IJCAI 2026 论文列表

Efficient transfer learning methods for large-scale vision–language models (e.g., CLIP) enable strong few-shot transfer, yet existing adaptation methods follow a fixed fine-tuning paradigm that implicitly assumes a uniform importance of the image and text branches, which has not been systematically studied in image classification. Through extensive analysis, we reveal a Branch Bias issue in vision–language image classification: adapting the image encoder does not always improve performance under out-of-distribution settings. Motivated by this observation, we propose A₃B₂, an Adaptive Asymmetric Adapter that alleviates Branch Bias in few-shot learning. A₃B₂ introduces Uncertainty-Aware Adapter Dampening (UAAD), which automatically suppresses image-branch adaptation when prediction uncertainty is high, enabling soft and data-driven control without manual intervention. Architecturally, A₃B₂ adopts a lightweight asymmetric design inspired by mixture-of-experts with Load Balancing Regularization. Extensive experiments demonstrate that A₃B₂ consistently outperforms 11 competitive prompt- and adapter-based baselines.

The deployment of Vision-Language Models (VLMs) faces memory and computational bottlenecks because of the massive parameters and intensive computations. While Post-Training Quantization (PTQ) can reduce these costs, existing methods often overlook the heterogeneity of multimodal input when applied to VLMs, leading to quantization accuracy degradation. In this paper, we categorize the Transformer linear layers into Expansion Space (responsible for feature expansion) and Projection Space (responsible for feature aggregation) and reveal a cross-modal structural sensitivity asymmetry in VLMs: The Expansion Space is more sensitive to quantization accuracy than the Projection Space in both LLM and ViT encoders. Based on this, we propose SeGO, a unified structural sensitivity-aware sparse optimization framework. First, it constructs search interval for scaling factors using channel-wise statistics of LLM and ViT weights to exclude invalid solutions. Then it applies scaling protection only to the sensitive Expansion Space while directly quantizing the robust Projection Space. Moreover, SeGO uses a Golden-Section Solver to speedup the search of optimal scaling factors. Experiments show that SeGO achieves the balance among model parameter amount, quantization accuracy and scaling factors' search efficiency on InternVL2 and LLaVA series. Under the W3/W4A16 configuration, SeGO improves the accuracy of 7B-26B models by 3.8%-17.3% and enhances calibration efficiency by dozens of times.

Remote sensing change detection (RSCD) aims to identify changed regions in bitemporal images. However, conventional one-step modeling suffers from performance degradation caused by imaging temporal differences (e.g., illumination disturbances, seasonal variations). To address this issue, we formulate the problem as an “adversarial attack and defense” paradigm. Then, instead of traditional adversarial training, we propose a brain-inspired deep network based on the drift diffusion model (DDM) decision-making mechanism for RSCD, which is rooted in cognitive neuroscience. By simulating brain decision processes, the network enhances model robustness significantly. Specifically, we design the multi-stage reasoning unit to simulate DDM’s iterative evidence generation, stochastic representation unit with reparameterization sampling to model stochastic diffusion in decision-making, decision aggregation unit to replicate DDM’s evidence accumulation, and loss-guided unit with multi-loss joint constraints to simulate directional drift. Experimental results verify the model’s superior performance on multiple datasets and their perturbed versions.

Low-Light Image Enhancement (LLIE) remains challenging due to severe noise, color artifacts, and structural degradation in extreme low-light environments. Existing methods primarily rely on photometric cues and often fail to preserve geometric consistency, resulting in blurred edges and distorted textures. To address this limitation, we introduce LAMP, a large-scale dataset of Low-light Aligned Multimodal Pairs for LLIE. LAMP contains over 18k high-quality aligned pairs and consists of two complementary subsets. LAMP-Real is captured using LiDAR-based depth sensors under physically controlled illumination, while LAMP-Synthetic is generated through a physics-calibrated degradation pipeline. Both subsets provide precisely aligned RGB images, depth maps, and pixel-wise semantic annotations. Building upon LAMP, we propose SPACE, a Structure Preservation Aware Cross-modal Enhancement framework that explicitly leverages geometric priors. SPACE introduces a Depth-Adaptive HVI Transformation to decouple luminance and chrominance under depth guidance, effectively suppressing color-space noise. Furthermore, a Depth-Manifold Modulated Attention mechanism constrains feature interactions within a learned depth manifold, ensuring structural coherence during enhancement. Extensive experiments demonstrate that SPACE consistently outperforms state-of-the-art methods in visual quality and structural fidelity. The code and data are available at https://github.com/YueCheong/SPACE.

