Model merging integrates multiple task-specific models into a single consolidated one. Recent research has made progress in improving merging performance for in-distribution or multi-task scenarios, but domain generalization in model merging remains underexplored. We investigate how merging models fine-tuned on distinct domains affects generalization to unseen domains. Through an analysis of parameter competition in the task matrix using singular value decomposition, we show that merging models trained under different distribution shifts induces stronger conflicts between their subspaces compared to traditional multi-task settings. To mitigate this issue, we propose SCORE (Subspace COnflict-Resolving mErging), a method designed to alleviate such singular subspace conflicts. SCORE finds a shared orthogonal basis by computing the principal components of the concatenated leading singular vectors of all models. It then projects each task matrix into the shared basis, pruning off-diagonal components to remove conflicting singular directions. SCORE consistently outperforms, on average, existing model merging approaches in domain generalization settings across a variety of architectures and model scales, demonstrating its effectiveness and scalability.
论文检索
输入标题、作者或关键词，从 9,654 篇学术成果中精准定位
ModularAgent: A Task-Aware Modular Framework for Joint Optimization of Multimodal Large Language Models and World Models
PDF ↗Building generalist embodied agents requires a unified system that can interpret multimodal goals, model environment dynamics, and execute reliable actions across diverse real-world tasks. Multimodal large language models (MLLMs) offer strong semantic priors and cross-modal generalization, while world models (WMs) provide actionable latent dynamics for prediction and control. Their combination holds promise for open-ended embodied intelligence, yet introduces two key challenges: (1) establishing a tight coupling between the semantic intent from MLLMs and the dynamic state representations within the WM's latent space, and (2) achieving task-aware adaptability that supports multi-task learning and cross-environment generalization. To address these limitations, we propose ModularAgent, a task-aware dynamic joint framework that enables bidirectional coupling between MLLMs and WMs. ModularAgent establishes two complementary pathways: a forward path that injects MLLM representations into the WMs latent space for semantically guided imagination, and a backward path where WM-generated feedback refines the MLLMs semantic space via dense text-conditioned rewards. This bidirectional interaction is realized through three synergistic components: Task-Aware Dynamic Joint Learning, Task-Aware Behavior Learning, and MLLM-WM Joint Optimization, which together harmonize semantic reasoning and dynamic prediction. Extensive experiments across multi-task and cross-environment settings demonstrate superior stability and generalization over state-of-the-art baselines, marking a step toward open-ended embodied learning.
Robust 3D semantic occupancy is essential for legged and humanoid robots, yet most Semantic Scene Completion (SSC) systems are built for wheeled platforms with forward-facing sensors. We present OneOcc, a vision-only panoramic SSC framework tailored to severe body jitter and 360deg continuity. OneOcc integrates four complementary modules: (i) Dual-Projection fusion (DP-ER), which jointly exploits the raw annular panorama and its equirectangular unfolding to preserve true 360deg continuity while enabling grid-aligned feature extraction and seam-aware context; (ii) Bi-Grid Voxelization (BGV), which reasons in Cartesian and polar/cylindrical voxel spaces to reduce discretization bias and better align with panoramic geometry, yielding sharper free/occupied boundaries; (iii) a lightweight decoder with Hierarchical AMoE-3D fusion that dynamically routes multi-scale 3D features to specialized experts, improving long-range context and occlusion handling; and (iv) a plug-and-play Gait Displacement Compensation (GDC) module that learns feature-level motion correction from gait, stabilizing representations without extra sensors. We also release two panoramic occupancy benchmarks: QuadOcc (real quadruped, first-person 360deg) and Human360Occ (H3O) (CARLA human-ego 360deg with RGB/Depth/semantic-occupancy and standardized within-/cross-city splits). OneOcc sets new SOTA: on QuadOcc it exceeds strong vision baselines and even popular LiDAR methods, and on H3O it improves within-city by +3.83 mIoU and cross-city by +8.08. The modules are lightweight, enabling deployable full-surround semantic perception for legged and humanoid robots.
