AAAI 2026 Papers

Recent advances in Multimodal Large Language Models (MLLMs) have spurred significant progress in Chain-of-Thought (CoT) reasoning. Building on the success of Deepseek-R1, researchers extended multimodal reasoning to post-training paradigms based on reinforcement learning (RL), focusing predominantly on mathematical datasets. However, existing post-training paradigms tend to neglect two critical aspects: (1) The lack of quantifiable difficulty metrics capable of strategically screening samples for post-training optimization. (2) Suboptimal post-training paradigms that fail to jointly optimize perception and reasoning capabilities. To address this gap, we propose two novel difficulty-aware sampling strategies: Progressive Image Semantic Masking (PISM) quantifies sample hardness through systematic image degradation, while Cross-Modality Attention Balance (CMAB) assesses cross-modal interaction complexity via attention distribution analysis. Leveraging these metrics, we design a hierarchical training framework that incorporates both GRPO-only and SFT+GRPO hybrid training paradigms, and evaluate them across six benchmark datasets. Experiments demonstrate consistent superiority of GRPO applied to difficulty-stratified samples compared to conventional SFT+GRPO pipelines, indicating that strategic data sampling can obviate the need for supervised fine-tuning while improving model accuracy.

Mixed-Integer Linear Programming (MILP) is a cornerstone of combinatorial optimization, yet solving large-scale instances remains a significant computational challenge. Recently, Graph Neural Networks (GNNs) have shown promise in accelerating MILP solvers by predicting high-quality solutions. However, we identify that existing methods misalign with the intrinsic structure of MILP problems at two levels. At the leaning objective level, the Binary Cross-Entropy (BCE) loss treats variables independently, neglecting their relative priority and yielding plausible logits. At the model architecture level, standard GNN message passing inherently smooths the representations across variables, msking the natural competitive relationships within constraints. To address these challenges, we propose CoCo-MILP, which explicitly models inter-variable Contrast and intra-constraint Competition for advanced MILP solution prediction. At the objective level, CoCo-MILP introduces the Inter-Variable Contrastive Loss (VCL), which explicitly maximizes the embedding margin between variables assigned one versus zero. At the architectural level, we design an Intra-Constraint Competitive GNN layer that, instead of homogenizing features, learns to differentiate representations of competing variables within a constraint, capturing their exclusionary nature. Experimental results on standard benchmarks demonstrate that CoCo-MILP significantly outperforms existing learning-based approaches, reducing the solution gap by up to 68.12% compared to traditional solvers.

Graph contrastive learning (GCL) aims to learn representations by bringing semantically similar graphs closer and pushing dissimilar ones farther apart without label supervision. Hard negatives, which refer to graphs that have different labels but similar embeddings to the target graph, play a key role in improving representation discrimination. However, current methods that generate both high-quality positives and hard negatives face two challenges: (1) Hard negative sample generation often suffers from class imbalance, resulting in unequal attention across classes and reduced discriminative power in the learned representations. (2) The typical binary positive sample generation approach, which divides the graph into important and unimportant semantic regions, overlooks regions that negatively impact semantics and mislead model predictions. To address these issues, we introduce a novel method named BalanceGCL, which enhance graph contrastive learning with balanced hard negatives and fine-grained semantic-aware positives. BalanceGCL comprises two modules: Balanced Hard Negative graphs generation (BHN) and Fine-grained Semantic-aware Positive graphs generation (FSP). Inspired by the counterfactual mechanism, BHN generates balanced hard negatives that remain structurally similar to the original graph while inducing a controlled semantic shift. To ensure class balance, BHN iteratively constructs one hard negative sample for each class, ensuring an even distribution of negative samples across all alternative categories. FSP leverages the semantic differences between original graphs and balanced hard negatives to identify positively contributing, negatively contributing, and unimportant regions. By enhancing the influence of positive contributors, suppressing negative ones, and perturbing unimportant areas, it generates more reliable and semantically complete positive samples. The proposed method outperforms state-of-the-art GCL techniques across 14 datasets in graph classification and transfer learning tasks, demonstrating its effectiveness in tackling class imbalance and identifying fine-grained semantic-aware regions.

