AAAI 2026 论文列表

Deploying Vision-Language Models (VLMs) on edge devices (e.g., smartphones and robots) is crucial for enabling low-latency and privacy-preserving intelligent applications. Given the resource constraints of these devices, quantization offers a promising solution by improving memory efficiency and reducing bandwidth requirements, thereby facilitating the deployment of VLMs. However, existing research has rarely explored aggressive quantization on VLMs, particularly for the models ranging from 1B to 2B parameters, which are more suitable for resource-constrained edge devices. In this paper, we propose SPEED-Q, a novel Staged Processing with EnhancEd Distillation framework for VLM low-bit weight-only quantization that systematically addresses the following two critical obstacles: (1) significant discrepancies in quantization sensitivity between vision (ViT) and language (LLM) components in VLMs; (2) training instability arising from the reduced numerical precision inherent in low-bit quantization. In SPEED-Q, a staged sensitivity adaptive mechanism is introduced to effectively harmonize performance across different modalities. We further propose a distillation-enhanced quantization strategy to stabilize the training process and reduce data dependence. Together, SPEED-Q enables accurate, stable, and data-efficient quantization of complex VLMs. SPEED-Q is the first framework tailored for quantizing entire small-scale billion-parameter VLMs to low bits. Extensive experiments across multiple benchmarks demonstrate that SPEED-Q achieves up to 6x higher accuracy than existing quantization methods under 2-bit settings and consistently outperforms prior on-device VLMs under both 2-bit and 4-bit settings.

In semi‑supervised multi‑view classification (SMVC), scarce labels and noisy unlabeled data impair feature aggregation and compromise prediction reliability, while existing methods lack principled guidance and interpretability. To overcome these limitations, we propose a novel unified SMVC framework, Neural Collapse Priors Driven Trust Semi-Supervised Multi-View Classification (NCPD-TSMVC), building upon neural collapse–derived prototype priors and evidential opinion fusion. Concretely, we rigorously prove under neural collapse theory that normalized classifier weights from the labeled‑data pre‑training stage coincide with class centroids in feature space, conferring maximal inter‑class separation and optimal within‑class compactness. These prototype priors permeate the entire learning pipeline, calibrating the representation learning of unlabeled samples to obtain highly discriminative embeddings. Simultaneously, our evidential learning module quantifies epistemic uncertainty and fuses view‑level opinions at the evidence level, yielding robust and transparent decision making. Extensive evaluations across diverse benchmarks demonstrate that NCPD‑TSMVC surpasses state‑of‑the‑art SMVC approaches in performance, robustness and interpretability.

We introduce Conformal Interquantile Regression (CIR), a conformal regression method that efficiently constructs near-minimal prediction intervals with guaranteed coverage. CIR leverages black-box machine learning models to estimate outcome distributions through interquantile ranges, transforming these estimates into compact prediction intervals while achieving approximate conditional coverage. We further propose CIR+ (Conditional Interquantile Regression with More Comparison), which enhances CIR by incorporating a width-based selection rule for interquantile intervals. This refinement yields narrower prediction intervals while maintaining comparable coverage, though at the cost of slightly increased computational time. Both methods address key limitations of existing distributional conformal prediction approaches: they handle skewed distributions more effectively than Conformalized Quantile Regression, and they achieve substantially higher computational efficiency than Conformal Histogram Regression by eliminating the need for histogram construction. Extensive experiments on synthetic and real-world datasets demonstrate that our methods optimally balance predictive accuracy and computational efficiency compared to existing approaches.

Recently, multi-modal embedding methods have flourished in entity alignment. As state-of-the-art approaches evolve rapidly, visual modality (i.e., images) missing emerges as a critical challenge. While visual modality typically offers the most informative signals in multi-modal entity alignment (MMEA), it is frequently unavailable for many entities. The existing methods commonly use dummy vectors to represent visual-missing embeddings, which negatively impacts both model training and inference. In this paper, we propose robust multi-modal entity alignment (rMMEA), which leverages ranking-based knowledge distillation and mutual information (MI) estimation to address missing modalities while enhancing noise robustness. Unlike conventional teacher-student distillation that requires the student to replicate teacher outputs, our rMMEA learns soft rankings from pure and complete modality sides while capturing implicit key semantics of teacher embeddings through mutual information maximization, allowing rMMEA to avoid strict point-to-point alignment. The experimental results across multiple benchmarks and settings demonstrate that rMMEA significantly outperforms the state-of-the-art anti-modality-missing methods in terms of effectiveness and efficiency.

