AAAI 2026 Papers

Image clustering is a classic problem in computer vision, which categorizes images into different groups. Recent studies utilize nouns as external semantic knowledge to improve clustering performance. However, these methods often overlook the inherent ambiguity of nouns, which can distort semantic representations and degrade clustering quality. To address this issue, we propose a hierarChical semAntic alignmEnt method for image clustering, dubbed CAE, which improves clustering performance in a training-free manner. In our approach, we incorporate two complementary types of textual semantics: caption-level descriptions, which convey fine-grained attributes of image content, and noun-level concepts, which represent high-level object categories. We first select relevant nouns from WordNet and descriptions from caption datasets to construct a semantic space aligned with image features. Then, we design a residual attention mechanism to further enhance the discriminability of this space. Finally, we combine the enhanced semantic and image features to perform clustering. Extensive experiments across 8 datasets demonstrate the effectiveness of our method, notably surpassing the state-of-the-art training-free approach with a 4.2% improvement in accuracy and a 2.9% improvement in adjusted rand index (ARI) on the ImageNet-1K dataset.

Developing high-performance GPU kernels is critical for AI and scientific computing, but remains challenging due to its reliance on expert crafting and poor portability. While large language models (LLMs) offer promise for automation, both general-purpose and finetuned LLMs suffer from two fundamental and conflicting limitations: correctness and efficiency. The key reason is that existing LLM-based approaches directly generate the entire optimized low-level programs, requiring exploration of an extremely vast space encompassing both optimization policies and implementation codes. To address the challenge of exploring an intractable space, we propose Macro Thinking Micro Coding (MTMC), a hierarchical framework inspired by the staged optimization strategy of human experts. It decouples optimization strategy from implementation details, ensuring efficiency through high-level strategy and correctness through low-level implementation. Specifically, Macro Thinking employs reinforcement learning to guide lightweight LLMs in efficiently exploring and learning semantic optimization strategies that maximize hardware utilization. Micro Coding leverages general-purpose LLMs to incrementally implement the stepwise optimization proposals from Macro Thinking, avoiding full-kernel generation errors. Together, they effectively navigate the vast optimization space and intricate implementation details, enabling LLMs for high-performance GPU kernel generation. Comprehensive results on widely adopted benchmarks demonstrate the superior performance of MTMC on GPU kernel generation in both accuracy and running time. On KernelBench, MTMC achieves near 100% and 70% accuracy at Levels 1-2 and 3, over 50% than SOTA general-purpose and domain-finetuned LLMs, with up to 7.3× speedup over LLMs, and 2.2× over expert-optimized PyTorch Eager kernels. On the more challenging TritonBench, MTMC attains up to 59.64% accuracy and 34× speedup. All models and datasets will be made publicly available.

Search and recommendation are pivotal for information access and are increasingly unified to exploit shared user-item interactions. Both tasks suffer from data sparsity, which joint modeling can mitigate by integrating behavioral data with or without explicit queries. However, existing unified frameworks rarely distinguish between users’ long- and short-term interests, despite their divergent temporal dynamics in search and recommendation. In this work, we propose a novel model, DHIM, which explicitly disentangles and integrates users' long- and short-term interests across both the search and recommendation scenarios. First, long- and short-term interests are independently extracted from search and recommendation using a unified extraction strategy. These interests are then adaptively integrated via a cross-scenario fusion module. A self‐supervised contrastive loss supervises the learning of both interest types within and across scenarios. The resulting representations are fed into downstream search and recommendation models for prediction. Extensive experiments on two public benchmarks demonstrate that our approach consistently outperforms single-scenario and state-of-the-art joint models, achieving superior accuracy and generalizability. To our knowledge, this is the first work to incorporate explicit dual-horizon interest modeling into a unified search and recommendation framework with self-supervised contrastive learning.