Modern LiDARs are rapidly transitioning from bulky, mechanically scanned systems to ultra-compact, low-cost, solid-state arrays. This miniaturization--while enabling scalability, affordability, and camera-like data structures--introduces a new and severe failure mode: internal-multipath glare. When light from a bright or retroreflective surface reflects and scatters within the LiDAR, light that should reach a single pixel spreads across the pixel array. The resulting artifacts create phantom objects, obscure real ones, and produce safety-critical "ghosts in the point clouds." This paper introduces a physically grounded sensing model and algorithmic techniques for addressing this effect. We show that internal glare can be represented as a linear, scene-independent operator--the Transient Glare Spread Function (TGSF)--acting on the transient measurements. Building on this model, we develop a training-free approach that operates on low-level LiDAR detections (or echoes) prior to point-cloud formation, leveraging knowledge of the glare spread function to reason about the likelihood of each detection arising from glare. The resulting approach is compatible with existing LiDAR signal-processing pipelines, and deployable on unmodified commercial sensors. Using experiments with real single-photon LiDAR hardware, we demonstrate substantial suppression of severe glare artifacts while preserving true scene structure.
CC-VQA: Conflict- and Correlation-Aware Method for Mitigating Knowledge Conflict in Knowledge-Based Visual Question Answering
PDF ↗Knowledge-based visual question answering (KB-VQA) demonstrates significant potential for handling knowledge-intensive tasks. However, conflicts arise between static parametric knowledge in vision language models (VLMs) and dynamically retrieved information due to the static model knowledge from pre-training. The outputs either ignore retrieved contexts or exhibit inconsistent integration with parametric knowledge, posing substantial challenges for KB-VQA. Current knowledge conflict mitigation methods primarily adapted from language-based approaches, focusing on context-level conflicts through engineered prompting strategies or context-aware decoding mechanisms. However, these methods neglect the critical role of visual information in conflicts and suffer from redundant retrieved contexts, which impair accurate conflict identification and effective mitigation. To address these limitations, we propose CC-VQA: a novel training-free, conflict- and correlation-aware method for KB-VQA. Our method comprises two core components: (1) Vision-Centric Contextual Conflict Reasoning, which performs visual-semantic conflict analysis across internal and external knowledge contexts; and (2) Correlation-Guided Encoding and Decoding, featuring positional encoding compression for low-correlation statements and adaptive decoding using correlation-weighted conflict scoring. Extensive evaluations on E-VQA, InfoSeek, and OK-VQA benchmarks demonstrate that CC-VQA achieves state-of-the-art performance, yielding absolute accuracy improvements of 3.3% to 6.4% compared to existing methods. Code will be released upon acceptance.
Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction
PDF ↗With the rapid advancement and widespread application of vision-language pre-training (VLP) models, their vulnerability to adversarial attacks has become a critical concern. In general, the adversarial examples can typically be designed to exhibit transferable power, attacking not only different models but also across diverse tasks. However, existing attacks on language-vision models mainly rely on static cross-modal interactions and focus solely on disrupting positive image-text pairs, resulting in limited cross-modal disruption and poor transferability. To address this issue, we propose a Semantic-Augmented Dynamic Contrastive Attack (SADCA) that enhances adversarial transferability through progressive and semantically guided perturbation. SADCA progressively disrupts cross-modal alignment through dynamic interactions between adversarial images and texts. This is accomplished by SADCA establishing a contrastive learning mechanism involving adversarial, positive and negative samples, to reinforce the semantic inconsistency of the obtained perturbations. Moreover, we empirically find that input transformations commonly used in traditional transfer-based attacks also benefit VLPs, which motivates a semantic augmentation module that increases the diversity and generalization of adversarial examples. Extensive experiments on multiple datasets and models demonstrate that SADCA significantly improves adversarial transferability and consistently surpasses state-of-the-art methods. The code is released at https://github.com/LiYuanBoJNU/SADCA.
