IJCAI 2026 论文列表

Recently, transformer-based trackers have become the leading approach, surpassing traditional CNN-based trackers in accuracy. However, their high computational demands hinder their deployment on edge platforms, necessitating efficient solutions. To address this, we propose a novel Neural Architecture Search (NAS) framework designed for transformer-based trackers, named NASTrack. This framework incorporates token pruning to optimize both transformer block structures and the layer-wise token keeping ratio, striking a balance between performance and efficiency. To handle the larger search space introduced by the keeping ratio, we propose a blacklist strategy and a matching-based distillation driven by the Token Overlap Ratio (TOR). Our method discovers hundreds of high-performing trackers, with FLOPs ranging from 1G to 18G. The searched trackers consistently outperform existing efficient state-of-the-art trackers such as CompressTracker and LiteTrack under comparable computational budgets.The code and models are available at https://github.com/Cyhoon84/NASTrack.git.

Few-shot medical image segmentation relies on dense, boundary-sensitive prototype matching, yet common pre-training objectives mainly optimize global alignment or reconstruction, creating an objective gap that hurts boundary delineation and increases adaptation cost. This raises the question: how to pre-train representations intrinsically matchable for episodic FSS while requiring minimal adaptation-induced re-organization? We introduce a Drift-Gap diagnostic to quantify intrinsic dense-matching misalignment and adaptation-induced feature drift. Guided by this lens, we propose BOG-PRETRAIN, combining Reliability-Gated Alignment to mitigate noisy report supervision, Semantic-Guided MIM to emphasize boundary-informative regions, and Dual Consistency Regularization to stabilize episodic metric geometry. Across five benchmarks (1/4/16-shot), BOG-PRETRAIN improves mean Dice by +15.5/+15.9/+12.5 points over best priors and reduces mean HD95 by 0.9/4.6/10.1; it achieves the lowest Drift (0.052 vs. 0.155 baseline) and Gap_PT (0.256), with ablations confirming the components' complementarity.

Large Vision-Language Models (LVLMs) often suffer from hallucinations, generating descriptions that include visual details absent from the input image. Recent preference alignment methods typically rely on supervision distilled from stronger models such as GPT. However, this offline paradigm introduces a Supervision-Perception Mismatch: the student model is forced to align with fine-grained details beyond its perceptual capacity, learning to guess rather than to see. To obtain reliable self-supervision for online learning, we identify a Generative-Discriminative Gap within LVLMs, where models exhibit higher accuracy on discriminative verification than open-ended generation. Leveraging this capability, we propose Online Self-CAlibRation (OSCAR), a framework that integrates Monte Carlo Tree Search with a Dual-Granularity Reward Mechanism to construct preference data and iteratively refines the model via Direct Preference Optimization. Extensive experiments demonstrate that OSCAR achieves state-of-the-art performance on hallucination benchmarks while preserving general multimodal capabilities.

Designing effective reward functions remains a major challenge in reinforcement learning (RL), particularly in open-ended environments where task goals are abstract and difficult to quantify. In this work, we present VLM-AR3L, a framework that leverages Vision-Language Models (VLMs) to provide both absolute and relative rewards for RL. VLM-AR3L interprets an agent’s visual observations in the context of a natural language task goal, and learns both absolute and relative rewards from VLM-generated preference labels. The absolute reward model predicts scalar evaluations for individual states, while the relative reward model compares consecutive observations to infer progress or regression toward the task goal. Their integration combines the stability of state-based evaluation with the robustness of comparative supervision. We evaluate VLM-AR3L across benchmarks spanning classic control, manipulation, and open-world embodied tasks, with a particular focus on Minecraft given its visual complexity and long-horizon decision-making requirements. Experimental results show that VLM-AR3L consistently outperforms prior VLM-based reward learning methods. Videos and code are available on the project website: https://vlm-ar3l.github.io/.