Despite recent advances in fairness-aware machine learning, predictive models often exhibit discriminatory behavior towards marginalized groups. Such unfairness might arise from biased training data, model design, or representational disparities across groups, posing significant challenges in high-stakes decision-making domains such as college admissions. While existing fair learning models aim to mitigate bias, achieving an optimal trade-off between fairness and accuracy remains a challenge. Moreover, the reliance on black-box models hinders interpretability, limiting their applicability in socially sensitive domains. To circumvent these issues, we propose integrating Kolmogorov-Arnold Networks (KANs) within a fair adversarial learning framework. Leveraging the adversarial robustness and interpretability of KANs, our approach facilitates stable adversarial learning. We derive theoretical insights into the spline-based KAN architecture that ensure stability during adversarial optimization. Additionally, an adaptive fairness penalty update mechanism is proposed to strike a balance between fairness and accuracy. We back these findings with empirical evidence on two real-world admissions datasets, demonstrating the proposed framework's efficiency in achieving fairness across sensitive attributes while preserving predictive performance.

Large language models (LLMs) exhibit diverse response behaviors, costs, and strengths, making it challenging to select the most suitable LLM for a given user query. We study the problem of adaptive multi-LLM selection in an online setting, where the learner interacts with users through multi-step query refinement and must choose LLMs sequentially without access to offline datasets or model internals. A key challenge arises from unstructured context evolution: the prompt dynamically changes in response to previous model outputs via a black-box process, which cannot be simulated, modeled, or learned. To address this, we propose the first contextual bandit framework for sequential LLM selection under unstructured prompt dynamics. We formalize a notion of myopic regret and develop a LinUCB-based algorithm that provably achieves sublinear regret without relying on future context prediction. We further introduce budget-aware and positionally-aware (favoring early-stage satisfaction) extensions to accommodate variable query costs and user preferences for early high-quality responses. Our algorithms are theoretically grounded and require no offline fine-tuning or dataset-specific training. Experiments on diverse benchmarks demonstrate that our methods outperform existing LLM routing strategies in both accuracy and cost-efficiency, validating the power of contextual bandits for real-time, adaptive LLM selection.

Current graph neural network (GNN) model-stealing methods rely heavily on queries to the victim model, assuming no hard query limits. However, in reality, the number of allowed queries can be severely limited. In this paper, we demonstrate how an adversary can extract a GNN with very limited interactions with the model. Our approach first enables the adversary to obtain the model backbone without making direct queries to the victim model and then to strategically utilize a fixed query limit to extract the most informative data. The experiments on eight real-world datasets demonstrate the effectiveness of the attack, even under a very restricted query limit and under defense against model extraction in place. Our findings underscore the need for robust defenses against GNN model extraction threats.

Growing concerns over data privacy underscore the need for deep learning methods capable of processing sensitive information without compromising confidentiality. Among privacy-enhancing technologies, Homomorphic Encryption (HE) stands out by offering post-quantum cryptographic security and end-to-end data protection, safeguarding data even during computation. Prior research on encrypted training has primarily focused on logistic regression, model fine-tuning, or relied on multi-party computation. This is largely due to the substantial computational overhead and algorithmic complexity involved in training deep Neural Networks (NNs) under HE. In this paper, we present ReBoot, the first framework to enable fully encrypted and non-interactive training of Multi-Layer Perceptrons (MLPs) using CKKS bootstrapping. ReBoot introduces a novel HE-compliant NN architecture based on local error signals, specifically designed to minimize multiplicative depth and reduce noise accumulation during training. It employs a tailored packing strategy that leverages real-number arithmetic through CKKS SIMD operations, significantly lowering both computational and memory overhead. We evaluate ReBoot on both image and tabular benchmarks, demonstrating up to +6.83% improvement in test accuracy over existing solutions, while reducing training latency by up to 8.83×. ReBoot is made available to the scientific community as a public repository.

The Gromov–Wasserstein (GW) distance and its fused extension (FGW) are powerful tools for comparing heterogeneous data. Their computation is, however, challenging since both distances are based on non-convex, quadratic optimal transport (OT) problems. Leveraging 1D OT, a sliced version of GW has been proposed to lower the computational burden. Unfortunately, this sliced version is restricted to Euclidean geometry and loses invariance to isometries, strongly limiting its application in practice. To overcome these issues, we propose a novel slicing technique for GW as well as for FGW that is based on an appropriate lower bound, hierarchical OT, and suitable quadrature rules for the underlying 1D OT problems. Our novel sliced FGW significantly reduces the numerical effort while remaining invariant to isometric transformations and allowing the comparison of arbitrary geometries. We show that our new distance actually defines a pseudo-metric for structured spaces that bounds FGW from below and study its interpolation properties between sliced Wasserstein and GW. Since we avoid the underlying quadratic program, our sliced distance is numerically more robust and reliable than the original GW and FGW distance; especially in the context of shape retrieval and graph isomorphism testing.

