IJCAI 2026 Papers

Cross-view geo-localization (CVGL) aims at localizing a ground-level query by retrieving its corresponding match from a database of geo-tagged satellite images. Existing multi-modal CVGL methods lack a structured design in the fusion stage, limiting their ability to fully exploit the information from multiple modalities. To overcome this limitation, we propose a Structural-Analysis-Based fusion principle that guides the design of network architecture. Following this principle, we present Decoupled Query Fusion (DQF), a novel fusion module that decouples feature interactions through role-specific learnable queries. These learnable queries aggregate features into distinct slots. Specifically, modality-specific queries capture unique information from each modality, while cross-modal queries extract redundant and synergistic information between different modalities. To better discriminate positive samples at varying geographic distances, we further propose Geo-aware Circle Loss, which adaptively weights supervision signal based on the geographical distance between samples and anchors. Extensive experiments on large-scale benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches. Comprehensive ablation studies further validate the effectiveness of each component. Our code and models will be made publicly available.

Test-time scaling (TTS) has demonstrated remarkable potential in enhancing the reasoning capabilities of Large Language Models (LLMs) and Large Vision-Language Models (LVLMs). However, its application has primarily been limited to domains such as mathematics and programming, owing to their reasoning-intensive nature and the ease of result verification. Its utility in other knowledge-intensive fields, such as medicine and general scientific research, remains underexplored. To bridge this gap and unlock the potential of TTS in broader domains, we propose Cross-Domain TTS, a novel framework that enables task-tailored scaling. This framework consists of two key components: a conformal prediction-based cold-start strategy and an information-gain-based dynamic reasoning adjustment. The CP-based cold-start strategy guides the model's initialization during test-time scaling based on conformal prediction theory, while the information-gain-based dynamic reasoning adjustment guides the model's reasoning progress through a progress vector according to the information gain of reasoning steps. We conducted experiments using LLMs and LVLMs on cross-domain benchmarks. Our results demonstrate that the proposed framework consistently improves performance across various domain-specific datasets. For instance, in the medical domain, it achieves an improvement of up to 17% in pass@1 accuracy while reducing inference latency and saving up to 30% in token consumption. Code is available at https://github.com/Yan0613/Cross-Domain-TTS.

Short video harmful content detection aims to automatically identify diverse anomalies from user-generated media. This task presents unique challenges due to frequent editing cuts and highly variable anomaly densities, limiting the effectiveness of traditional surveillance-based approaches. Moreover, existing Vision-Language Model-based approaches typically rely on rigid instance selection mechanisms that fail to adapt to the unpredictable duration of anomalies in such unconstrained videos. To address these issues, we propose SVLA, a Shot-conditioned Vision-Language Adaptation framework, for effectively detecting harmful contents from online short videos. Our approach introduces a novel π-adaptive strategy to dynamically estimate shot-level anomaly density, replacing rigid selection with calibrated supervision. Furthermore, we employ a shot-conditioned temporal encoder to respect video hierarchy and adopt a dual-path contextual adapter to resolve semantic ambiguity. To benchmark this task, we construct a new dataset (SVA) covering more genuine online short videos that involve seven anomaly categories. Experiments on the SVA dataset demonstrate that SVLA can achieve the state-of-the-art performance and outperform its competitors across diverse scenarios. Codes and datasets are available at: https://github.com/xushuai7/IJCAI-SVLA.

Zinc-based alloys are indispensable emerging absorbable metallic biomaterials, and their macroscopic performance is governed by microstructural characteristics. Intermediate phases—key microstructural constituents—are pivotal in regulating mechanical and functional properties. However, intermediate phase segmentation in zinc alloy microstructures faces formidable challenges: scarce annotated datasets, low contrast, difficulty detecting small targets, and heterogeneous morphologies. To this end, we construct IPSM-Bench, the largest high-quality dataset for zinc-alloy intermediate phase segmentation. Furthermore, we propose SCoP-SAM, a new Spatial Context Prior-guided SAM method that leverages the gradient structure and grayscale properties of intermediate phases to capture spatial context priors and incorporates them into the entire SAM encoding-decoding process, improving segmentation performance. Based on the proposed IPSM-Bench, we establish a new benchmark for intermediate phase segmentation to systematically evaluate state-of-the-art (SOTA) methods and advance research on zinc alloy microstructure analysis. Extensive experiments on IPSM-Bench and additional public alloy benchmarks demonstrate that our SCoP-SAM not only achieves SOTA performance for zinc-alloy intermediate phase segmentation but also generalizes remarkably well to other alloy scenarios.