Diffusion planning is a promising method for learning high-performance policies from offline data. To avoid the impact of discrepancies between planning and reality on performance, previous works generate new plans at each time step. However, this incurs significant computational overhead and leads to lower decision frequencies, and frequent plan switching may also affect performance. In contrast, humans might create detailed short-term plans and more general, sometimes vague, long-term plans, and adjust them over time. Inspired by this, we propose the Temporal Diffusion Planner (TDP) which improves decision efficiency by distributing the denoising steps across the time dimension. TDP begins by generating an initial plan that becomes progressively more vague over time. At each subsequent time step, rather than generating an entirely new plan, TDP updates the previous one with a small number of denoising steps. This reduces the average number of denoising steps, improving decision efficiency. Additionally, we introduce an automated replanning mechanism to prevent significant deviations between the plan and reality. Experiments on D4RL show that, compared to previous works that generate new plans every time step, TDP significantly improves the decision-making frequency by 11-24.8 times while achieving higher or comparable performance.

In this paper, we explore the transferability of SSL by addressing two central questions: (i) what is the representation transferability of SSL, and (ii) how can we effectively model this transferability? Transferability is defined as the ability of a representation learned from one task to support the objective of another. Inspired by the meta-learning paradigm, we construct multiple SSL tasks within each training batch to support explicitly modeling transferability. Based on empirical evidence and causal analysis, we find that although introducing task-level information improves transferability, it is still hindered by task conflict. To address this issue, we propose a Task Conflict Calibration method to alleviate the impact of task conflict. Specifically, it first splits batches to create multiple SSL tasks, infusing task-level information. Next, it uses a factor extraction network to produce causal generative factors for all tasks and a weight extraction network to assign dedicated weights to each sample, employing data reconstruction, orthogonality, and sparsity to ensure effectiveness. Finally, the method calibrates sample representations during SSL training and integrates into the pipeline via a two-stage bi-level optimization framework to boost the transferability of learned representations. Experimental results on multiple downstream tasks demonstrate that our method consistently improves the transferability of SSL models.

Federated learning (FL) protects data privacy by enabling distributed model training without direct access to client data. However, its distributed nature makes it vulnerable to model and data poisoning attacks. While numerous defenses filter malicious clients using statistical metrics, they overlook the role of model redundancy, where not all parameters contribute equally to the model and attack performance. Current attacks manipulate all model parameters uniformly, making them more detectable, while defenses focus on the overall statistics of client updates, leaving gaps for more sophisticated attacks. We propose an attack-agnostic augmentation method to enhance the stealthiness and effectiveness of existing poisoning attacks in FL, exposing flaws in current defenses and highlighting the need for fine-grained FL security. Our three-stage methodology, including pill construction, pill poisoning, and pill injection, injects poison into a compact subnet (i.e., pill) of the global model during the iterative FL training. Experimental results show that FL poisoning attacks enhanced by our method can bypass 8 state-of-the-art (SOTA) defenses, gaining an up to 7x error rate increase, as well as on average a more than 2x error rate increase on both IID and non-IID data, in both cross-silo and cross-device FL systems.

With recent advancements in next-generation data storage, especially in biological molecule-based storage, insertion, deletion, and substitution (IDS) error-correcting codes have garnered increased attention. However, a universal method for designing tailored IDS-correcting codes across varying channel settings remains underexplored. We present an autoencoder-based approach, THEA-code, aimed at efficiently generating IDS-correcting codes for complex IDS channels. In the work, a disturbance-based discretization is proposed to discretize the features of the autoencoder, and a simulated differentiable IDS channel is developed as a differentiable alternative for IDS operations. These innovations facilitate the successful convergence of the autoencoder, producing channel-customized IDS-correcting codes that demonstrate commendable performance across complex IDS channels, particularly in realistic DNA-based storage channels.