While leveraging pseudo-labels has become a common paradigm in untargeted gray-box graph poisoning attacks, it suffers from two critical limitations: the use of brittle hard pseudo-labels that overlook uncertainty and can amplify surrogate model errors, and static guidance that progressively becomes stale as the graph is perturbed. To resolve these issues, we propose MetaDist, a novel framework that reframes the attack as an adversarial self-knowledge distillation process. Here, a "teacher" model provides continuously refined soft pseudo-labels to a "student" model, with the attack objective being to maximize the divergence between them. MetaDist makes two synergistic innovations. It employs the Reverse KL (RKL) divergence as a more strategic attack loss that efficiently converts uncertain nodes into robust, high-confidence errors. Concurrently, it introduces the Online Adaptive Teacher (OAT) mechanism, which adapts the teacher via student feedback to ensure the guidance signal remains relevant. Extensive experiments demonstrate that MetaDist consistently and significantly outperforms strong baselines across multiple datasets, proving its effectiveness and transferability even against advanced graph defenses.

Cross-modal hashing (CMH) facilitates efficient retrieval across different modalities (e.g., image and text) by encoding data into compact binary representations. While recent methods have achieved remarkable performance, they often rely heavily on fully annotated datasets, which are costly and labor-intensive to obtain. In real-world scenarios, particularly in multi-label datasets, label noise is prevalent and severely degrades retrieval performance. Moreover, existing CMH approaches typically overlook the partial semantic overlaps inherent in multi-label data, limiting their robustness and generalization. To tackle these challenges, we propose a novel framework named Semantic-Consistent Bidirectional Contrastive Hashing (SCBCH). The framework comprises two complementary modules: (1) Cross-modal Semantic-Consistent Classification (CSCC), which leverages cross-modal semantic consistency to estimate sample reliability and reduce the impact of noisy labels; (2) Bidirectional Soft Contrastive Hashing (BSCH), which dynamically generates soft contrastive sample pairs based on multi-label semantic overlap, enabling adaptive contrastive learning between semantically similar and dissimilar samples across modalities. Extensive experiments on four widely-used cross-modal retrieval benchmarks validate the effectiveness and robustness of our method, consistently outperforming state-of-the-art approaches under noisy multi-label conditions.

Automated evaluation of generative text-to-image models remains a challenging problem. Recent works have proposed using multimodal LLMs to judge the quality of images, but these works offer little insight into how multimodal LLMs make use of concepts relevant to humans, such as image style or composition, to generate their overall assessment. In this work, we study what attributes of an image--specifically aesthetics, lack of artifacts, anatomical accuracy, compositional correctness, object adherence, and style--are important for both LLMs and humans to make judgments on image quality. We first curate a dataset of human preferences using synthetically generated image pairs. We use inter-task correlation between each pair of image quality attributes to understand which attributes are related in making human judgments. Repeating the same analysis with LLMs, we find that the relationships between image quality attributes are much weaker. Finally, we study individual image quality attributes by generating synthetic datasets with a high degree of control for each axis. Humans are able to easily judge the quality of an image with respect to all of the specific image quality attributes (e.g. high vs. low aesthetic image), however we find that some attributes, such as anatomical accuracy, are much more difficult for multimodal LLMs to learn to judge. Taken together, these findings reveal interesting differences between how humans and multimodal LLMs perceive images.

Finding the optimal configuration of Sparse Mixture-of- Experts (SMoE) that maximizes semantic differentiation among experts is essential for exploiting the full potential of MoE architectures. However, existing SMoE frameworks either heavily rely on hyperparameter tuning or overlook the importance of diversifying semantic roles across experts when adapting the expert pool size. We propose Mixture-of-Experts for Adaptive Semantic Specialization (MASS), a semantic-aware MoE framework for adaptive expert expansion and dynamic routing. MASS introduces two key advancements: (i) a gradient-based semantic drift detector that prompts targeted expert expansion when the existing expert pool lacks capacity to capture the full semantic diversity of the data, and (ii) an integration of adaptive routing strategy that dynamically adjusts expert usage based on token-level routing confidence mass. We first demonstrate that MASS reliably converges to the point of optimal balance between cost-performance trade-off with notably improved sematic specialization in a highly controlled synthetic setup. Further empirical results on real-world datasets across language and vision domains show that MASS consistently outperforms a range of strong MoE baselines, demonstrating its domain robustness and enhanced expert specialization.