With the rising demand for trustworthy AI in clinical practice, strong interpretability is now a critical requirement as well as accuracy. However, the modality gap for medical visual question answering is quite severe when continuous visual signals are forcibly projected into discrete text space for reasoning, and the loss of necessary diagnostic information leads to low precision and black-box opacity. To address this problem, we propose MedVCoT, which incorporates latent visual reasoning into the medical visual question answering(VQA) domain. Rather than merely integrating modules, MedVCoT utilizes the specialized expertise of MedSAM to train a large vision-language model so that it can autonomously generate consistent and continuous latent visual tokens within Visual Chain-of-Thought. This mechanism forces the model to explicitly "see" the lesion in the latent space before formulating a textual diagnosis, ensuring answers are causally rooted in verifiable visual evidence rather than statistical hallucination. We achieve this through a progressive 3-stage training procedure: medical feature alignment, visual reasoning learning by utilizing latent tokens generated, and instruction tuning for complex clinical scenarios. Extensive experiments show that MedVCoT can achieve state-of-the-art performance on multiple benchmarks, outperforming other methods by large margins. Meanwhile, it provides pixel-level segmentation masks to validate its diagnostic reasoning. Our demo is available at https://zhuqh19.github.io/MedVCoT.

Retinal angiography provides critical vascular information, yet optical coherence tomography angiography (OCTA) acquisition remains slower and less accessible than conventional structural OCT. A key open question is: how angiographic signals can be recovered? To address the issues of artifacts and structural blurring in existing generative methods, we propose a wavelet flow matching-based model, FlowOCT. This method operates directly on OCT wavelet coefficients, learning frequency-specific velocity fields to map them to the OCTA wavelet distribution. To ensure anatomical accuracy of the vasculature, we introduce constraints such as depth-wise contrastive alignment and two-dimensional projection consistency. Given the sparse nature of vascular signals in OCTA data, a structure-focused loss function and corresponding evaluation metrics are designed. Experiment results show that FlowOCT outperforms existing methods in terms of both image quality and perceptual similarity. Downstream diagnostic tasks further validate its superior authenticity and generalization capability. Code is available at https://github.com/cnu-medilab/FlowOCT.

Spiking Neural Networks (SNNs) offer notable advantages in biological plausibility and energy efficiency, making them promising candidates for building low-power Transformers. However, existing Spiking Transformers largely adhere to a passive reactive paradigm, which struggles to focus on task-relevant information and incurs substantial computational overhead when processing redundant visual data. To overcome this fundamental yet underexplored limitation, we propose SAFformer, a novel Spiking Transformer architecture based on an active predictive filtering paradigm. Inspired by the brain’s predictive coding mechanism, SAFformer actively suppresses predictable signals and focuses on salient visual features. Extensive experiments show that SAFformer establishes new state-of-the-art performance on CIFAR-10/100 and CIFAR10-DVS. Remarkably, on ImageNet-1K, it achieves 80.44% Top-1 accuracy with only 26.58M parameters and an energy consumption of 5.88 mJ, demonstrating an exceptional balance between accuracy and efficiency.

Source-Fully-Free Domain Adaptation (SFF-DA) has emerged as a strategic paradigm to adapt Vision-Language Models (VLMs) without any access to source data or task-specific source models. However, we identify a critical "Dual Semantic Drift" that hinders this process: static drift caused by the stagnation of frozen hand-crafted templates, and dynamic drift resulting from noisy, instance-level generated captions. These issues lead to a severe stability-plasticity dilemma, manifesting as semantic misalignment and class collapse. To address this, we propose DSSG (Dual-Stream Semantic Guidance), an end-to-end framework that reconciles fine-grained plasticity with global stability. Our core contribution is the Dual Semantic Guidance (DSG) module, which integrates a captioning stream for detailed domain nuances with a template stream to anchor global categorical consistency. Furthermore, a Dynamic Cross-Modal Knowledge Distillation (CMKD) module is introduced to adaptively optimize teacher-student alignment across modalities. Extensive experiments demonstrate that DSSG significantly outperforms current state-of-the-art methods across multiple benchmarks, providing a robust solution for the unsupervised transfer of foundation models. The code is available on https://github.com/mrmenand/DSSG.

