IJCAI 2026 Papers

Gaze target estimation aims to infer the position of a person's gaze within a scene. Within mainstream design logic, multi-branch methods require extra supervision and annotations, while streamlined designs prioritize low-level visual saliency over true gaze intent. The former leads to a high annotation burden and hinders domain transfer, whereas the latter causes misalignment between predicted attention and actual gaze targets. To address this issue, we propose TextGaze, a unified cross-modal architecture that leverages a Large Vision-Language Model (LVLM) as scalable semantic guidance to balance the two design paradigms. The model extracts visual features from a frozen encoder and utilizes an LVLM to obtain gaze-aligned textual cues. We design a transformer-based fusion module with hierarchical text supervision to preserve task semantics. Lightweight decoding heads enable the joint prediction of gaze heatmaps and in-/out-of-frame status. We evaluate our method on four mainstream datasets, and the results show competitive performance across key metrics with robust cross-dataset generalisation without extra fine-tuning. Overall, we provide a streamlined alternative to traditional designs and highlight the potential of LVLMs as accessible auxiliary guidance for gaze estimation. All contents are available at: https://github.com/idremo/TextGaze-IJCAI2026.

The huge parameters and extensive training corpora of Large Language Models (LLMs) empower them to tackle complex tasks. Yet, their massive size incurs substantial inference costs, hindering deployment on resource‑constrained edge devices. Low‑bit quantization provides a way to compress models and improve their practicality for edge-side deployment. Nonetheless, many edge applications do not require the full generalization abilities of LLMs and instead only depend on their knowledge in specific domains. This paper introduces Gradient Enhancement Task Aware Post-training Quantization, i.e., GTAQ, to address the generalization issue. Concretely, GTAQ locates task‑critical weights via gradient‑based validation and retrieval, and then amplifies these salient weights. For that purpose, GTAQ preserves and strengthens weights related to the target tasks, enabling task‑aware uniform‑bit quantization. We extensively evaluate the LLaMA family of language models on WikiText, C4, and MMLU. Our experiments show that GTAQ consistently delivers notable performance improvements across tasks, surpassing both general‑purpose and task‑aware quantization baselines. At the same time, GTAQ yields more than 3.5× speedup in model inference and substantially cuts storage requirements. https://github.com/YihuaJerry/GTAQ.git.

While 3D Gaussian Splatting (3DGS) has excelled in dynamic scene reconstruction, it struggles with multi-view object motion blur, where view-dependent non-uniform blur violates fundamental multi-view geometric constraints. Existing methods fail to balance complex motion fitting with physical consistency across different views. To address the challenge, we propose TaylorMoDe-GS, the first 3DGS framework tailored for multi-view dynamic object deblurring. Specifically, we shift the modeling paradigm from displacement fitting to velocity driven modeling. This is achieved by analytically deriving instantaneous 3D velocities via Taylor series. To map 3D physical motion onto the 2D image plane, we propose a velocity splatting technique. Building upon this, we introduce a neural Peano remainder network to compensate for high frequency non-linear dynamics, effectively resolving the conflict between physical priors and fitting flexibility. Combined with a learnable physical blur synthesis mechanism, our framework ensures rigorous spatiotemporal and view consistency. Extensive experimental results validate that our method achieves notable advancements in restoring high-fidelity dynamic scenes and physically consistent motion fields, effectively addressing the challenges of multi-view dynamic object deblurring. The code is released at https://github.com/Chiffin-0816/TaylorMoDe-GS.

Recent deep learning-based ISP methods are primarily constrained by mimicking fixed camera pipelines, consequently struggling to achieve expert-level aesthetic quality. To this end, we propose AesISP, the first expert-level aesthetic ISP framework formulated as a reward flow model. Addressing the ill-posed nature of ISP, AesISP establishes a Privileged Prior Distillation paradigm. By leveraging expert-image-guided latent proxies, we decompose the intractable task into a tractable, proxy-driven learning process. Subsequently, we propose MeanFlow++, which reformulates the flow matching objective via target-prediction parameterization and a Progressive Temporal Curriculum. By evolving from learning instantaneous to long-range average velocities, this mechanism rectifies transport trajectories and uses deterministic mapping to impose explicit constraints, anchoring the generation trajectory to the expert manifold. Finally, we introduce a multi-dimensional reward-driven reinforcement learning approach.Leveraging Group Relative Policy Optimization (GRPO) to balance trade-offs across dimensions such as color and lighting, it steers the model to converge precisely on expert-level aesthetic standards. Experiments demonstrate that AesISP outperforms state-of-the-art methods in both quantitative metrics and aesthetic quality.