Camouflaged Object Detection (COD) aims to segment objects that are hidden within complex backgrounds. Due to the low visual contrast of camouflaged objects, annotations are costly, motivating unsupervised COD (UCOD) to eliminate labeling expenses. Most UCOD methods follow the “MLLMs + other foundation models + SAM” paradigm, which relies on spatial interactions that are unreliable in camouflaged scenarios, leading to fundamental performance bottlenecks. In this paper, we propose a novel UCOD framework that leverages SAM3 through category-level interaction with MLLMs, bypassing unreliable spatial prompts. To address SAM3’s sensitivity to category granularity, we introduce Fine-grained Category Query, guiding MLLMs to generate full-granularity category chains for robust category prompting. To mitigate suboptimal segmentation caused by high camouflage and background confusion, we propose Semantic–Geometric Dual Confirmation, which jointly validates segmentation masks from semantic and spatial perspectives. Furthermore, we introduce Semantic–Geometric Reasoning Injection, which injects critical semantic and geometric cues into MLLMs to refine category reasoning and progressively correct segmentation errors under extreme camouflage or MLLM hallucinations. Extensive experiments show that our method significantly outperforms existing UCOD approaches and achieves performance comparable to weakly supervised COD.

Multimodal sentiment analysis relies on textual, acoustic, and visual signals, yet real-world data often suffer from modality missing and quality imbalance. Existing methods generate features for modality missing from available ones, but differences in expression mechanisms and sentiment dynamics across modalities may cause the generated features to deviate from true distributions and mislead prediction. In addition, unreliable modalities may dominate fusion, resulting in representation shift across modality combinations and unstable sentiment representations. To address these challenges, we propose a two-level reference alignment framework. The framework introduces stable references at the feature representation and sentiment decision levels to improve robustness under modality missing. First-level reference alignment leverages complete-modality samples to constrain representations and align different modality combinations into a shared sentiment space. Second-level reference alignment enforces cross-modal consistency at the decision level by suppressing unreliable modalities through prototype retrieval and voting. As a result, the framework maintains stable and reliable sentiment predictions under diverse missing-modality patterns. Experiments on CMU-MOSI and CMU-MOSEI show consistent improvements across various missing-modality settings. Under full-modality input, the proposed method achieves state-of-the-art performance, with ACC of 86.28% and 85.88%, and F1 of 86.24% and 85.86%.

Video captioning models convert frames into visual tokens and generate descriptions with large language models (LLMs). Since encoding all frames is prohibitively expensive, uniform sampling is the default choice, but it enforces equal temporal coverage while ignoring the uneven events distribution. This motivates a Learnable Frame Selector (LFS) that selects temporally diverse and event-relevant frames. LFS explicitly models temporal importance to balance temporal diversity and event relevance, and employs a stratified strategy to ensure temporal coverage while avoiding clustering. Crucially, LFS leverages caption feedback from frozen video-LLMs to learn frame selection that directly optimizes downstream caption quality. Additionally, we identify the gap between existing benchmark and human's cognition. Thus, we introduce ICH-CC built from carefully designed questions by annotators that reflect human-consistent understanding of video. Experiments indicate that LFS consistently improves detailed video captioning across two representative community benchmarks and ICH-CC, achieving up to 2.0% gains on VDC and over 4% gains on ICH-CC. Moreover, we observe that enhanced captions with LFS leads to improved performance on video question answering. Overall, LFS provides an effective and easy-to-integrate solution for detailed video captioning.

Real-world systems increasingly require coherent reasoning and generation over diverse data modalities simultaneously. Current generative frameworks rely on complex, multi-stage training, resulting in low efficiency due to iterative inference and high computational cost. They also struggle with unified multimodal representation, failing to balance fine-grained details with global structures and long-term dependencies, which limits generative quality and practical usability. To overcome these limitations, we introduce the Inverse Heat Mean Flow (IHMF), a general-purpose solver that is compatible with a wide range of model backbones. Without requiring pre-training, IHMF directly learns an average velocity field through an inverse heat formul ation. By exploiting the inherent scale-space properties of the inverse heat process, IHMF explicitly decomposes multi-scale complexities, thereby simplifying trajectory learning and enabling adaptive topological alignment. Extensive experiments show that IHMF achieves a competitive performance with state-of-the-art methods, providing a robust and unified mathematical framework for various generative tasks. The code is publicly available at https://github.com/CHBonline/IHMF.