Continual unsupervised domain adaptation (UDA) for multi-domain cardiac image segmentation is crucial for clinical deployment under strict privacy constraints, where data from different centers cannot be stored or revisited. However, existing methods still suffer from catastrophic forgetting and training instability caused by noisy pseudo-labels and representation drift. In this study, we propose a novel continual UDA framework that enforces dual-level alignment at the input and feature levels. At the input level, the Domain-Style Adapting Generative Replay (DSA-GR) module employs trajectory-consistent diffusion distillation and style adaptation to synthesize high-fidelity, domain-consistent replay images without storing raw data. At the feature level, the Cross-domain Prototype-guided Feature Consolidation (CPFC) mechanism mitigates representation drift by maintaining a cross-domain prototype memory bank that aggregates features from the current and replay streams. These prototypes act as robust anchors for prototype-guided contrastive learning, rectifying the feature space and suppressing pseudo-label drift. Extensive experiments on two public datasets demonstrate that our framework consistently outperforms state-of-the-art methods. Code will be released at https://github.com/hlyf-xs/CUDA_Seg.

With the increasing demand of privacy protection, Machine Unlearning (MU) appears to remove private data from an already trained model without retraining from scratch. Most current works suffer from overly unlearning (low fidelity) or incomplete unlearning (low effectiveness). To identify the issues behind, we conduct causal analysis to obtain a resolvable route, i.e., disentangling shared knowledge into attribute-level semantics to remove it as the confounder. We further perform MU loss analysis to reformulate it as balanced form of constraints, thus guaranteeing high fidelity and effectiveness. Based on theoretical analysis, we propose disentangled knowledge forgetting constrained by the reformulated MU loss, which disentangles knowledge with variational auto-encoder and refines knowledge with counterfactual inference. Extensive experimental results demonstrate that our method achieves state-of-the-art performance.

Federated Learning (FL) faces two major challenges when deploying large vision-language models such as contrastive language-image pretraining (CLIP) model, namely high computation/communication overhead and performance degradation due to data heterogeneity. To address these, we propose FedENC, a novel federated adaptive CLIP framework. FedENC keeps the CLIP encoder frozen and introduces a lightweight communicable global expert adaptation module for learning general knowledge across clients and two local experts for client-specific adaptation to reduce computation and communication overhead. To mitigate data heterogeneity, we introduce neural collapse (NC) alignment to encourage the model to learn unified and balanced image feature representations. To enrich the performance, we propose a cosine and Kullback–Leibler (KL) divergence based alignment function to improve the prediction consistency between the global and local experts. Finally, we ensemble their outputs for robust inference. Extensive experiments on six datasets (4 natural + 2 medical) in challenging settings (e.g., domain shift) show that FedENC improves the test accuracy (e.g., by +5.37% on DomainNet compared to LoRA) with reasonable communication load (e.g., 15.2x lower than LoRA, with 3.3GB GPU memory usage). The code is available at https://github.com/AIPMLab/FedENC.

Existing AI-generated image (AIGI) detectors perform well in-domain but degrade severely under distribution shift. We observe that this failure is mainly caused by content shortcuts, where detectors spuriously couple forgery artifacts with semantic content, such as object categories or demographic attributes, learning content–label correlations instead of generalizable forgery patterns. To address this issue, we propose PURE (Purging Unrelated Representations for Content-Agnostic Forgery Detection), which achieves content-agnostic detection through two complementary components: a Causal Semantic Generative (CSG) mechanism that disentangles semantic representations from forgery-irrelevant nuisance factors, and a Gaussian Mixture Model (GMM)-based prototype alignment module that suppresses category-specific content bias. Extensive experiments on CIFAKE, GenImage, and AlFace show that PURE achieves superior generalization under spurious correlation reversal. The code is available at https://github.com/wuxinyu519/PURE.