As Convolutional Neural Networks (CNNs) continue to gain traction in deep learning, Winograd convolution has emerged as a key algorithm to enhance computational efficiency. Although ARM-based CPUs are increasingly prevalent in mobile devices, embedded systems and HPC servers, existing 2D Winograd convolution implementations for ARM often leave room for improvement in transformation efficiency, computational throughput, and overall versatility. Furthermore, the lack of tailored 3D Winograd convolution implementations for ARM architectures stems from the additional complexity of supporting higher-dimensional kernels. AirWino introduces a set of novel optimizations covering transformations, data layouts, micro-kernel computations, and parallelization strategies for both 2D and 3D Winograd convolution. It supports FP32 and FP16 precisions with filter sizes of 3 and 5, targeting a broad range of applications. Evaluations on four distinct ARM platforms show that AirWino consistently outperforms state-of-the-art libraries across various experimental scenarios and hardware configurations, highlighting its efficiency and portability.

This paper presents a novel method, called Deformable Polygonal Flow Matching (DPFM), for the generation of polygonal arrangements such as jigsaw puzzles and floor plans. DPFM is a Flow Matching framework that enables the generation process to deform, rotate, and translate polygons while decoupling these transformations, allowing to toggle them individually. Able to combine the spatial reasoning capabilities of arrangement models with the flexibility of position-based models, it covers a wide range of applications within a unified formulation, from noiseless puzzle solving using rigid alignments to unconstrained floor plan generation.We represent data using a hierarchical graph composed of a topological subgraph encoding connectivity information and semantics (such as room types for floor plans), and a geometrical subgraph encoding the 1D polygonal loop of each shape. DPFM also leverages Flow Matching's arbitrary prior distributions for geometric constraints by designing priors with domain knowledge. Rather than starting the generation process from uninformed distributions, the generation is constrained through the informed priors at the initialization stage. The qualitative and quantitative evaluations of our method, ran on the RPLAN and jigsaw puzzle datasets, demonstrate strong performance. DPFM outperforms task-specific methods, becoming the new state-of-the-art for 2D arrangement generation. Our results show that DPFM is able to solve novel tasks, such as puzzle denoising, where pieces are reconstructed from noisy versions and arranged into a valid puzzle in parallel.

Multi-view graph clustering (MVGC) for remote sensing data has gained increasing attention due to its ability to integrate complementary information across modalities while capturing spatial dependencies in heterogeneous data. Although current methods based on graph contrastive learning achieve strong performance, they often misidentify intra-cluster samples as negatives, leading to class conflicts and reduced clustering accuracy. Graph masked autoencoders have recently shown promising potential in learning robust representations through masked reconstruction, but their application to remote sensing data remains underexplored. This challenge is especially notable in the multi-view remote sensing setting, where high heterogeneity and complex spatial structures increase the difficulty of effective representation learning. To address these issues, we propose Clustering-Guided graph Mask AutoEncoder (CG-MAE), the first framework to extend graph masked autoencoders to multi-view remote sensing clustering. We introduce a clustering-guided masking strategy that selectively masks nodes near cluster centers and intra-cluster edges, which are crucial for capturing key structural information. By reconstructing these masked components, the model is encouraged to focus on learning features that are highly relevant to clustering. To further improve training stability and efficiency, we design an easy-to-hard node masking strategy that enables the model to gradually learn from increasingly challenging patterns. Additionally, we propose a dual self-adaptive learning mechanism that encourages the model to align more closely with the underlying semantic distributions. Extensive experiments on four widely used multi-view remote sensing datasets demonstrate that CG-MAE consistently outperforms state-of-the-art methods in both clustering accuracy and representation quality.