Few-shot learning has attracted extensive attention, with metric-based approaches such as Prototypical Networks establishing strong baselines. These methods construct class prototypes from support samples and classify query samples via distance metrics, but their performance is highly sensitive to label noise. To tackle this challenge, we propose a novel Graph Attention Prototypical Network (GAPNet) for robust few-shot classification. GAPNet first extracts local and global features via a classic CNN backbone and a group attention broad learning module, respectively. To mitigate the impact of label noise, the intra-class and inter-class relationships between support and query samples are explicitly modeled via a pseudo-label guided graph constructor, and then processed by an edge-aware graph attention module to capture topological correlations. Furthermore, an adaptive noise-robust prototype generator is introduced to dynamically suppress the contributions of noisy samples, substantially improving the reliability of class prototypes. Extensive experiments demonstrate the effectiveness and robustness of GAPNet to label noise. Compared to state-of-the-art approaches, GAPNet improves accuracy in the 5-way 5-shot setting by 3% ~ 8% on three general image benchmarks and one fine-grained classification dataset.
Cross-modal Identity Mapping: Minimizing Information Loss in Modality Conversion via Reinforcement Learning
PDF ↗Large Vision-Language Models (LVLMs) often omit or misrepresent critical visual content in generated image captions. Minimizing such information loss will force LVLMs to focus on image details to generate precise descriptions. However, measuring information loss during modality conversion is inherently challenging due to the modal gap between visual content and text output. In this paper, we argue that the quality of an image caption is positively correlated with the similarity between images retrieved via text search using that caption. Based on this insight, we further propose Cross-modal Identity Mapping (CIM), a reinforcement learning framework that enhances image captioning without requiring additional annotations. Specifically, the method quantitatively evaluates the information loss from two perspectives: Gallery Representation Consistency and Query-gallery Image Relevance. Supervised under these metrics, LVLM minimizes information loss and aims to achieve identity mapping from images to captions. The experimental results demonstrate the superior performance of our method in image captioning, even when compared with Supervised Fine-Tuning. Particularly, on the COCO-LN500 benchmark, CIM achieves a 20% improvement in relation reasoning on Qwen2.5-VL-7B. The codes are released on GitHub.
TSTM: Temporal Segmentation for Task-relevant Mask in Visual Reinforcement Learning Generalization
PDF ↗Achieving strong policy generalization to unseen environments remains a core challenge in visual reinforcement learning, and segmenting task-relevant regions to mitigate the influence of irrelevant visual cues has emerged as a promising direction. However, existing methods rely solely on the current observation, lack temporal information, and fail to exploit preceding observations, leaving learned policies susceptible to task-irrelevant background variations and ultimately degrading policy performance. In this paper, we propose temporal segmentation for task-relevant mask in visual reinforcement learning, named TSTM, which extracts task-relevant regions from sequential observations by exploiting temporal information, thereby producing more reliable masks and improving policy generalization. TSTM introduces a temporal segmentation network with an encoder-temporal-decoder architecture, where a convolutional LSTM module captures temporal dependencies across observations. To reduce inference overhead, we further develop a lightweight student network as an efficient substitute for the teacher network. The resulting task-relevant masks are encoded by a CNN-based encoder, and invariant representation learning is employed to improve robustness by enforcing consistency between representations of the original and augmented observation sequences. With these task-relevant representations, we train an actor-critic agent to learn a policy with strong generalization capability. Experimental results demonstrate that TSTM achieves superior generalization performance over existing state-of-the-art methods on most visual RL tasks.