Few-shot learning aims to recognize novel categories from limited labeled samples, where prototypes estimated from 1--5 supports per class are often unreliable. Semantic-based approaches alleviate this by introducing class-level priors, but they often ignore instance-level cues and rarely optimize queries under the inductive protocol. We propose PMCE, a Probabilistic framework that leverages Multi-granularity semantics with Caption-guided Enhancement for few-shot classification. On base classes, we build a knowledge bank with class-wise visual statistics and class-name embeddings. At test time, the semantic embedding of a novel class retrieves a few similar base classes whose visual priors are aggregated into a class-specific prior and combined with the support-based prototype via MAP estimation. Simultaneously, we train a lightweight enhancer on base classes that fuses frozen BLIP captions with visual features, and apply it to both supports and queries without using query-set statistics or novel labels. A simple caption-consistency regularizer further improves robustness to noisy captions. Experiments on four standard benchmarks with ResNet-12 and Swin-T backbones show that PMCE outperforms state-of-the-art few-shot baselines, achieving up to 7.71% gains over the strongest competitor on MiniImageNet in the challenging 1-shot setting. Our code is available at https://github.com/channa419/PMCE.

Unsupervised 3D semantic segmentation is vital for label-free open-world perception. However, current methods typically struggle with two core limitations: fixed category assumptions that restrict adaptability in complex scenes, and geometric incompleteness caused by irregular point sampling, which degrade the integrity of local geometric descriptors for fine-grained or long-tail objects. To address these, we propose DyG-Seg, a Dynamic Geometric-Semantic Alignment framework. First, targeting the challenge of incomplete geometry, the Geometric Manifold Rectification (GMR) leverages a diffusion model to reconstruct integral object geometry and recover the underlying manifold topology, thereby enhancing feature discriminability. Utilizing this geometric prior, the Dynamic Prototype Contrastive Clustering (DPCC) applies non-parametric Bayesian inference to automatically estimate the optimal number of clusters and generate pseudo-labels. Finally, our Iterative Optimization with Dynamic Constraints integrates Dynamic Class Balancing (DCB) to mitigate long-tail bias. Capitalizing on the recovered manifold topology, we enforce Spatial Geometric Consistency to rectify structural incompleteness and ensure spatially coherent semantic predictions. Extensive experiments on ScanNet and S3DIS demonstrate that DyG-Seg discovers latent semantic structures and significantly outperforms state-of-the-art unsupervised methods.

While Vision-Language Models (VLMs) excel at semantic understanding, they struggle to comprehend 3D spatial relationships from limited views. Their reliance on implicit geometric encoding often leads to severe hallucinations and inconsistencies in spatial reasoning tasks. To address this, we introduce GeoMind, a model-then-reason framework that employs a single LLM to autoregressively generate an explicit Geometric Description Language (GDL) map, serving as a grounded context to derive the final answer. This intermediate GDL map provides an explicit and queryable world representation. Leveraging this explicit representation, we enforce a strict referential constraint, compelling the model to ground reasoning solely on the instantiated entities to ensure referential integrity and auditability. Specifically, we lift multi-view observations into object-centric tokens using frozen geometric priors and instance masks. The LLM is trained via a two-stage curriculum with programmatic supervision to generate the GDL map as a prerequisite for answering. On five spatial understanding benchmarks in both image and video settings, GeoMind delivers average accuracy gains of +6.9% (2B) and +9.8% (8B) over Qwen3-VL baselines. Our results suggest that explicit geometric grounding enables robust spatial reasoning without human annotation, providing a scalable and practical route to stronger spatial intelligence in large VLMs.