Multimodal Large Language Models (MLLMs) have played an increasingly important role in multimodal intelligence. However, the existing fine-tuning methods often ignore cross-modal heterogeneity, limiting their full potential. In this work, we propose a novel fine-tuning strategy by injecting beneficial random noise, which outperforms previous methods and even surpasses full fine-tuning, with minimal additional parameters. The proposed Multimodal Noise Generator (MuNG) enables efficient modality fine-tuning by injecting customized noise into the frozen MLLMs. Specifically, we reformulate the reasoning process of MLLMs from a variational inference perspective, upon which we design a multimodal noise generator that dynamically analyzes cross-modal relationships in image-text pairs to generate task adaptive beneficial noise. Injecting this type of noise into the MLLMs effectively suppresses irrelevant semantic components, leading to significantly improved cross-modal representation alignment and enhanced performance on downstream tasks. Experiments on two mainstream MLLMs, QwenVL and LLaVA, demonstrate that our method surpasses full parameter fine-tuning and other existing fine-tuning approaches, while requiring adjustments to only about 1~2% additional parameters.

We propose a physics-informed learning framework, called Koopman-PINN, to estimate the parameters of the Heston stochastic volatility model with high-frequency price data in financial markets. The method integrates a nonparametric volatility estimation (known as ART-filter in the literature), moment-based parameter initialization, and a neural Koopman operator constrained by the infinitesimal generator of the underlying stochastic differential equation. By incorporating a generator-based loss, the model bridges Koopman theory and neural modeling to handle partially observed coupled stochastic dynamics in a manner consistent with continuous-time evolution. Across diverse parameter combinations reflecting varying market conditions, Koopman-PINN consistently achieves accurate and robust five-parameter recovery, outperforming existing estimators under a minimal set of initialization assumptions.

Prompting is fundamental to unlocking the full potential of large language models. To automate and enhance this process, automatic prompt optimization (APO) has been developed, demonstrating effectiveness primarily in text-only input scenarios. However, extending existing APO methods to multimodal tasks—such as video-language generation—introduces two core challenges: (i) visual token inflation, where long visual-token sequences restrict context capacity and result in insufficient feedback signals; (ii) a lack of process-level supervision, as existing methods focus on outcome-level supervision and overlook intermediate supervision, limiting prompt optimization. We present UniAPO: Unified Multimodal Automated Prompt Optimization, the first framework tailored for multimodal APO. UniAPO adopts an EM-inspired optimization process that decouples feedback modeling and prompt refinement, making the optimization more stable and goal-driven. To further address the aforementioned challenges, we introduce a short-long term memory mechanism: historical feedback mitigates context limitations, while historical prompts provide directional guidance for effective prompt optimization. UniAPO achieves consistent gains across text, image, and video benchmarks, establishing a unified framework for efficient and transferable prompt optimization.

Schrödinger Bridge-based diffusion models have demonstrated promising performance in signal denoising. However, since ground truth signals are unavailable during the sampling process, neural networks must be employed to learn the mapping, which breaks the theoretical coupling between diffusion and sampling processes. This paper reveals a critical inconsistency between the theoretical diffusion path and the learned sampling trajectory across different frequency bands. This diffusion-sampling inconsistency directly undermines denoising effectiveness. To address this limitation, we propose the Frequency-Dependent Scheduled Schrödinger Bridge (FDSSB), which leverages power spectral density to adaptively schedule diffusion processes across frequencies. This mechanism assigns asynchronous diffusion schedules to different frequency components, correcting the diffusion schedule to better match the sampling process. As a result, FDSSB effectively mitigates the mismatch and enhances the consistency between diffusion and sampling processes. Extensive experiments demonstrate that FDSSB achieves state-of-the-art performance, with an average scale-invariant signal-to-noise ratio improvement of 7.9066 dB over competitive approaches.

Irregular multivariate time series (IMTS), characterized by uneven sampling and inter-variate asynchrony, fuel many forecasting applications yet remain challenging to model efficiently. Canonical Pre-Alignment (CPA) has been widely adopted in IMTS modeling by padding zeros at every global timestamp, thereby alleviating inter-variate asynchrony and unifying the series length, but its dense zero-padding inflates the pre-aligned series length, especially when numerous variates are present, causing prohibitive compute overhead. Recent graph-based models with patching strategies sidestep CPA, but their local message passing struggles to capture global inter-variate correlations. Therefore, we posit that CPA should be retained, with the pre-aligned series properly handled by the model, enabling it to outperform state-of-the-art graph-based baselines that sidestep CPA. Technically, we propose KAFNet, a compact architecture grounded in CPA for IMTS forecasting that couples (1) a Pre-Convolution module for sequence smoothing and sparsity mitigation, (2) a Temporal Kernel Aggregation module for learnable compression and modeling of intra-series irregularity, and (3) Frequency Linear Attention blocks for low-cost inter-series correlation modeling in the frequency domain. Experiments on multiple IMTS datasets show that KAFNet achieves state-of-the-art forecasting performance, with a 7.2× parameter reduction and an 8.4× training–inference acceleration.