Parameter-efficient adaptation methods, such as adapters and prompt learning, have become popular for transferring pre-trained vision-language models (VLMs) to downstream tasks. However, under limited supervision during adaptation, overly aggressive cross-modal alignment can distort the intrinsic structure of modality-specific representations, leading to degraded generalization. In this paper, we propose Co-adaptive Dual-path Alignment (CoDA), a new adaptation framework that explicitly disentangles and coordinates cross-modal semantic alignment and intra-modal structural consistency. In CoDA, we also propose a parent class labeling strategy that injects hierarchical semantic priors into textual prompts, further stabilizing alignment under limited supervision. Extensive experiments on eleven benchmark datasets show that CoDA outperforms state-of-the-art parameter-efficient methods, particularly under few-shot learning and distribution-shift scenarios. Our code is available at https://github.com/yizhang-ac/CoDA.

Physical-layer device identification serves as a critical security mechanism for mitigating cyberattacks and enhancing network resilience. However, its deployment in modern full-duplex Ethernet networks faces two major challenges: (1) the presence of Out-of-Distribution (OOD) data caused by temporal distribution shift and diverse local transmitters, and (2) the difficulty of environment-agnostic device identification, where the unique features of target devices are often obscured in mixed signals. To address these challenges, we propose GenID, which consists of two core components. First, GenID constructs prototypes to emulate the full communication environment through a two stage process: offline and online. Based on the prototype context, it then introduces an invariant feature extraction method to disentangle mixed signals and improve generalization under OOD conditions.Extensive experiments demonstrate that GenID significantly outperforms state-of-the-art baselines, achieving an increase of at least 7.94\% in Macro-F1 score under OOD environments. Our code is available on https://github.com/zyw2004/GenID_IJCAI2026.

Existing image deraining models struggle to handle complex real-world scenarios with rain-mist coupled degradation. The core challenge stems from the high visual similarity and spatial coupling between rain streaks and mist, making it difficult to accurately model their joint degradation patterns. Consequently, both specialized deraining models and all-in-one models for multiple degradations are ineffective in rain-mist coexistence scenarios. To address these challenges, we propose a novel Rain-Mist Removal (RMR) framework. It effectively utilizes single degradation priors from existing deraining and dehazing datasets to model the joint degradation, thereby achieving effective rain streak removal while preserving background structures obscured by mist. To enhance the generalization to real-world scenarios, we leverage text prompts trained in the CLIP perceptual space to drive the generated results toward real samples. Extensive experiments demonstrate that the proposed RMR outperforms state-of-the-art methods in rain-mist coexistence scenarios.

Unified remote sensing multimodal models exhibit a pronounced spatial reversal curse: although they can accurately recognize and describe object locations in images, they often fail to faithfully execute the same spatial relations during text-to-image generation, where such relations constitute core semantic information in remote sensing. Motivated by this observation, we propose Uni-RS, the first unified multimodal model tailored for remote sensing, to explicitly address the spatial asymmetry between understanding and generation. Specifically, we first introduce explicit Spatial-Layout Planning to transform textual instructions into spatial layout plans, decoupling geometric planning from visual synthesis. We then impose Spatial-Aware Query Supervision to bias learnable queries toward spatial relations explicitly specified in the instruction. Finally, we develop Image–Caption Spatial Layout Variation to expose the model to systematic geometry-consistent spatial transformations. Extensive experiments across multiple benchmarks show that our approach substantially improves spatial faithfulness in text-to-image generation, while maintaining strong performance on multimodal understanding tasks like image captioning, visual grounding, and VQA tasks.