Low-Light Image Enhancement (LLIE) has long been a challenging problem in low-level vision, as insufficient illumination often leads to low contrast, detail loss, and noise. Recent studies show that deep learning-based Retinex theory can effectively decouple illumination and reflectance. However, existing methods frequently suffer from over-enhancement or color distortion, and often assume uniform noise or ideal lighting. To address these limitations, we propose InterLight, a novel framework that systematically excavates and operationalizes intrinsic illumination priors for LLIE. Our core insight is that robust enhancement requires not just estimating illumination, but constructing an illumination-aware pipeline. We first inject sensor-level illumination-response priors via physics-guided augmentation, then represent the degradation through adaptive prompts conditioned on the scene's latent illumination state. This explicit representation directly guides a luminance-gated intrinsic memory mechanism to selectively compensate for information loss, prioritizing reconstruction in dark regions while preserving fidelity in bright ones. Finally, the entire process is regularized by a self-supervised consistency objective that distills illumination-invariant features. By deeply exploiting intrinsic illumination priors, our method achieves clearer textures and more visually coherent enhancement results. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our approach. Code is available at: https://github.com/House-yuyu/InterLight.

Pansharpening aims to fuse a high-resolution panchromatic (PAN) image with a low-resolution multispectral (MS) image to generate a highresolution multispectral output that preserves both fine spatial details and faithful spectral responses. However, enhancing spatial textures without introducing spectral distortion remains challenging due to the inherent spectral–spatial trade-off. To address this issue, we propose a two-stage pansharpening framework that tackles the problem from both modeling and optimization perspectives. In the first stage, we formulate pansharpening as spectral-prior conditioned residual diffusion, where a stable spectral base constrains the generation process and allows the diffusion model to focus on PAN-guided high-frequency details, leading to improved training stability and reduced spectral drift. To better capture the coupled spatial and frequency characteristics of PAN–MS fusion, we adopt a complex-valued denoising network to enhance spectral–spatial interaction modeling. In the second stage, to bridge the gap between distortionoriented training objectives and practical quality preferences, we introduce Group Relative Preference Optimization (GRPO) to fine-tune the diffusion model using multi-objective preference signals, explicitly balancing spectral fidelity, texture sharpness, and perceptual quality. Extensive experiments on standard benchmarks demonstrate that the proposed method achieves a more favorable trade-off between fidelity and perceptual quality compared to competitive end-to-end and diffusionbased approaches.

Continual Visual Question Answering (VQA) requires learning from non-stationary streams of visual inputs and questions while preserving past knowledge. Most prior methods adapt by updating a largely shared parameter set. This often leads to cross-level task interference, hindering accurate adaptation to the current task and object. To address this limitation, we propose HyLoVQA. It maintains a drift-resilient memory bank of anchors. The bank stores the content of visual objects and textual tasks, and they are updated using current input features. Conditioned on retrieved anchors, a hypernetwork generates lightweight Low-Rank Adaptation (LoRA) adapters. This ensures parameter efficiency, allowing the model to adapt to each task and object dynamically. Additionally, we formulate an alignment loss that aligns semantic discrepancies in the feature space with functional changes in the parameter space, thereby constraining LoRA adapters to remain focused on the current task and object. Extensive experiments on VQA v2 and NExT-QA under both standard and compositional settings demonstrate the superiority of HyLoVQA over prior state-of-the-art methods. The code is available at https://github.com/HubuKG/HyLoVQA.

We present S³-TIR (Coupled Spatial-Spectral Splatting), a physics-inspired framework for high-fidelity thermal infrared (TIR) novel view synthesis. Existing 3D Gaussian splatting methods struggle to capture the unique duality of TIR imagery: the coexistence of diffusive smoothness from heat conduction and structural sharpness from material discontinuities. To address this issue, we leverage the intrinsic frequency properties of thermal scenes, where conduction creates smooth fields while discontinuities cause sharp signal jumps. Accordingly, S³-TIR couples adaptive spatial support with internal spectral modulation via a unified generalized-exponential Gabor primitive. This design utilizes spatial support as a physical truncation controller, confining high-frequency textures within sharp boundaries while maintaining smooth diffusion elsewhere. Furthermore, we propose a frequency-aware deferred splatting scheme that enforces multi-view frequency consistency, anchoring frequency distributions to the 3D structure rather than view-dependent appearance. Experiments on diverse benchmarks demonstrate that S³-TIR achieves state-of-the-art performance, notably reducing LPIPS by 42.5% and improving PSNR by 3.2% on the RGBT-Scenes dataset, while yielding structurally stable thermal sequences free from jittering artifacts.