Vision-Language Models (VLMs) have enhanced traditional LLMs with visual capabilities through the integration of vision encoders. While recent works have explored various combinations of vision encoders and LLMs, there still lacks a principled understanding of what makes a vision encoder suitable for VLM alignment. In this paper, we systematically investigate this question via comprehensive experiments on a curated collection of 19 pre-trained vision encoders from diverse sources. We first demonstrate that common practices, such as choosing encoders with the largest size or highest zero-shot accuracy, consistently fail to identify optimal models. In fact, these metrics show only weak to moderate correlation with VLM performance. This intriguing finding begs a fundamental question: What factors of vision-encoders matter in VLM? Through comprehensive analysis, we identify that the structural similarity across modalities plays a crucial but previously overlooked role in vision-encoder selection, which we measure using the Gromov-Wasserstein distance as a proxy. From a theoretical perspective, we show that the learnability of cross-modality mapping can be provably associated with the Gromov-Wasserstein distance. Empirical verification on 60+ full VLM training runs shows that our proposed inference-only metric performs significantly better than alternative model selection strategies and exhibits a much stronger correlation with final VLM performance, thereby enabling efficient and effective prediction of VLM performance before full training.
From Pixel to Precision: Enhancing Handwritten Mathematical Expression Recognition with Image-Level Reward
PDF ↗Handwritten mathematical expression recognition is hindered by a fundamental misalignment between the dual representations of LaTeX formulas: the symbolic text and the rendered visual image. This discrepancy means that textually distinct LaTeX sequences can produce visually identical outputs, while minor textual errors can cause catastrophic rendering failures. As a result, text-level reward mechanisms cannot perfectly assess the quality of model predictions, failing to effectively guide the model towards optimal performance during training. To overcome this limitation, we introduce the Image Matching Score (IMS), a lightweight yet effective reward based on the structural edit distance of column-wise image projections, which robustly quantifies the visual fidelity between rendered formulas. Leveraging IMS, we then propose Image-Matching driven Policy Optimization (IMPO), a training framework built upon Group Relative Policy Optimization (GRPO). This approach facilitates stable policy learning directly from our sequence-level visual reward, notably without the need for a separate value function network. Extensive experiments demonstrate that IMPO yields consistent performance gains across various backbone models on the challenging CROHME, HME100K, and M2E datasets. Our framework establishes new state-of-the-art results, improving the Expression Recognition Rate by an average of 1.1% and up to 1.37% over strong prior methods.
Generative models are often deployed to make decisions on behalf of users, such as vision-language models (VLMs) identifying which person in a room is a doctor to help visually impaired individuals. Yet, VLM decisions are influenced by the perceived demographic attributes of people in the input, which can lead to biased outcomes like failing to identify women as doctors. Moreover, when reducing bias leads to performance loss, users may have varying needs for balancing bias mitigation with overall model capabilities, highlighting the demand for methods that enable controllable bias reduction during inference. Activation steering is a popular approach for inference-time controllability that has shown potential in inducing safer behavior in large language models (LLMs). However, we observe that current steering methods struggle to correct biases, where equiprobable outcomes across demographic groups are required. To address this, we propose Direct Steering Optimization (DSO) which uses reinforcement learning to find linear transformations for steering activations, tailored to mitigate bias while maintaining control over model performance. We demonstrate that DSO achieves state-of-the-art trade-off between fairness and capabilities on both VLMs and LLMs, while offering practitioners inference-time control over the trade-off. Overall, our work highlights the benefit of designing steering strategies that are directly optimized to control model behavior, providing more effective bias intervention than methods that rely on pre-defined heuristics for controllability.
Diffusion Transformers achieve impressive generative quality but remain computationally expensive due to iterative sampling. Recently, dynamic resolution sampling has emerged as a promising acceleration technique by reducing the resolution of early sampling steps. However, existing methods rely on heuristic re-noising at every resolution transition, injecting noise that breaks cross-stage consistency and forces the model to relearn global structure. In addition, these methods indiscriminately upsample the entire latent space at once without checking which regions have actually converged, causing accumulated errors, and visible artifacts. Therefore, we propose Fresco, a dynamic resolution framework that unifies re-noise and global structure across stages with progressive upsampling, preserving both the efficiency of low-resolution drafting and the fidelity of high-resolution refinement, with all stages aligned toward the same final target. Fresco achieves near-lossless acceleration across diverse domains and models, including 10 times speedup on FLUX, and 5 times on HunyuanVideo, while remaining orthogonal to distillation, quantization and feature caching, reaching 22 times speedup when combined with distilled models.