Denoising Diffusion Codebook Models (DDCM) have emerged as a promising framework for zero-shot image compression by replacing stochastic sampling with discrete selection from a reproducible Gaussian codebook. By greedily picking noise vectors that best match the target image, DDCM encodes the generative trajectory into a compact sequence of indices for image compression. However, the target-guided noise selection provides only coarse guidance as the discrepancy between the denoiser's output and the target image is substantial. This discrepancy inevitably results in trajectory drift, necessitating massive codebooks to span the enlarged search space and extensive inference steps to gradually compensate for the deviation. In this paper, we propose the Trajectory-Consistent Diffusion Codebook Model (TC-DDCM) to address these inefficiencies. Rather than approximating the distant target, TC-DDCM navigates the generative process by strictly tracking a pre-computed inversion trajectory. This paradigm shift yields two critical advantages: 1) By minimizing the deviation to the proximal reference state along the trajectory, the search space is drastically shrunk, resulting in significantly smaller codebooks. 2) By aligning the generative steps with inversion methods, TC-DDCM achieves optimal reconstruction in a few inference steps. Extensive experiments demonstrate that TC-DDCM significantly outperforms state-of-the-art zero-shot methods in rate-distortion-perception performance at ultra-low bitrates, making zero-shot diffusion-based compression practical.

Automatic radiology report generation (RRG) aims to simulate the workflow of radiologists, assisting them in clinical diagnosis. However, existing methods often fall short in utilizing all information relevant to the examination, as is typically done in clinical practice. Although some works attempt to incorporate multi-view images and historical data, these additional inputs may sometimes lead to avoidable diagnostic errors on the contrary. To address these challenges, we introduce a decision-making stage after report generation for the first time and propose a Posterior Diagnostic Decision framework (PDD-RRG) to integrate potentially conflicting diagnoses. Specifically, we create various subsets of input data and utilize an existing RRG model to generate reports from different perspectives. Then the Bayesian posterior probability and the learned thresholds for each clinical observation are calculated to obtain an aggregated diagnostic conclusion, which is subsequently used to refine the generated report. Experiments on MIMIC-CXR demonstrate that our proposed PDD-RRG can effectively enhance the clinical efficacy of existing RRG models without any retraining.

Vision-Language-Action (VLA) models have demonstrated strong performance in robotic manipulation, yet their closed-loop deployment is hindered by the high latency and compute cost of repeatedly running large vision-language backbones at every timestep. We observe that VLA inference exhibits structured redundancies across temporal, spatial, and depth dimensions, and that most existing efficiency methods ignore action context, despite its central role in embodied tasks. To address this gap, we propose Action-Context-aware Adaptive Computation for VLA models (AC2-VLA), a unified framework that conditions computation on current visual observations, language instructions, and previous action states. Based on this action-centric context, AC2-VLA adaptively performs cognition reuse across timesteps, token pruning, and selective execution of model components within a unified mechanism. To train the adaptive policy, we introduce an action-guided self-distillation scheme that preserves the behavior of the dense VLA policy while enabling structured sparsification that transfers across tasks and settings. Extensive experiments on robotic manipulation benchmarks show that AC2-VLA achieves up to a 1.79 X speedup while reducing FLOPs to 29.4% of the dense baseline, with comparable task success.

Neural network pruning is crucial for efficient deployment on resource-constrained devices, yet achieving high sparsity often leads to significant robustness degradation against adversarial perturbations and corruptions. Recent works typically rely on single-model fine-tuning along a fixed optimization trajectory, which renders the network susceptible to local optima and noise while failing to restore the multiple robustness properties compromised during compression. In this paper, we propose Mutual Heterogeneous Learning (MHL), a framework enabling robust pruning via single-model inference. MHL instantiates heterogeneity through two complementary mechanisms: layer-wise Lipschitz regularization for intermediate feature smoothness and adaptive margin objective for difficulty-aware boundary separation. To guide these diverse experts to converge, we employ entropy-based mutual distillation with a strategic schedule that shifts the optimization trajectory from exploring diverse feature subspaces to consolidating a unified robust model. Extensive experiments on 4 clean and corruption benchmarks and adversarial attacks demonstrate that MHL significantly outperforms single-model baselines in both adversarial robustness (+5%) and corruption robustness (+2.6%) while maintaining competitive clean accuracy.