Bandit multiple hypothesis testing has broad applications in biological sciences, clinical testing for drug discovery, and online A/B/n testing. The framework utilizes an adaptive sampling strategy for multiple testing which aims to maximize statistical power while ensuring anytime false discovery rate control. This paper proposes a robust approach for bandit multiple testing, allowing for at most an epsilon fraction of arbitrary distribution corruption, as in Huber’s contamination model. Specifically, we introduce two adaptive sampling strategies designed to minimize the number of samples required to exceed a target true positive rate, while providing anytime control over the false discovery rate. We analyze the sample complexity of our proposed methods and perform numerical simulations to demonstrate their efficiency and robustness. Furthermore, we extend our methods to address scenarios where distributions have infinite variance and situations involving multiple agents collaborating on the same bandit task.

The goal of distributionally robust learning is to learn models capable of performing well against distributional shifts, such as latent heterogeneous subpopulations, unknown covariate shifts, or unmodeled temporal effects. Recently, Duchi and Namkoong (2021) have proven an upper bound for the excess risk of distributionally robust learning through the lens of covering number argument. However, there are situations where the covering argument fails. This motivates us to study the generalization bound through the lens of Rademacher complexity. More specifically, we consider the Cressie-Read divergence, f sub k of t is proportional to t to the k minus one. Our theoretical results indicate that the excess risk is of the order big O sub P of n to the negative one over two k star, where k star equals k over k minus one. The decay rate of the excess risk increases with increasing k. As illustrative examples, we consider three learning settings: 1) linear classifier; 2) Gaussian reproducing kernel Hilbert space; 3) one-hidden-layer networks. The empirical results validate our theoretical findings.

Homeostatic mechanisms play a crucial role in maintaining optimal functionality within the neural circuits of the brain. By regulating physiological and biochemical processes, these mechanisms ensure the stability of an organism’s internal environment, enabling it to better adapt to external changes. Among these mechanisms, the Bienenstock, Cooper, and Munro (BCM) theory has been extensively studied as a key principle for maintaining the balance of synaptic strengths in biological systems. Despite the extensive development of spiking neural networks (SNNs) as a model for bionic neural networks, no prior work in the machine learning community has integrated biologically plausible BCM formulations into SNNs to provide homeostasis. In this study, we propose a Dynamic Weight Adaptation Mechanism (DWAM) for SNNs, inspired by the BCM theory. DWAM can be integrated into the host SNN, dynamically adjusting network weights in real time to regulate neuronal activity, providing homeostasis to the host SNN without any fine-tuning. We validated our method through dynamic obstacle avoidance and continuous control tasks under both normal and specifically designed degraded conditions. Experimental results demonstrate that DWAM not only enhances the performance of SNNs without existing homeostatic mechanisms under various degraded conditions but also further improves the performance of SNNs that already incorporate homeostatic mechanisms.

A fundamental challenge in visual reinforcement learning (RL) is achieving robust generalization across environments with varying visual distractions. Current RL methods struggle with generalization due to their inability to differentiate foreground and background features during augmentation,while their Q-consistency mechanisms rely on outdated actions from replay buffers that drift from the current policy.In this paper, we present PQDA, a novel framework that addresses generalization challenges in RL through two key innovations: (1) Foreground-Background Decoupled Augmentation leverages Gaussian mixture model-based segmentation to efficiently generate and cache masks in replay buffers, applying differentiated augmentation strategies to foreground and background regions, thereby enhancing data diversity while maintaining task-relevant features. (2) Policy-Aligned Q-Consistency enforces policy alignment by sampling actions from the current policy for Q-regularization, achieving faster and more stable convergence. Notably, PQDA eliminates auxiliary tasks entirely through a unified architecture that co-optimizes the encoder and RL components directly. Extensive experiments on DMControl benchmarks (including our newly proposed CVDMC benchmark) and robotic manipulation tasks demonstrate PQDA's superior generalization performance, outperforming state-of-the-art methods.

