bayes mixture of experts. We derive a generalization bound that cleanly separates the router's estimation error from the reduction in a kernel-mismatch penalty achieved via synthesis. MoE² matches or exceeds state-of-the-art ensemble baselines on major DG benchmarks while using only a single, compact model. MoE² thus provides a theoretically-grounded and lightweight alternative to large-scale ensembles for robust domain generalization.

We introduce the Probabilistic Coin Change Problem (PCCP), a novel variant of the classical Combination Coin Change Problem (CCCP), motivated by a real-world scientific inverse task. The goal of CCCP is to enumerate all unordered combinations of coin denominations that sum to a given target. In PCCP, each coin type’s value follows a discrete probability distribution, and the aggregate value of a combination of coins is thus stochastic. Given a set of such coin types and noisy observations of total sums, the task is to infer the most likely latent coin combination. To address the combinatorial and probabilistic complexity of PCCP, we propose DeepProReasoner (Deep Combinatorial Probabilistic Reasoning with Embedded Representations), an unsupervised, end-to-end, deep-learning framework that integrates combinatorial reasoning, latent-space modeling, and differentiable probabilistic reasoning. The model is trained using a reconstruction loss between the observed empirical distribution and a decoded probability mass function (PMF), enabling efficient gradient-based search over a continuous relaxation of the combinatorial space. We evaluate DeepProReasoner on two instances of PCCP: (1) a synthetic Candy Mix problem for ablation studies, and (2) a real-world task of molecular formula inference from ultrahigh resolution mass spectrometry (MS) data. Besides the two given instances, PCCP captures a wide range of inverse settings in biology, chemistry, environmental sciences, and medicine, where latent combinatorial structures give rise to noisy aggregate observations through stochastic processes. Our results show that DeepProReasoner achieves high accuracy and robustness, outperforming state-of-the-art methods.

In this paper, we rethink model agent behaviors from a geometric structure perspective in multi-agent reinforcement learning. Modeling agent behaviors is essential for understanding how agents interact and facilitating effective decisions. The key lies in capturing the dependencies and sequential relationships among agent decisions. Since each decision influences the subsequent choices, this forms a hierarchical and nested tree-like structure of interdependencies. While modeling tree-like data in Euclidean spaces could cause distortion, which results in a loss of agent decision structure information. Motivated by this, we reconsider model agent behaviors in hyperbolic space and propose the Hyperbolic Multi-Agent Representations (HMAR) method, which projects the agent behaviors into a Poincaré ball and leverages hyperbolic neural networks to learn agent policy representations. Additionally, we designed a contrastive loss function to train this network, minimizing the distance in feature space between different representations of the same agent while maximizing the distance between representations of distinct agents. Experimental results provide empirical evidence for the effectiveness of the HMAR method in cooperative and competitive environments, demonstrating the potential of hyperbolic agent representations for effective decision-making in multi-agent environments.

Vision–language models (VLMs) such as CLIP have unlocked powerful zero-shot transfer, yet efficient adaptation to downstream tasks remains challenging. Existing methods often depend on graph structures and dataset-specific tuning, making them sensitive to modality gaps and computationally costly at scale. In this paper, we propose IOTA (Inverse Optimal Transport Adaptation), a lightweight algorithm that reformulates VLMs inference from the perspective of inverse optimal transport (IOT), providing a unified view of training and inference. Under the IOT framework, IOTA enhances zero-shot alignment via a theory-guided unbalanced OT strategy and refines textual prototypes using OT-based pseudo-labels with a marginal-aware adaptive threshold, enabling reliable supervision without gradient updates. The framework naturally extends to few-shot scenarios through a label-guided masking mechanism. By decoupling image–text interactions from other inter-modal dependencies, IOTA avoids task-specific tuning and expensive affinity construction. Extensive experiments on standard benchmarks show that IOTA consistently improves zero-shot and few-shot performance while reducing memory and computation overhead, validating both its theoretical insight and plug-and-play practicality.