Supervised infrared-visible image fusion (IVIF) often overfits limited training distributions, creating a critical generalization gap under open-world degradations (rain, haze, low light, noise, blur). To address this issue, we propose AIR-Fusion, a parameter-efficient adaptation of a frozen, restoration-capable latent diffusion backbone for degraded IVIF without full fine-tuning, transferring restoration priors for robust fusion. A Cross-Modal Bridging Adapter (CMBA) aligns infrared cues and textual instructions with the frozen diffusion conditioning space and injects them into multi-scale denoising features to steer instruction-guided restoration-aware fusion. In addition, a Trajectory-Constrained Rectifier (TCR) regularizes stochastic sampling via a pixel-latent closed loop, rectifying intermediate predictions with source-referenced structures and re-encoding them to stabilize the denoising trajectory and recover fine details suppressed by latent compression. Experiments across multiple datasets and degradation settings show consistent improvements in restoration quality and fusion fidelity, with strong generalization under complex and compounded degradations.

Visual grounding aims to localize target objects based on natural language descriptions, and the core challenge lies in the cross-modal gap, which is partly caused by the significant differences in semantic structure between language and vision. Existing methods typically rely on holistic sentence-level semantic representations to modulate visual features, while overlooking the inherent structure of textual prompts. In this work, we propose a Syntactic Structure-guided Visual Grounding framework, referred to as SSVG. Specifically, to inject syntactic priors into unstructured visual representations, we design a semantic structure-based feature refinement module to adaptively modulate subject-centric and contextual visual features. To perform cross-modal alignment, we further incorporate a visual semantic consistency verification module, which leverages a subject-aware contrastive learning strategy to constrain and verify the semantic correspondence between the visual prediction and subject-level textual representation, thereby enhancing the model's robustness against semantically similar distractors. Extensive experiments demonstrate that our approach consistently outperforms state-of-the-art methods, and detailed analyses verify the effectiveness of each component.

In this paper, we revisit multimodal few-shot 3D point cloud semantic segmentation (FS-PCS), identifying a conflict in "Fuse-then-Refine" paradigms: the "Plasticity-Stability Dilemma." In addition, Contrastive Language-Image Pre-training (CLIP)'s inter-class confusion can result in semantic blindness. To address these issues, we present the Decoupled-experts Arbitration Few-Shot SegNet (DA-FSS), a model that effectively distinguishes between semantic and geometric paths and mutually regularizes their gradients to achieve better generalization. DA-FSS employs the same backbone and pre-trained text encoder as the baseline MultiModal Few-Shot SegNet (MM-FSS) to generate text embeddings, which can increase cost-free modalities' utilization rate and better leverage each modality's information space. To achieve this, we propose a Parallel Expert Refinement module to generate each modal correlation. We also propose a Stacked Arbitration Module (SAM) to perform convolutional fusion and arbitrate correlations for each modality pathway. The Parallel Experts decouple two paths: a Geometric Expert maintains plasticity, and a Semantic Expert ensures stability. They are coordinated via a Decoupled Alignment Module (DAM) that transfers knowledge without propagating confusion. Experiments on popular datasets (S3DIS, ScanNet) demonstrate the superiority of DA-FSS over MM-FSS. Meanwhile, geometric boundaries, completeness, and texture differentiation are all superior to the baseline. The code is available at: https://github.com/MoWenQAQ/DA-FSS/.