With the explosive growth of text generated by large language models (LLMs) on the internet, technologies for detecting LLM-generated text are increasingly important. However, existing technologies often require frequent access to LLMs during detection, significantly restricting their application in scenarios where only the text is accessible. To tackle this challenge, we identify that human-written texts exhibit higher entropy than LLM-generated texts. Motivated by this, we propose a novel method named IED, which leverages the Information Entropy for generative text Detection. Specifically, IED first transforms the given text into a vector of which each dimension represents the information entropy of each word, and then adopts a classifier to conduct the detection. Further, we propose IEGD which adopts the information gain to construct the vector by enhancing the differentiation between human-written and LLM-generated texts. To evaluate the effective of the proposed method, we construct a new large-scale dataset comprised of generated texts using various LLMs (LLaMA-7B, GPT-NeoX, etc.). Extensive experiments show that the proposed methods achieve improvement over state-of-the-art methods by up to 10.81%.

Active learning aims to maximize model performance with minimal annotation costs by selecting the most informative samples from large unlabeled pools, which often face a budget dilemma: uncertainty-based methods induce redundancy under low budgets, while representativeness-based methods struggle to mine challenging samples under high budgets. Although some heuristic parameter interpolation schemes attempt to bridge this gap, such strategies suffer from a misalignment between theoretical assumptions and real-world distributions, failing to achieve a good exploration-exploitation balance across all scenarios. In this paper, we propose a general active learning framework based on Distribution-Aware Energy Minimization, which reformulates sample selection as minimizing the energy function for the distributional discrepancy between the selected subset and the global uncertainty field. This physical-inspired perspective naturally derives a Hamiltonian comprising attractive terms and repulsive terms, mathematically achieving an intrinsic and dynamic balance between uncertainty and representativeness. Furthermore, we transform the optimization objective to the ground-state search of an Ising Model, enabling efficient solutions via Coherent Ising Machine. Extensive numerical experiments show that our method outperforms prior state-of-the-art methods across multi-budget regimes on several benchmark datasets. Validation on a real quantum hardware also demonstrates the potential for quantum computer in future large-scale selection tasks.

Single-frame Infrared Small Target Detection (ISTD) aims to localize weak targets under heavy background clutter, yet dense pixel-wise annotations are expensive. Point supervision with online label evolution reduces annotation cost; however, lightweight CNN detectors often lack sufficient semantics, leading to noisy pseudo-masks and unstable optimization. To address this, we propose a hierarchical VFM-driven knowledge distillation framework that uses a frozen Vision Foundation Model (VFM) during training. We formulate point-supervised learning as a bilevel optimization process: the inner loop adapts a VFM-embedded teacher on reweighted training samples, while the outer loop transfers validation-guided knowledge to a lightweight student to mitigate pseudo-label noise and training-set bias. We further introduce Semantic-Conditioned Affine Modulation (SCAM) to inject VFM semantics into CNN features at multiple layers. In addition, a dynamic collaborative learning strategy with cluster-level sample reweighting enhances robustness to imperfect pseudo-masks. Experiments on diverse challenging cases across multiple ISTD backbones demonstrate consistent improvements in detection accuracy and training stability. Our code is available at https://github.com/yuanhang-yao/semantic-prior.