With the rapid development of multimodal large language models (MLLMs), deploying them on low-resource devices remains challenging. Beyond the model size, long multimodal inputs cause substantial memory overhead in the KV cache, making efficient cache management critical. In this paper, we propose DAVID, a KV cache eviction strategy that adapts to the degree of modality fusion across layers. By analyzing the feature distributions of vision and text tokens, we observe low fusion in early layers and high fusion in deeper layers. Based on this observation, DAVID adopts a decoupled eviction strategy in shallow layers and a super-modal eviction strategy in deeper layers. To support this dynamic switching, we design a lightweight metric that quantifies cross-modal fusion and uses a threshold to determine which layers require decoupling. Experimental results show that DAVID achieves state-of-the-art performance on multiple benchmarks and offers a new perspective on KV cache eviction for MLLMs.

Universal multimodal embedding models are essential in various tasks. Existing approaches typically use in-batch mining to identify hard negatives by measuring the similarity of query-candidate pairs. However, these methods often struggle to capture subtle semantic differences among candidates and lack diversity in negative samples. Moreover, the embeddings exhibit limited discriminative ability in distinguishing false and hard negatives. In this paper, we leverage the advanced understanding capabilities of MLLMs to enhance representation learning, and present a novel Universal Multimodal Embedding(UniME-V2) model. Our approach first constructs a potential hard negative set through global retrieval. We then introduce the MLLM-as-a-Judge mechanism, which utilizes MLLMs to assess the semantic alignment of query-candidate pairs and generate soft semantic matching scores. These scores serve as a foundation for hard negative mining, mitigating the impact of false negatives and enabling the identification of diverse, high-quality hard negatives. Furthermore, the semantic matching scores are used as soft labels to mitigate the rigid one-to-one mapping constraint. By aligning the similarity matrix with the soft semantic matching score matrix, the model learns semantic distinctions among candidates, significantly enhancing its discriminative capacity. To further improve performance, we propose UniME-V2, a reranking model trained on our mined hard negatives through a joint pairwise and listwise optimization approach. We conduct comprehensive experiments on the MMEB benchmark and multiple retrieval tasks, demonstrating that our method achieves state-of-the-art performance across all tasks.

Traditional post-training quantization (PTQ) is considered an effective approach to reduce model size and accelerate inference of large-scale language models (LLMs). However, existing low-rank PTQ methods require costly fine-tuning to determine a compromise rank for diverse data and layers in large models, failing to exploit their full potential. Additionally, the current SVD-based low-rank approximation compounds the computational overhead. In this work, we thoroughly analyze the varying effectiveness of low-rank approximation across different layers in representative models. Accordingly, we introduce Flexible Low-Rank Quantization (FLRQ), a novel solution designed to quickly identify the accuracy-optimal ranks and aggregate them to achieve minimal storage combinations. FLRQ comprises two powerful components, Rank1-Sketch-based Flexible Rank Selection (R1-FLR) and Best Low-rank Approximation under Clipping (BLC). R1-FLR applies the R1-Sketch with Gaussian projection for the fast low-rank approximation, enabling outlier-aware rank extraction for each layer. Meanwhile, BLC aims at minimizing the low-rank quantization error under the scaling and clipping strategy through an iterative method. FLRQ demonstrates strong effectiveness and robustness in comprehensive experiments, achieving state-of-the-art performance in both quantization quality and algorithm efficiency.

Identifying in-vehicle electronic control units based on voltage characteristics has been the subject of extensive research in cybersecurity. However, the results reported so far generally depend on restricted datasets and supervised learning. In this work, we show that clustering, i.e., unsupervised learning, of voltage characteristics, is in fact more challenging when done on a larger pool of electronic control units as several out-of-the-box clustering methods and metrics will fail to determine the correct number of clusters when exerted over a large dataset. To overcome this issue, we propose a new methodology that takes advantage of domain-specific constraints, which guide the search toward the correct number of electronic control units in a car, or even in a larger pool of units from several cars. We introduce two new metrics: correctness, which measures the success ratio with respect to the constraints, and divergence, which measures the consistency of the clustering, and show that they provide a strong indication for the optimal number of clusters. In this specific context, both metrics prove to be more reliable than the widely used Silhouette score, Davies-Bouldin and Calinski-Harabas indexes. We successfully test our methodology on the largest dataset available today for in-vehicle voltage characteristics and discover new insights regarding the number of devices.