Diffusion models show significant potential for low-light image enhancement. However, this task requires satisfying human perceptual preferences and content fidelity transcending simple brightness and color improvement. Existing methods rely on heuristic physical priors or incorporate perceptual metrics directly into timestep-wise training objectives. Such proxy constraints fail to provide reliable trajectory-level guidance for perceptual alignment. Furthermore, they often lead to artifacts or unnatural visual effects in this ill-posed inverse problem. To address these issues, we propose an unsupervised low-light enhancement framework based on Group Relative Policy Optimization (GRPO), which utilizes perceptual preferences to directly optimize the diffusion policy. We introduce a Sliding Window Hybrid ODE-SDE Sampling strategy that confines stochasticity to dynamic sub-intervals, thereby achieving efficient coarse-to-fine exploration and precise advantage attribution. To meet strict fidelity constraints, we construct a Synergistic Perception-Fidelity Reward and introduce an Independent Advantage Estimation strategy to mitigate signal collapse and gradient suppression in multi-objective optimization. Furthermore, we design a Temporal-Aware Dynamic Weighting Mechanism that adaptively adjusts perceptual weights across denoising stages to balance visual enhancement with structural preservation. Extensive experiments on multiple real-world benchmarks demonstrate that our method effectively improves the perceptual performance of existing models, yielding results that better align with human aesthetics.

Test-Time Adaptation (TTA) aims to adapt pretrained models to unseen test data, which is crucial for resource-constrained edge devices that must handle distribution shifts on the fly without human supervision. However, conventional TTA methods often fail in realistic scenarios characterized by continuous shifts and severe class imbalances while incurring prohibitive memory overheads. To enable continuous adaptation to test data on edge devices under a realistic TTA setting, we propose a Bias-Resistant Online Statistical Alignment (BOSA) method. BOSA facilitates unbiased adaptation via a discrepancy-aware statistical alignment mechanism integrated with a class-balanced memory bank. To ensure memory efficiency, we further design a saliency-guided activation sparsification and gradient reconstruction scheme, which drastically reduces memory overhead without sacrificing gradient integrity. Extensive evaluations demonstrate that BOSA achieves superior accuracy with a compact memory consumption compared to state-of-the-art methods under realistic TTA settings.

The rapid progress of identity-feature-based face-swapping technology has raised concerns about impersonation and privacy violations. Although proactive defenses aim to block identity extraction at the source, existing methods suffer from perceptible visual artifacts, poor generalization across diverse deepfake models, and vulnerability to post-processing techniques (e.g., diffusion purification, image compression, and transformations). This work proposes a robust, generalizable proactive face-swapping defense via semantic gradient divergence (SGD-Guard) to address these challenges. It introduces an integrated feature gallery that uses CLIP features and a generalized identity feature, obtained by iteratively refining heterogeneous identity features into a homogeneous representation. This framework facilitates our semantic distortion attack by leveraging consensus weighting to target specific facial attributes within a CLIP-identity joint embedding space, disrupting deepfake generation while preserving visual fidelity. Furthermore, to ensure robustness against purification and post-processing, this method incorporates a module that prioritizes critical transformations by exploiting directional discrepancies. Comprehensive experiments demonstrate that the method effectively defends against diverse face-swapping models with high cross-model transferability.

Safety alignment of large language models (LLMs) has been gaining increasing attention. However, current safety-aligned LLMs suffer from the fragile and imbalanced safety mechanisms, which can still be induced to generate unsafe responses, exhibit over-safety by rejecting safe user inputs, and fail to preserve general utility after safety alignment. To this end, we propose a novel post safety alignment (PSA) method to address these inherent and emerging safety challenges, including safety enhancement, over-safety mitigation, and utility preservation. In specific, we introduce SafePatching, a novel framework for comprehensive PSA, where two distinct safety patches are developed on the harmful data to enhance safety and mitigate over-safety concerns, and then seamlessly integrated into the target LLM backbone without compromising its utility. Extensive experiments on four representative aligned LLMs, including LLaMA-2/3, Gemma and Mistral, show that SafePatching achieves a more comprehensive PSA than baseline methods, further optimizing the balance between being helpful and harmless in current aligned LLMs. Also, SafePatching demonstrates its superiority in continual PSA scenarios.