Automated radiology report generation aims to create clear and clinically correct diagnostic reports from medical images. Existing retrieval enhancement methods primarily focus on reusing textual knowledge, neglecting the crucial role of local visual pattern memory in clinical diagnosis. Furthermore, cross-modal retrieval lacking explicit clinical semantic constraints can easily introduce irrelevant pathological information, thereby reducing the clinical effectiveness of the generated reports. To address these challenges, we propose ClinAlign—a memory-based retrieval framework aligned with clinical workflow, drawing inspiration from clinical diagnostic workflows. Visually, we construct a disease-aware visual memory bank and enhance local patch representations through proposed Memory‑based Patch Pattern Augmentation (MPPA), thereby improving the perception and discrimination of pathological regions. On the textual side, we construct a disease-aware textual memory bank and introduce Classification-Guided Prompt Augmentation (CGPA), where disease state predictions are converted into structured diagnostic prompts to provide explicit semantic guidance for textual memory retrieval.Extensive experiments on two medical report generation benchmarks, MIMIC-CXR and IU X-Ray, demonstrate the effectiveness and practical value of our proposed method.

Multi-modality MRI lesion segmentation leverages complementary contrasts across modalities to better characterize lesion morphology and tissue alterations, thereby improving lesion delineation and reducing misdiagnosis. Existing correlation-driven methods estimate fusion weights by converting cross-modality correlation scores into adaptive fusion weights. However, in multi-center settings, domain shift changes cross-modal correlation patterns, which breaks the learned correlation-to-weight mapping and leads to modality-weight bias that amplifies center-specific shortcuts. To address this issue, we propose an Intervention-based Weight Calibration Network (IWCNet) that models domain shift as a controllable embedding and leverages explicit interventions to mitigate its impact on fusion weight estimation. IWCNet constructs original and intervened branches by performing interventions on the domain embedding, measuring inter-branch prediction effects to calibrate fusion weights. IWCNet further introduces Operator-level Domain Modulation (OLDM) to construct structure-consistent domain-variant feature pairs via operator-level domain intervention, and Effect-Driven Consistency Calibration (EDCC) to jointly leverage intervention effects and cross-branch prediction consistency to calibrate weights. Extensive experiments on multi-center multi-modality MRI datasets demonstrate that IWCNet substantially improves segmentation accuracy and robustness under domain shift.

Prototype-based federated learning enables efficient knowledge sharing by exchanging class prototypes rather than full model parameters. However, heterogeneous client data and limited local samples increase prototype estimation variance, making many client prototypes unreliable. Existing methods usually treat prototypes as deterministic point estimates and cannot quantify their reliability, which may contaminate global prototypes and cause negative transfer. To address these challenges, we propose FedUP, an uncertainty-aware personalized federated learning framework that models prototypes as probability distributions. FedUP captures both aleatoric and epistemic uncertainty on a probabilistic simplex to guide local training and global aggregation. It further uses global probabilistic prototypes as class-conditional priors for feature augmentation, alleviating client-side data sparsity. Finally, FedUP aggregates prototype distributions through reliability-informed barycenters on the statistical manifold, suppressing unreliable contributions while preserving geometric structure. Experiments on natural and medical benchmarks demonstrate that FedUP consistently outperforms state-of-the-art methods.

Hyperspectral imaging (HSI) provides detailed spectral information but is often impractical due to high cost, size, and hardware complexity. While learning-based methods attempt to recover hyperspectral images from RGB sensors, they are constrained by limited spectral measurements and noise. We propose a unified optical–computational approach that converts any standard RGB camera into a snapshot hyperspectral imaging system. Our method introduces a diffractive optical adapter that replaces the conventional lens with a diffractive lens array optimized for spectral encoding. To reconstruct hyperspectral images from the resulting measurements, we design a neural network specialized for diffractively encoded RGB data, capable of compensating for optical distortions and recovering high-quality spectra. The proposed system achieves up to 10 dB improvement in PSNR over RGB-based hyperspectral reconstruction and enhances the performance of existing state-of-theart models by up to 6 dB when used with the proposed adapter. Our results demonstrate that diffractive optical encoding combined with learned reconstruction enables practical and scalable hyperspectral imaging using commodity RGB cameras.