While vision-language foundation models (VLMs) achieve remarkable performance when fine-tuned on downstream in-distribution (ID) data, this process compromises their generalization ability on out-of-distribution (OOD) data that deviate from the downstream tasks due to overfitting. To address this, we propose ProLoG, a new adaptation method that effectively fine-tunes VLMs on downstream tasks while achieving high OOD performance. Specifically, we design a unique integration of prompt tuning and LoRA, offering a robust hybrid platform to improve performance. During training, we propose an augmentation-based regularization loss that enhances the generalization of our hybrid network by using augmented image features aligned with LLM-generated texts containing key attributes of each class. By leveraging our hybrid design, we also introduce an adaptive inference strategy that flexibly applies trained prompts and LoRA based on a task similarity score to effectively handle both ID and OOD data. Experimental results demonstrate that our proposed method outperforms existing works on various datasets, confirming its advantages.

The goal of fair clustering is to find clusters such that the proportion of sensitive attributes (e.g., gender, race, etc) in each cluster is similar to the proportion of the entire data. Various fair clustering algorithms have been proposed, which modify standard K-means clustering to satisfy a given fairness constraint. A critical limitation of several existing fair clustering algorithms is that the number of parameters to be learned is proportional to the sample size because the cluster assignment of each datum should be optimized simultaneously with the cluster center, and thus scaling up the algorithms is difficult. In this paper, we propose a new fair clustering algorithm based on finite mixture model called Fair Model-based Clustering (FMC). A main advantage of FMC is that the number of learnable parameters is independent to the sample size and thus can be scaled up easily. In particular, a mini-batch learning is possible to obtain clusters that are approximately fair. Moreover, FMC can be applied to non-metric data (e.g., categorical data) as long as the likelihood is well-defined. Theoretical and empirical justifications of the superiority of the proposed algorithm are provided.

Vision-Language Models (VLMs) have been widely used in various visual recognition tasks due to their remarkable generalization capabilities. As these models grow in size and complexity, fine-tuning becomes costly, emphasizing the need to reuse adaptation knowledge from 'weaker' models to efficiently enhance 'stronger' ones. However, existing adaptation transfer methods exhibit limited transferability across models due to their model-specific design and high computational demands. To tackle this, we propose Transferable Model-agnostic adapter (TransMiter), a light-weight adapter that improves vision-language models 'without backpropagation'. TransMiter captures the knowledge gap between pre-trained and fine-tuned VLMs, in an 'unsupervised' manner. Once trained, this knowledge can be seamlessly transferred across different models without the need for backpropagation. Moreover, TransMiter consists of only a few layers, inducing a negligible additional inference cost. Notably, supplementing the process with a few labeled data further yields additional performance gain, often surpassing a fine-tuned stronger model, with a marginal training cost. Experimental results and analyses demonstrate that TransMiter effectively and efficiently transfers adaptation knowledge while preserving generalization abilities across VLMs of different sizes and architectures in visual recognition tasks.

Vision large language models (VLLMs) are focusing primarily on handling complex and fine-grained visual information by incorporating advanced vision encoders and scaling up visual models. However, these approaches face high training and inference costs, as well as challenges in extracting visual details, effectively bridging across modalities. In this work, we propose a novel visual framework, MoCHA, to address these issues. Our framework integrates four vision backbones (i.e., CLIP, SigLIP, DINOv2 and ConvNeXt) to extract complementary visual features and is equipped with a sparse Mixture of Experts Connectors (MoECs) module to dynamically select experts tailored to different visual dimensions. To mitigate redundant or insufficient use of the visual information encoded by the MoECs module, we further design a Hierarchical Group Attention (HGA) with intra- and inter-group operations and an adaptive gating strategy for encoded visual features. We train MoCHA on two mainstream LLMs (e.g., Phi2-2.7B and Vicuna-7B) and evaluate their performance across various benchmarks. Notably, MoCHA outperforms state-of-the-art open-weight models on various tasks. For example, compared to CuMo (Mistral-7B), our MoCHA (Phi2-2.7B) presents outstanding abilities to mitigate hallucination by showing improvements of 3.25% in POPE and to follow visual instructions by raising 153 points on MME. Finally, ablation studies further confirm the effectiveness and robustness of the proposed MoECs and HGA in improving the overall performance of MoCHA.