Medical image segmentation is still challenging, especially in semi-supervised scenarios where limited annotations are expected to support both accurate boundary delineation and coherent anatomical structures. We propose LR-GCF, a Local-Rotation-Driven Global Consistency Framework that couples strong local geometric perturbations with global consistency regularization in a dual-view architecture. Given an image, LR-GCF transforms it to different variants through patch-wise jigsaw-rotated perturbations. Such variants are decoded using parallel Swin Transformer decoders and perform consistency regularization by virtue of cross pseudo supervision. A rotation prediction head is designed to predict the rotation of each local patch. This makes locally discriminative representations robust to rotations and helps the model recognize global anatomical structure under strong perturbations. To further stabilize global modeling, we introduce a rotation-aligned attention consistency loss that encourages the inter-patch correlation across different variants to be as similar as possible. Moreover, we design a Position Encoding Duplicating scheme to propagate positional encodings from the low-resolution token map to high-resolution token maps, which helps alleviate the misalignment problem in rotation-aligned attention consistency. Extensive experiments on multiple medical image segmentation benchmarks demonstrate that LR-GCF consistently achieves superior performance under limited supervision, by jointly optimizing local detail extraction and stable global anatomical modeling. Source code is available at https://github.com/shuaiaihang/LR-GCF.

Transformer-based architectures have demonstrated strong performance in 3D point cloud understanding, yet many existing methods generate attention weights mainly from semantic feature similarity. In deep networks, feature-centric attention may become less selective as point features are progressively smoothed, a behavior associated with feature homogenization and rank collapse, which can weaken the structural discrimination of local aggregation.We propose PointGP, a geometry-primed framework that uses rectified local geometric topology as the primary cue for attention generation. PointGP introduces a Semantic-Guided Manifold Rectifier to predict feature-conditioned local coordinate offsets, and a Dual-Stream Geometric Kernel to compute attention logits from both raw and rectified geometric cues. By reducing reliance on explicit query-key feature matching while implicitly incorporating semantic guidance through geometric rectification, PointGP provides an effective and efficient mechanism for local point aggregation.Experiments across five benchmarks covering classification, part segmentation, and indoor scene segmentation show that PointGP achieves competitive accuracy with strong parameter and computational efficiency compared with representative strong baselines.

Whole-slide images (WSIs) provide gigapixel-scale visual evidence for cancer diagnosis, yet diagnostically relevant regions are typically sparse and embedded within large amounts of weakly informative tissue. Existing multiple instance learning methods often aggregate all patches by using attention mechanisms or sequence modeling, leading to redundant computation and limited adaptability to slides with highly variable sizes and lesion burdens. We propose Proportional Memory-guided Differential-attention Multiple Instance Learning (PMDMIL) network, a scalable MIL framework that follows a “less is more” principle for WSI classification. PMDMIL retains a fixed ratio of patches, pre-filters noise with a class-wise memory bank, and fuses the sparse survivors through differential attention to amplify subtle yet decisive morphological differences. Experiments on five public WSI benchmarks demonstrate that PMDMIL consistently outperforms representative MIL and sequence-based baselines across multiple evaluation metrics. Notably, on the DHMC dataset, our model can achieve competitive performance, while processing only 10% of patches, indicating that effective instance selection is more important than exhaustive aggregation for WSI classification.

Deep hashing is widely used for large-scale image retrieval but remains vulnerable to adversarial examples, raising practical security concerns. Existing targeted adversarial attacks on deep hashing typically rely on lp-norm constrained perturbations, which struggle to balance attack effectiveness and imperceptibility, often requiring perceptible noise and limiting their practicality in real-world retrieval scenarios. We propose UTDHA, the first unrestricted targeted attack for deep hashing models using contrastive-guided latent diffusion. UTDHA generates adversarial examples with a latent diffusion model and performs optimization in the latent space rather than the pixel space, enabling semantic manipulation while preserving image naturalness. Through contrastive guidance, the attack pulls adversarial examples toward the target label while pushing them away from non-target labels. Meanwhile, UTDHA enforces structural and perceptual consistency, producing adversarial examples that are both imperceptible and visually natural. Extensive experiments on three benchmarks demonstrate that UTDHA outperforms existing targeted adversarial attack baselines for deep hashing models in both attack effectiveness and imperceptibility.