High dynamic range (HDR) imaging is crucial for capturing real-world lighting conditions. HDR imaging is traditionally achieved either by fusing multiple exposure frames or via inverse tone mapping from a single SDR image. However, the multi-exposure HDR method is prone to motion-induced artefacts and imposes demanding hardware requirements, limiting its practical applicability. Traditional inverse tone-mapping techniques primarily rely on pixel-wise regression methods, which ignore semantic scene contexts and thus frequently introduce halo artefacts and structural distortions. To address this limitation, this paper proposes MLLM-ITM, a novel inverse tone-mapping framework incorporating multimodal large language models (MLLMs). MLLM-ITM utilizes cross-modal features extracted from a frozen MLLM to simultaneously encode visual features and semantic understanding. These features are integrated into a downstream HDR reconstruction backbone through lightweight adapters, enabling content-aware dynamic-range expansion. Furthermore, the decoupled design between MLLM and HDR backbone avoids costly fine-tuning of MLLMs and remains model-agnostic, allowing effortless substitution with emerging multimodal architectures. Extensive experiments on public benchmarks demonstrate that the proposed MLLM-ITM achieves state-of-the-art performance compared with existing inverse tone-mapping methods, highlighting the effectiveness of cross-modal semantic priors in enhancing HDR imaging performance.

In this work, we introduce XSIM, a sensor simulation framework for autonomous driving. XSIM extends 3DGUT splatting with a generalized rolling-shutter modeling tailored for autonomous driving applications. Our framework provides a unified and flexible formulation for appearance and geometric sensor modeling, enabling rendering of complex sensor distortions in dynamic environments. We identify spherical cameras, such as LiDARs, as a critical edge case for existing 3DGUT splatting due to cyclic projection and time discontinuities at azimuth boundaries leading to incorrect particle projection. To address this issue, we propose a phase modeling mechanism that explicitly accounts temporal and shape discontinuities of Gaussians projected by the Unscented Transform at azimuth borders. In addition, we introduce an extended 3D Gaussian representation that incorporates two distinct opacity parameters to resolve mismatches between geometry and color distributions. As a result, our framework provides enhanced scene representations with improved geometric consistency and photorealistic appearance. We evaluate our framework extensively on multiple autonomous driving datasets, including Waymo Open Dataset, Argoverse 2, and PandaSet. Our framework consistently outperforms strong recent baselines and achieves state-of-the-art performance across all datasets.

FlashAttention improves efficiency through tiling, but its online softmax still relies on floating-point arithmetic for numerical stability, making full quantization difficult. We identify three main obstacles to integer-only FlashAttention: (1) scale explosion during tile-wise accumulation, (2) inefficient shift-based exponential operations on GPUs, and (3) quantization granularity constraints requiring uniform scales for integer comparison. To address these challenges, we propose QFlash, an end-to-end integer FlashAttention design that performs softmax entirely in the integer domain and runs as a single Triton kernel. On seven attention workloads from ViT, DeiT, and Swin models, QFlash achieves up to 6.73x speedup over I-ViT and up to 8.69x speedup on Swin, while reducing energy consumption by 18.8% compared to FP16 FlashAttention, without sacrificing Top-1 accuracy on ViT/DeiT and remaining competitive on Swin under per-tensor quantization. Our code is publicly available at https://github.com/EfficientCompLab/qflash.