Federated learning (FL) enables collaborative model training without centralizing data. In multi-domain scenarios with non-identically and independently distributed (non-IID) data, prediction performance is often hindered by catastrophic forgetting of specialized local knowledge and negative transfer from conflicting client updates. To address these challenges, we propose a personalized FL framework with dual-branch (pFedDB) structure and a two-phase training protocol. The dual-branch architecture separates the model into a shared branch for cross-client aggregation and a private branch that remains on each local client. The private branch is never overwritten by server updates, which prevents the catastrophic forgetting of domain-specific knowledge. This structure also significantly reduces communication overhead per round as only the shared branch is transmitted. To mitigate negative transfer, our two-phase protocol first establishes a personalized knowledge anchor by training a single-branch expert model on each client’s local data. In the second phase, the locally trained model is cloned to initialize private and shared branches. Only the shared branch is aggregated in federated training. This process enables the shared branch to learn a general representation that complements the established local expertise. This design consistently improves the performance of every client over its single-domain baseline, overcoming the challenge of negative transfer among clients. Experiments on our new Chest-X-Ray-4 suite and three public benchmarks show that the proposed pFedDB method obtains 30% saving in communication overhead per round and competitive or better accuracy performance than recent FL methods.

Reproducibility is a cornerstone of scientific validation and of the authority it confers on its results. Reproducibility in machine learning evaluations leads to greater trust, confidence, and value. However, the ground truth responses used in machine learning often necessarily come from humans, among whom disagreement is prevalent, and surprisingly little research has studied the impact of effectively ignoring disagreement in these responses, as is typically the case. One reason for the lack of research is that budgets for collecting human-annotated evaluation data are limited, and obtaining more samples from multiple raters for each example greatly increases the per-item annotation costs. We investigate the trade-off between the number of items (N) and the number of responses per item (K) needed for reliable machine learning evaluation. We analyze a diverse collection of categorical datasets for which multiple annotations per item exist, and simulated distributions fit to these datasets, to determine the optimal (N, K) configuration, given a fixed budget (N x K), for collecting evaluation data and reliably comparing the performance of machine learning models. Our findings show, first, that accounting for human disagreement may come with N x K at no more than 1000 (and often much lower) for every dataset tested on at least one metric. Moreover, this minimal N x K almost always occurred for K > 10. Furthermore, the nature of the tradeoff between K and N, or if one even existed, depends on the evaluation metric, with metrics that are more sensitive to the full distribution of responses performing better at higher levels of K. Our methods can be used to help ML practitioners get more effective test data by finding the optimal metrics and number of items and annotations per item to collect to get the most reliability for their budget.

The use of learned dynamics models, also known as world models, can improve the sample efficiency of reinforcement learning. Recent work suggests that the underlying causal graphs of such dynamics models are sparsely connected, with each of the future state variables depending only on a small subset of the current state variables, and that learning may therefore benefit from sparsity priors. Similarly, temporal sparsity, i.e. sparsely and abruptly changing local dynamics, has also been proposed as a useful inductive bias. In this work, we critically examine these assumptions by analyzing ground truth dynamics from a set of robotic reinforcement learning environments in the MuJoCo Playground benchmark suite, aiming to determine whether the proposed notions of state and temporal sparsity actually tend to hold in typical reinforcement learning tasks. We study (i) whether the causal graphs of environment dynamics are sparse, (ii) whether such sparsity is state-dependent, and (iii) whether local system dynamics change sparsely. Our results indicate that global sparsity is rare, but instead the tasks show local, state-dependent sparsity in their dynamics and this sparsity exhibits distinct structures, appearing in temporally localized clusters (e.g., during contact events) and affecting specific subsets of state dimensions. These findings challenge common sparsity prior assumptions in dynamics learning, emphasizing the need for grounded inductive biases that reflect the state-dependent sparsity structure of real-world dynamics.