The rapid advancement of Large Vision Language Models (LVLMs) has demonstrated excellent abilities in various visual tasks. Building upon these developments, the thinking with images paradigm has emerged, enabling models to dynamically edit and re-encode visual information at each reasoning step, mirroring human visual processing. However, this paradigm introduces significant challenges as diverse errors may occur during reasoning processes. This necessitates Process Reward Models (PRMs) for distinguishing positive and negative reasoning steps, yet existing benchmarks for PRMs are predominantly text-centric and lack comprehensive assessment under this paradigm. To address these gaps, this work introduces the first comprehensive benchmark specifically designed for evaluating PRMs under the thinking with images paradigm. Our main contributions are: (1) Through extensive analysis of reasoning trajectories and guided search experiments with PRMs, we define 7 fine-grained error types and demonstrate both the necessity for specialized PRMs and the potential for improvement. (2) We construct a comprehensive benchmark comprising 1,206 manually annotated thinking with images reasoning trajectories spanning 4 categories and 16 subcategories for fine-grained evaluation of PRMs. (3) Our experimental analysis reveals that current LVLMs fall short as effective PRMs, exhibiting limited capabilities in visual reasoning process evaluation with significant performance disparities across error types, positive evaluation bias, and sensitivity to reasoning step positions. These findings demonstrate the effectiveness of our benchmark and establish crucial foundations for advancing PRMs in LVLMs.

Vision-Language Models (VLMs), exemplified by CLIP, have emerged as foundational for multimodal intelligence. However, their capacity for logical understanding remains significantly underexplored, resulting in critical **''logical blindspots''** that limit their reliability in practical applications. To systematically diagnose this, we introduce **LogicBench**, a comprehensive benchmark with over 50,000 vision-language pairs across 9 logical categories and 4 diverse scenarios: images, videos, anomaly detection, and medical diagnostics. Our evaluation reveals that existing VLMs, even the state-of-the-art ones, fall at over 40 accuracy points below human performance, particularly in challenging tasks like Causality and Conditionality, highlighting their reliance on surface semantics over critical logical structures. To bridge this gap, we propose **LogicCLIP**, a novel training framework designed to boost VLMs' logical sensitivity through advancements in both data generation and optimization objectives. LogicCLIP utilizes logic-aware data generation and a contrastive learning strategy that combines coarse-grained alignment, a fine-grained multiple-choice objective, and a novel logical structure-aware objective. Extensive experiments demonstrate LogicCLIP's substantial improvements in logical comprehension across all LogicBench domains, significantly outperforming baselines. Moreover, LogicCLIP retains, and often surpasses, competitive performance on general vision-language benchmarks, demonstrating that the enhanced logical understanding does not come at the expense of general alignment. We believe LogicBench and LogicCLIP will be important resources for advancing VLM logical capabilities.

Sharpness-aware minimization (SAM) is widely recognized for enhancing the generalization performance of deep neural networks. However, recent works have challenged the statement that flatness implies generalization, demonstrating that it is insufficient as the indicator of generalization. In this paper, we reveal an insightful phenomenon: among minima of similar sharpness, stochastic optimization algorithms tend to prefer those with lower nonuniformity. We define nonuniformity by both the magnitude and structure of the gradient noise, and show that it fundamentally differs from sharpness and plays a critical role in generalization. Specifically, we first theoretically prove that the expected generalization gap of models trained via stochastic optimization algorithm is positively correlated with nonuniformity (the magnitude of the gradient noise). Empirically, we show that nonuniformity exhibits a stronger correlation with generalization than sharpness, especially in Transformer models. Furthermore, we demonstrate that the nonuniformity (the structure of the gradient noise) more effectively guides the algorithm towards sparser solutions and exhibits better generalization performance than sharpness-based methods in the high-dimensional sparse regression problem. Finally, extensive experiments on various datasets and models confirm the advantages of nonuniformity for generalization: (1) optimization guided by nonuniformity achieves better generalization compared to those achieved through flatness (including standard training, transfer learning, hyperparameter sensitivity and robustness to label noise); (2) model architecture (such as depth and width) is closely related to nonuniformity.