Multimodal reasoning requires a path that retains integrity over a wide range of constraints, from visual grounding to logic consistency. However, the current Process Reward Models (PRMs) focus on heuristically defined rewards that equally weigh these factors, which may lead to the concealment of individual dimension failures (such as visual hallucinations) by the dominating factors, without guaranteeing the validity of the reasoning process in general. Therefore, to overcome the limitation, the paper proposes the concept of Multimodal Multi-Dimensional Scalarization Process Reward Modeling (MMS-PRM), a paradigm specifically developed to enforce the worst dimension’s robustness in multimodal reasoning. Specifically, a hierarchical fine-grained reward space is developed to represent the multimodal risks in the reasoning tasks, and a Chebyshev-based Monte Carlo Tree Search (MCTS) algorithm is introduced, in which the primary focus during the path searching is given to the worst-performing dimension. Moreover, a curriculum-based Direct Preference Optimization (DPO) approach is developed to gradually learn the balanced reasoning skills in the policy. The experimental results show that, without the dimension collapse issue, the MMS-PRM approach significantly improves the reliability of the multimodal reasoning performance and reaches competitive results in various challenging tasks. The code is available at https://github.com/leibniz-Man/MMS-PRM

Cross-modal hashing has been extensively adopted in cross-modal retrieval tasks owing to its high storage efficiency and fast retrieval capability. However, in practical applications, multimodal data often suffer from modality missing issues, which cause semantic incompleteness and thus severely impair both cross-modal alignment and hashing-based retrieval performance. To address these issues, this paper proposes a novel incomplete cross-modal retrieval framework, called retrieval-guided completion hashing with token–patch alignment (RGCH-TPA). Specifically, it first constructs a dynamic memory bank based on the features of complete samples and uses available modalities to retrieve and locate similar regions of incomplete samples. Thus, it achieves intra-modal reconstruction of missing data under the condition of available modalities using retrieved expert signals. Moreover, a token-patch level cross-modal alignment mechanism is introduced to model the interaction between text tokens and image patches, while global representations are jointly integrated to further enhance the discriminability and robustness of the learned cross-modal representations. Extensive evaluations on three benchmark datasets demonstrate that the proposed RGCH-TPA method consistently outperforms other state-of-the-art incomplete cross-modal hashing methods across missing-modality ratios. The source code is available at https://github.com/szq0816/RGCH-TPA.

Diffusion-based image compression has exhibited robust performance. However, most existing methods primarily emphasize spatial domain, causing frequency dependencies to be learned only implicitly. We revisit this problem from a frequency perspective, and observe pronounced heterogeneity between global and local frequency dependencies. Global frequency patterns describe the energy distribution over the entire image, whereas patch level frequency relations govern the coupling of local details. These two forms of dependency differ substantially in both scale and semantic level, yet most methods overlook this distinction. To address this issue, we propose a learning heterogeneous global local frequency dependencies in diffusion-based image compression which uses Fourier and Mamba jointly models both global and local frequency correlations(FMDiff). The core of FMDiff is the dual branch FMBlock. In the frequency branch, features are decomposed into magnitude and phase, then fed into Mamba separately. Its scanning mechanism captures cross frequency coupling within each patch while aggregating long range frequency context along the sequence. Magnitude phase interactive modulation is then used to explicitly restore spectral information corrupted by compression. The spatial branch provides high level semantic constraints and is fused with the frequency branch during diffusion reconstruction. Extensive experiments show that FMDiff consistently improves performance across multiple datasets.

Spatiotemporal fusion (STF) bridges the gap between temporal and spatial resolutions in satellite imagery, enabling effective monitoring of Earth's surface dynamics. However, existing methods rely on cloud-free reference images, a constraint that fails in realistic, cloud-prone scenarios. To overcome this, we propose the Cloud-Aware Wavelet Generative Adversarial Network (CLAW-GAN), a novel framework for high-fidelity reconstruction under cloud-contaminated conditions. CLAW-GAN introduces Regression Wavelet Analysis (RWA) to decouple spectral backgrounds from structural details. While a change-aware gated mechanism accounts for land-cover changes, the Frequency-Separated Fusion (FSF) module then independently integrates these components. To ensure visual realism, a multi-scale discriminator operates in the wavelet domain, enforcing consistency across high-frequency subbands to minimize artifacts. Evaluated on the newly introduced Global Cloud-shrouded Agricultural Regions (GCAR) benchmark and the simulated Daxing dataset, CLAW-GAN achieves state-of-the-art performance and demonstrates superior robustness across varying cloud coverage.