Lexicographic multi-objective problems, which consist of multiple conflicting subtasks with explicit priorities, are common in real-world applications. Despite the advantages of Reinforcement Learning (RL) in single tasks, extending conventional RL methods to prioritized multiple objectives remains challenging. In particular, traditional Safe RL and Multi-Objective RL (MORL) methods have difficulty enforcing priority orderings efficiently. Therefore, Lexicographic Multi-Objective RL (LMORL) methods have been developed to address these challenges. However, existing LMORL methods either rely on heuristic threshold tuning with prior knowledge or are restricted to discrete domains. To overcome these limitations, we propose Lexicographically Projected Policy Gradient RL (LPPG-RL), a novel LMORL framework which leverages sequential gradient projections to identify feasible policy update directions, thereby enabling LPPG-RL broadly compatible with all policy gradient algorithms in continuous spaces. LPPG-RL reformulates the projection step as an optimization problem, and utilizes Dykstra's projection rather than generic solvers to deliver great speedups, especially for small- to medium-scale instances. In addition, LPPG-RL introduces Subproblem Exploration (SE) to prevent gradient vanishing, accelerate convergence and enhance stability. We provide theoretical guarantees for convergence and establish a lower bound on policy improvement. Finally, through extensive experiments in a 2D navigation environment, we demonstrate the effectiveness of LPPG-RL, showing that it outperforms existing state-of-the-art continuous LMORL methods.

Decentralized online convex optimization (D-OCO), where multiple agents within a network collaboratively learn optimal decisions in real-time, arises naturally in applications such as federated learning, sensor networks, and multi-agent control. In this paper, we study D-OCO under unknown, time- and agent-varying feedback delays. While recent work has addressed this problem~\citep{nguyen2024handling}, existing algorithms assume prior knowledge of the total delay over agents and still suffer from suboptimal dependence on both the delay and network parameters. To overcome these limitations, we propose a novel algorithm that achieves an improved regret bound of Õ(N √d_tot + N √( T / √(1 − σ₂) )), where d_tot denotes the average total delay across agents, N is the number of agents, and 1 − σ₂ is the spectral gap of the network. We also prove a lower bound showing that our upper bound is tight up to logarithmic factors. Our approach builds upon recent advances in D-OCO~\citep{wan2024nearly}, but crucially incorporates an adaptive learning rate mechanism via a decentralized communication protocol. This enables each agent to estimate delays locally using a gossip-based strategy without the prior knowledge of the total delay. We further extend our framework to the strongly convex setting and derive a sharper regret bound. Experimental results validate the effectiveness of our approach, showing improvements over existing benchmark algorithms.

Online continual learning (OCL) aims at learning a non-stationary data stream in a way of reading each data sample only once, and hence suffers from the trade-off of catastrophic forgetting and insufficient learning. In this work, we firstly analytically establish relationship between loss functions and model parameters from the Bayesian perspective. Based on our analysis, we subsequently propose a parameter merging method with gradient-guided supermasks. Our method leverages 1-order and 2-order gradient information to construct supermasks that determine the merging weights between the old and new models. Our method performs direct arithmetic operations on parameters to update models, beyond traditional gradient descent. We further discover that a widely-used premise that 1-order gradients can be negligible is invalid in OCL, due to slow convergence incurred by insufficient learning. Additionally, we utilize a dual-model dual-view distillation strategy that can align output distributions of the new and merged models for each sample, further enhancing model performance. Extensive experiments are conducted on four benchmarks in OCL settings, including CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-100. Experimental results demonstrate that our method is effective, and achieves a substantial boost over previous methods.

Visual Autoregressive (VAR) models adopt a next-scale prediction paradigm, offering high-quality content generation with substantially fewer decoding steps. However, existing VAR models suffer from significant attention complexity and severe memory overhead due to the accumulation of key-value (KV) caches across scales. In this paper, we tackle this challenge by introducing KV cache compression into the next-scale generation paradigm. We begin with a crucial observation: attention heads in VAR models can be divided into two functionally distinct categories: Contextual Heads focus on maintaining semantic consistency, while Structural Heads are responsible for preserving spatial coherence. This structural divergence causes existing one-size-fits-all compression methods to perform poorly on VAR models. To address this, we propose HACK, a training-free Head-Aware KV cache Compression frameworK. HACK utilizes an offline classification scheme to separate head types, enabling it to apply pattern-specific compression strategies with asymmetric cache budgets for each category. By doing so, HACK effectively constrains the average KV cache length within a fixed budget B, reducing the theoretical attention complexity from O(n4) to O(Bn2). Extensive experiments on multiple VAR models across text-to-image and class-conditional tasks validate the effectiveness and generalizability of HACK. It achieves up to 70% KV cache compression without degrading output quality, resulting in memory savings and faster in- ference. For example, HACK provides a 1.75× memory reduction and a 1.57× speedup on Infinity-8B.