Generalization to unseen environments is a significant challenge in the field of robotics and control. In this work, we focus on contextual reinforcement learning, where agents act within environments with varying contexts, such as self-driving cars or quadrupedal robots that need to operate in different terrains or weather conditions than they were trained for. We tackle the critical task of generalizing to out-of-distribution (OOD) settings, without access to explicit context information at test time. Recent work has addressed this problem by training a context encoder and a history adaptation module in separate stages. While promising, this two-phase approach is cumbersome to implement and train. We simplify the methodology and introduce SPARC: single-phase adaptation for robust control. We test SPARC on varying contexts within the high-fidelity racing simulator Gran Turismo 7 and wind-perturbed MuJoCo environments, and find that it achieves reliable and robust OOD generalization.

Model-free deep reinforcement learning (RL) algorithms have achieved tremendous success on a range of challenging tasks. However, safety concerns remain when these methods are deployed on real-world applications, necessitating risk-aware agents. A common utility for learning such risk-aware agents is the entropic risk measure, but current policy gradient methods optimizing this measure must perform high-variance and numerically unstable updates. As a result, existing risk-sensitive model-free approaches are limited to simple tasks and tabular settings. In this paper, we provide a comprehensive theoretical justification for policy gradient methods on the entropic risk measure, including on- and off-policy gradient theorems for the stochastic and deterministic policy settings. Motivated by theory, we propose risk-sensitive exponential actor-critic (rsEAC), an off-policy model-free approach that incorporates novel procedures to avoid the explicit representation of exponential value functions and their gradients, and optimizes its policy w.r.t. the entropic risk measure. In this way, we show that rsEAC produces more numerically stable updates compared to existing approaches and reliably learns risk-sensitive policies in challenging risky variants of continuous tasks in MuJoCo.

Resource-constrained weight deployment is a task of immense practical importance. Recently, there has been interest in the specific task of Delta Compression, where parties each hold a common base model and only communicate compressed weight updates. However, popular parameter efficient updates such as Low Rank Adaptation (LoRA) face inherent representation limitations - which are especially pronounced when combined with aggressive quantization. To overcome this, we build on recent work that improves LoRA representation capacity by using fixed-frequency sinusoidal functions to increase stable rank without adding additional parameters. We extend this to the quantized setting and present the first theoretical analysis showing how stable rank evolves under quantization. From this, we introduce SineLoRA∆, a principled and effective method for delta compression that improves the expressivity of quantized low-rank adapters by applying a sinusoidal activation. We validate SineLoRA∆ across a diverse variety of domains - including language modeling, vision-language tasks, and text-to-image generation - achieving up to 66% memory reduction with similar performance. We additionally provide a novel application of the canonical Bjøntegaard Delta metric to consistently compare adapter compression changes across the rate-distortion curve.

Many reinforcement learning algorithms, particularly those that rely on return estimates for policy improvement, can suffer from poor sample efficiency and training instability due to high-variance return estimates. In this paper we leverage new results from off-policy evaluation; it has recently been shown that well-designed behaviour policies can be used to collect off-policy data for provably lower variance return estimates. This result is surprising as it means collecting data on-policy is not variance optimal. We extend this key insight to the online reinforcement learning setting, where both policy evaluation and improvement are interleaved to learn optimal policies. Off-policy RL has been well studied (e.g., IMPALA), with correct and truncated importance weighted samples for de-biasing and managing variance appropriately. Generally these approaches are concerned with reconciling data collected from multiple workers in parallel, while the policy is updated asynchronously, mismatch between the workers and policy is corrected in a mathematically sound way. Here we consider only one worker - the behaviour policy, which is used to collect data for policy improvement, with provably lower variance return estimates. In our experiments we extend two policy-gradient methods with this regime, demonstrating better sample efficiency and performance over a diverse set of environments.