Current remote sensing (RS) perception systems suffer from task heterogeneity, necessitating distinct architectures for classification, localization, and reasoning. While vision--language models (VLMs) offer a route toward unification, their computational cost can hinder deployment. In this work, we propose RS-Florence, a compact unified model that addresses these tasks through a Prompt-Driven Sequence-to-Sequence framework. Unlike traditional approaches that segregate semantic understanding and geometric localization, RS-Florence maps images and task-specific prompts into a unified sequence of natural language and discrete geometric tokens. This formulation helps bridge high-level semantics and low-level pixel perception. Experiments across 4 task families and 8 benchmarks show that our 0.23B model remains competitive with task-specific specialists. These results also suggest that multitask joint training improves performance on several benchmarks over single-task fine-tuning, indicating that a shared prompt-driven interface can serve both language and geometry tasks.

The Segment Anything Model (SAM) provides a new paradigm for visual perception tasks through large-scale pre-training and interactive prompts. However, the quality of mask generation is limited by the difficulty of aligning the optimization direction between the backbone and the prompt, which is especially significant in non-natural domain images with low contrast and indistinct semantic edge structure features, especially in SAR images. Therefore, this paper proposes a coupling optimization mechanism that integrates the originally independent prompt generation process with the backbone into a closed loop through the Point Impact Decomposition (PID) module to guide the iterative optimization of the prompt. Specifically, this paper introduces the concept of PID in the decoder to quantify the contribution of prompt points to the masks, constructing an interpretable reward variable to drive the reinforcement learning prompt optimizer. This optimizer designs a reward function based on hypothesis testing ideas, achieving iterative updates of prompt points, and providing feedback on the effects of mask generation based on prompt points. Experiments show that this mechanism can provide a prompt that fits image features; its performance exceeds the existing baseline, which effectively improves the generalization ability of SAM on SAR images.Code is available at https://github.com/Fm336/PID-SAM.

Adapting large, adversarially pre-trained models to specialized domains via transfer learning is a promising path toward building secure AI systems. However, a critical challenge arises when fine-tuning on limited downstream data: models often suffer from catastrophic forgetting of robustness, where the generalizable robust features from pre-training are erased, leading to severe robust overfitting. To address this, we first establish the underlying principle that adversarial vulnerability is not diffuse, but is concentrated in a low-dimensional subspace defined by the principal components of the model's weight matrices. We then introduce Principal component-based Implicit regularization with Low-rank Optimization (PILO), a novel parameter-efficient framework that operationalizes this insight. PILO employs a principled decoupling of learning objectives: a spectrally-guided adversarial branch, initialized via Singular Value Decomposition (SVD), performs a surgical update on the identified vulnerable subspace to enhance robustness. This is complemented by a transient, standard low-rank branch that gently adapts the model to natural data, preserving clean accuracy. This dual-branch design acts as a form of implicit spectral regularization, anchoring the model to its robust pre-trained state. Extensive experiments show that PILO significantly outperforms state-of-the-art full-parameter and parameter-efficient methods in robust accuracy across multiple benchmarks, all while reducing trainable parameters by up to 71%. Our work thus establishes a new, more effective paradigm for robust transfer learning through principled and targeted parameter optimization.

Although Large Vision-Language Models (LVLMs) have demonstrated remarkable reasoning capabilities across various downstream multimodal tasks, they are proven to be vulnerable to carefully designed adversarial examples. Existing LVLM attackers show that exploring external components of adversarial guidance (e.g., forcing adversarial alignment, resembling harmful features) can help improve adversarial effects. However, leveraging the intrinsic patterns of LVLMs to induce adversarial perturbation generation by exploring how LVLMs perceive images has not been deeply studied. Inspired by the cognitive science, in this paper, we make the first attempt to investigate the interference of adversarial perturbation from the perspectives of image phase, and find that LVLMs are sensitive to the phase-aware image structure. Motivated by this, we propose a novel LVLM attack method called BadPhase with further backdoor designs, to implant adversarial phase as triggers into any image inputs via data poisoning so as to control the LVLMs’ predictions. A textual trigger and a backdoor perturbation switcher are also introduced to activate the malicious behavior only when both triggers are present. The whole backdoor optimization is implemented at the test-time to reduce the resource reliance. Experiments on four popular LVLMs and three benchmarks demonstrate the effectiveness of our proposed method.