While Multimodal Large Language Models (MLLMs) show immense promise for achieving truly human-like interactions, progress is hindered by the lack of fine-grained evaluation frameworks for human-centered scenarios, encompassing both the understanding of complex human intentions and the provision of empathetic, context-aware responses. Here we introduce HumanSense, a comprehensive benchmark designed to evaluate the human-centered perception and interaction capabilities of MLLMs, with a particular focus on deep understanding of extended multimodal contexts and the formulation of rational feedback. Our evaluation reveals that leading MLLMs still have considerable room for improvement, particularly for advanced interaction-oriented tasks. Supplementing visual input with audio and text information yields substantial improvements, and Omni-modal models show advantages on these tasks.Furthermore, grounded in the observation that appropriate feedback stems from a contextual analysis of the interlocutor's needs and emotions, we posit that reasoning ability serves as the key to unlocking it. We devise a multi-stage, modality-progressive reinforcement learning approach, resulting in HumanSense-Omni-Reasoning, which substantially enhances performance on higher-level understanding and interactive tasks. Additionally, we observe that successful reasoning processes appear to exhibit consistent thought patterns. By designing corresponding prompts, we also enhance the performance of non-reasoning models in a training-free manner.

Cross-modal retrieval is a fundamental application of multi-modal learning that has achieved remarkable success with large-scale well-paired data. However, in practice, it is costly to collect large-scale well-paired data. To alleviate the dependence on the amount of paired data, in this paper, we study a practical learning paradigm: semi-paired cross-modal learning (SPL), which utilizes both a small amount of paired data and a large amount of unpaired data to enhance cross-modal learning directly and is more accessible in practice. To achieve this, we take image-text retrieval as an example and propose a novel Robust Cross-modal Semi-paired Learning method (RCSL) by addressing two challenges. To be specific, i) to overcome the under-optimization issue caused by too little paired data, we present Semi-paired Discriminative Learning (SDL) to fully learn visual-semantic associations from a small amount of image-text pairs by preserving the alignment and uniformity of modality representations. ii) To mine visual-semantic correspondences from unpaired data, RCSL first constructs pseudo-paired correlations across different modalities by nearest neighbor association. However, this may introduce noisy correspondences (NCs) due to inaccurate pseudo signals, which could degrade the model's performance. To tackle NCs, we devise Robust Cross-correlation Mining (RCM) based on the risk minimization criterion to robustly and explicitly learn visual-semantic associations from pseudo-paired data, thus boosting cross-modal learning. Finally, we conduct extensive experiments on four datasets, i.e., three widely used benchmark datasets of Flickr30K, MS-COCO, CC152K, and a newly constructed real-world dataset Drone-SP, to demonstrate the effectiveness of RCSL under semi-paired and noisy settings.

Learning on multi-view data is a fundamental task, which integrates the information from different views to improve the final performance. It is also a basic task for learning on the long-tailed data in real applications, followed by the downstream tasks, i.e., classification. The existing works for trusted classification on multi-view data or long-tailed data usually aim to improve the final performance and dynamically consider the confidence of prediction for the data which is crucial in cost-sensitive domains. However, these methods pay few attentions to the pairwise trusted problem which considers the trusted pairs instead of trusted annotated data points. Besides, the problem of classification on long-tailed multi-view data has never been studied so far. In this work, we focus on the pairwise trusted problem on long-tailed multi-view classification and give a general framework, which considers the trusted pairs instead of trusted annotated data points. We then construct a specific example under the general framework and introduce a novel Enhanced Normal-Inverse Gamma distribution (ENIG). ENIG is a joint probabilistic distribution built on Dirichlet distribution and NIG. A novel combination rule based on ENIG for long-tailed multi-view data is also given, which adaptively integrates the long-tailed data from different views to achieve a consensus one at the level of evidence and effectively produces a trusted long-tailed multi-view classification result. Our method is robust and able to be dynamically aware of the uncertainty for the long-tailed data from each view. The accurate uncertainty can be induced by the proposed learning framework, leading to both robustness and reliability for classification on long-tailed multi-view data. Experimental results on different long-tailed multi-view datasets demonstrate the effectiveness of our method in terms of accuracy, robustness and reliability.

Irregular time series (IRTS) are prevalent in real-world applications, where uneven sampling and missing data pose fundamental challenges to deep learning-based feature modeling. Although existing methods attempt to retain timestamp information, they often overlook the structured patterns embedded within the missingness itself, and tend to perform poorly when confronted with class imbalance exacerbated by data incompleteness. Specifically, temporal irregularity hinders the modeling of long-range dependencies and local patterns, while sparse observations limit representational capacity, disproportionately impairing minority classes and leading to severe classification bias. To address these deeply coupled challenges, we propose SPECTRA (Structured Pattern and Enriched Context-aware Temporal Representation Architecture), a unified framework for robust IRTS classification. SPECTRA introduces a frequency-guided observation encoder that reconstructs temporal dependencies in a stable manner, mitigating spectral distortion and information corruption. Complementarily, a missingness pattern encoder explicitly captures the dynamic evolution of missing data and leverages it as a discriminative signal. In addition, a prototype-constrained classification paradigm directly optimizes the geometric structure of the feature space, enhancing intra-class compactness and alleviating generalization bottlenecks caused by class imbalance. Extensive experiments on three public IRTS datasets—P12, P19, and PAM—demonstrate the superior performance of SPECTRA under both missing and imbalanced conditions.