Fault Diagnosis (FD) on sequential data suffers from irregular sampling (with missing values), limited training data, and varying underlying environments. In response, this paper proposes FD by adjoint learning in continuous-time model space. Model-Space Learning employs well-fitted models that capture data's dynamics (i.e., changing information) as more stable and concise representations of the original data. The Continuous-Time Reservoir Computing Network (CT-Res) is first introduced, which embeds Ordinary Differential Equation (ODE) within the reservoir-based hidden layer to govern continuous-time hidden-state evolution, naturally handling irregular sampling without relying on fixed time steps and effectively capturing intrinsic data dynamics. By fitting each sequence via CT-Res and representing it with the fitted model, the original sequences are mapped from the data space into the continuous-time model space. We further develop an adjoint learning strategy by incorporating a discrete-time "adjoint Echo State Network (ESN)" that shares structure and parameters with CT-Res, thus enabling efficient training by bypassing the computationally intensive ODE solver, with joint optimization of fitting accuracy and class discrimination in the model space. Experiments on multiple FD benchmarks highlight the effectiveness and efficiency of our study, particularly with missing values and scarce training data.

This work introduces Text-based Aerial-Ground Person Retrieval (TAG-PR), which aims to retrieve person images from heterogeneous aerial and ground views with textual descriptions. Unlike traditional Text-based Person Retrieval (T-PR), which focuses solely on ground-view images, TAG-PR introduces greater practical significance and presents unique challenges due to the large viewpoint discrepancy across images. To support this task, we contribute: (1) TAG-PEDES dataset, constructed from public benchmarks with automatically generated textual descriptions, enhanced by a diversified text generation paradigm to ensure robustness under view heterogeneity; and (2) TAG-CLIP, a novel retrieval framework that addresses view heterogeneity through a hierarchically-routed mixture of experts module to learn view-specific and view-agnostic features and a viewpoint decoupling strategy to decouple view-specific features for better cross-modal alignment. We evaluate the effectiveness of TAG-CLIP on both the proposed TAG-PEDES and existing T-PR benchmarks.

The success of modern machine learning hinges on access to high-quality training data. In many real-world scenarios, such as acquiring data from public repositories or sharing across institutions, data is naturally organized into discrete datasets that vary in relevance, quality, and utility. Selecting which repositories or institutions to search for useful datasets, and which datasets to incorporate into model training are therefore critical decisions, yet most existing methods select individual samples and treat all data as equally relevant, ignoring differences between datasets and their sources. In this work, we formalize the task of dataset selection: selecting entire datasets from a large, heterogeneous pool to improve downstream performance under resource constraints. We propose Dataset Selection via Hierarchies (DaSH), a dataset selection method that models utility at both dataset and group (e.g., collections, institutions) levels, enabling efficient generalization from limited observations. Across two public benchmarks (Digit-Five and DomainNet), DaSH outperforms state-of-the-art data selection baselines by up to 26.2% in accuracy, while requiring significantly fewer exploration steps. Ablations show DaSH is robust to low-resource settings and lack of relevant datasets, making it suitable for scalable and adaptive dataset selection in practical multi-source learning workflows.

Marked Temporal Point Processes (MTPPs) provide a principled framework for modeling asynchronous event sequences by conditioning on the history of past events. However, most existing MTPP models rely on channel-mixing strategies that encode information from different event types into a single, fixed-size latent representation. This entanglement can obscure type-specific dynamics, leading to performance degradation and increased risk of overfitting. In this work, we introduce ITPP, a novel channel-independent architecture for MTPP modeling that decouples event type information using an encoder-decoder framework with an ODE-based backbone. Central to ITPP is a type-aware inverted self-attention mechanism, designed to explicitly model inter-channel correlations among heterogeneous event types. This architecture enhances effectiveness and robustness while reducing overfitting. Comprehensive experiments on multiple real-world and synthetic datasets demonstrate that ITPP consistently outperforms state-of-the-art MTPP models in both predictive accuracy and generalization.