Large Language Model (LLM)-based navigation systems have commonly constructed expli cit spatial representations (e.g., topological graphs, semantic raster maps) and translated them into textual descriptions as LLMs’ inputs. However, the linguistic structures of such text-based spatial representations and the choices of contextual features (e.g., topology, geometry) they contain are often treated as neutral engineering decisions rather than key factors that shape LLMs' behavior. To address this gap, we propose a dual-interventional framework that disentangles linguistic structures from different contextual cues to evaluate the linguistic inductive bias of LLMs for navigation planning. In the framework, representation intervention varies linguistic format and the degree of linguistic compression, clarifying when linguistic representations support or inhibit navigation planning; context intervention, combined with contextual feature combination and conflict probing, explicitly clarifies the preferences and weaknesses of LLMs when processing different contextual cues. Experiments across diverse spatial reasoning tasks and multiple model scales reveal a consistent pattern: topological information is a sturdy shield and the backbone of robust planning; linguistic format is a double-edged sword whose effect depends on model size, task demands, and the compression level; and semantic information is a fatal Achilles' heel---incorrect semantic cues can systematically derail the planning process. Overall, our study shows that effective text-based spatial representations in LLM-based navigation should preserve topological integrity, calibrate representational compression to model capacity, and ensure semantic correctness, rather than simply adopting a single representation.

Achieving fairness in machine learning models while maintaining high accuracy is an important but complex task, especially when handling multiple sensitive attributes. Traditional fairness methods often struggle to eliminate bias within subgroups divided by sensitive attributes. Several key challenges have been identified in this context: (1) Multiple sensitive attribute scalability challenge, where methods fail to ensure fairness as the number of sensitive attributes increases, despite scenarios with multiple sensitive attributes being prevalent in real-world applications; (2) Multiple objective optimization conflict challenge, where simultaneously optimizing for accuracy, fairness, and other relevant objectives leads to conflicting gradient updates, causing suboptimal performance. To address these challenges, we propose BAMFair, a Barycenter Aligned Mediation framework for fairness across multiple sensitive attributes. It comprises two core modules: a Global Barycentric Alignment (GBA) module and a Nash Fairness Mediator (NFM) module. Specifically, GBA innovatively introduces a global fair barycenter and minimizes the distances from subgroups divided by sensitive attributes to it, providing a scalable and efficient solution for fairness optimization across multiple sensitive attributes. Subsequently, NFM negotiates an agreement among inconsistent gradient updates between different objectives. Extensive experiments on four real-world datasets validate that BAMFair outperforms state-of-the-art methods in scenarios with multiple sensitive attributes.

Text-to-Image (T2I) diffusion models have been widely adopted due to their strong generative capabilities, while Low-Rank Adaptation (LoRA) has emerged as an efficient mechanism for customizing these models for diverse creative and commercial applications. This trend has fostered LoRA-centric service platforms that that enable the customization and commercial distribution of LoRA modules according to user requirements. However, the growing prevalence of LoRA and its critical role in customized AI services have raised urgent concerns about LoRA copyright protection. To address this gap, we propose LoRA2D, an authorized dual-watermarking framework specifically designed to protect LoRA modules in T2I diffusion models. LoRA2D integrates license-based authorization control with explicit watermarks as visible deterrents for unauthorized or trial usage, which can be removed upon valid authorization, while persistently embedding an implicit watermark for robust black-box ownership verification. Extensive experiments on multiple image-generation datasets demonstrate the effectiveness and practicality of LoRA2D for securing copyrights in LoRA-adapted T2I diffusion models.

Large Language Models (LLMs) achieve strong performance in many applications but remain limited in handling graph-structured data due to their reliance on textual context. Recent approaches integrate Graph Neural Networks (GNNs) to enhance structural modeling, yet they largely overlook fairness, leaving models vulnerable to bias amplification across graph and text modalities. To address this issue, we propose FairGEnt, a disentangled graph-enhanced large language model for fair graph learning. FairGEnt separates sensitive-related and sensitive-invariant factors in both graph and textual representations to mitigate bias while preserving task-relevant information, and further aligns the two modalities through a fairness-aware integration module. In addition, FairGEnt incorporates fair graph-enhanced instruction tuning to improve LLM understanding of complex graph structures. Experiments on multiple benchmark datasets demonstrate that FairGEnt consistently outperforms existing methods in both fairness and predictive performance.