Machine unlearning, as a post-hoc processing technique, has gained widespread adoption in addressing challenges like bias mitigation and robustness enhancement. However, existing non-privacy unlearning-based solutions persist in using a binary data removal framework designed for privacy-driven motivation, even when repurposed for fairness or robustness improvements. This leads to significant utility loss, a phenomenon known as “over-unlearning”. While over-unlearning has been largely described in many studies as primarily causing utility degradation, we investigate deeper insights in this work through counterfactual leave-one-out analysis. Based on insights, we introduce a soft weighting strategy that assigns tailored weights to each sample by solving a convex quadratic programming problem analytically, which enables fine-grained model adjustments to address the over-unlearning. We demonstrate that the proposed soft-weighted scheme can be seamlessly integrated into most existing unlearning algorithms. Extensive experiments show that in fairness- and robustness-driven tasks, the soft-weighted scheme significantly outperforms hard-weighted schemes in fairness/robustness metrics and alleviates the decline in utility metric, thereby enhancing unlearning algorithm as an effective correction solution.

Compositional reasoning is a critical capability for multimodal models, enabling systematic understanding of complex scenes through structured combinations of objects, attributes, and relations. However, existing research on this ability primarily focuses on vision-language models (VLMs, e.g., CLIP and SigLIP), with limited exploration of multimodal large language models (MLLMs). To address this gap, we introduce CR³, a novel framework that enhances compositional reasoning abilities of MLLMs via rule-based reinforcement learning. CR³ leverages rule-based rewards to optimize the MLLM's policy on systematically curated multimodal instruction-following tasks, guided by a model-adaptive dynamic task mixing strategy. Our approach boosts performance by over 19% on three compositional reasoning benchmarks, significantly outperforming supervised fine-tuning (SFT) by at least 12%. Crucially, CR³ demonstrates superior generalization by improving performance on out-of-domain benchmarks where SFT methods degrade, highlighting its effectiveness and data efficiency.

Diffusion and flow matching models have recently emerged as promising approaches for peptide binder design. Despite their progress, these models still face two major challenges. First, categorical sampling of discrete residue types collapses their continuous parameters into one-hot assignments, while continuous variables (e.g., atom positions) evolve smoothly throughout the generation process. This mismatch disrupts the update dynamics and results in suboptimal performance. Second, current models assume unimodal distributions for side-chain torsion angles, which conflicts with the inherently multimodal nature of side-chain rotameric states and limits prediction accuracy. To address these limitations, we introduce PepBFN, the first Bayesian flow network for full-atom peptide design that directly models parameter distributions in fully continuous space. Specifically, PepBFN models discrete residue types by learning their continuous parameter distributions, enabling joint and smooth Bayesian updates with other continuous structural parameters. It further employs a novel Gaussian mixture-based Bayesian flow to capture the multimodal side-chain rotameric states and a Matrix Fisher-based Riemannian flow to directly model residue orientations on the SO(3) manifold. Together, these parameter distributions are progressively refined via Bayesian updates, yielding smooth and coherent peptide generation. Experiments on side-chain packing, reverse folding, and binder design tasks demonstrate the strong potential of PepBFN in computational peptide design.

With the rapid advancement and strong generalization capabilities of large language models (LLMs), they have been increasingly incorporated into the active learning pipelines as annotators to reduce annotation costs. However, considering the annotation quality, labels generated by LLMs often fall short of real-world applicability. To address this, we propose a novel active learning framework, Mixture of LLMs in the Loop Active Learning, replacing human annotators with labels generated through a Mixture-of-LLMs-based annotation model, aimed at enhancing LLM-based annotation robustness by aggregating the strengths of multiple LLMs. To further mitigate the impact of the noisy labels, we introduce annotation discrepancy and negative learning to identify the unreliable annotations and enhance learning effectiveness. Extensive experiments demonstrate that our framework achieves performance comparable to human annotation and consistently outperforms single-LLM baselines and other LLM-ensemble-based approaches. Moreover, our framework is built on lightweight LLMs, enabling it to operate fully on local machines in real-world applications.