IMU-HOI: A Symbiotic Framework for Coherent Human-Object Interaction and Motion Capture via Contact-Conscious Inertial Fusion
PDF ↗Capturing full-body human motion with object interactions is crucial for AR/VR and robotics applications, yet it remains challenging for conventional vision-based methods due to occlusions and constrained capture volumes. Inertial measurement units (IMUs) offer a compelling alternative without line-of-sight requirements, but existing IMU-based motion capture assumes an isolated human and ignores object contacts and dynamics. To bridge this gap, we present IMU-HOI, a novel framework that jointly recovers full-body human pose and 6-DoF object trajectory from sparse IMUs on the body and object, explicitly modeling human-object Interaction.Our approach first infers probabilistic hand-object contacts directly from IMU streams and uses them as a high-level signal to route between kinematic and inertial reasoning. These contact cues drive a three-stage fusion pipeline that refines human pose and root translation, and fuses hand-based forward kinematics with object-IMU integration for object motion, yielding coherent, drift-resilient trajectories for both human and object. Experiments on challenging human-object interaction scenarios demonstrate substantial accuracy gains over prior inertial motion capture methods. Moreover, IMU-HOI can be plugged into existing sparse-IMU mocap backbones with minimal changes, effectively extending the scope of purely inertial motion capture from isolated humans to full human-object interaction and joint motion estimation.
CLIP achieves strong zero-shot image-text retrieval by aligning global vision and text representations, yet it falls behind on fine-grained tasks even when fine-tuned on long, detailed captions. In this work, we propose b-CLIP, a multi-granular text-conditioned contrastive learning framework designed to achieve hierarchical alignment between multiple textual granularities--from full captions to sentences and phrases--and their corresponding visual regions. For each level of granularity, b-CLIP uses cross-attention to dynamically pool image patches into contextualized visual embeddings. To address the semantic overlap inherent in this hierarchy, we introduce the b-Contextualized Contrastive Alignment Loss (b-CAL), which parameterizes the trade-off between query-specific matching and intra-image contextualization under both Cross-Entropy and Binary Cross-Entropy formulations. Through extensive experiments, we demonstrate that b-CLIP significantly improves dense alignment, achieving 91.8% T2I 92.3% I2T at R@1 on Urban1K and 30.9% on FG-OVD (Hard), setting state-of-the-art among methods trained without hard negatives. b-CLIP establishes a robust, adaptive baseline for dense vision-language correspondence. The code and models are released at \href https://github.com/fzohra/B-CLIP https://github.com/fzohra/B-CLIP .
Vision-language models (VLMs) demonstrate powerful capabilities in multimodal tasks. However, the large number of visual tokens imposes a significant computational cost. In this paper, we propose QuietPrune, a QUery-guIded Early Token Pruning method to remove redundant visual tokens within VLMs, thereby enhancing computational efficiency. Unlike previous late pruning methods, we recognize that implementing early pruning within the vision transformer (ViT) can achieve benefits in both latency reduction and accuracy maintenance. To address the semantic loss problem in early pruning, we design a lightweight adapter by performing an inverse transformation of the projector in VLMs. The proposed adapter converts the contextual query into a visual domain [Q-CLS] (Query [CLS]) token, providing textual guidance for ViT pruning. During pruning, we further introduce a semi-structured pruning scheme based on visual-textual relevance. Specifically, we group spatially adjacent 2 x2 tokens to accommodate the visual token merging operation prevalent in mainstream VLMs. We use the mean attention scores between the [Q-CLS] token and the visual tokens as the relevance metric for each group, avoiding additional computation. Pruning is then applied at the group level based on the relevance score, preserving positional continuity. After pruning, we aggregate the redundant tokens into a single token to maintain context cues. Our method achieves up to 19.0% reduction in prefill latency while outperforming 4.2% in accuracy on the recent Qwen3-VL and InternVL3 series compared to existing late pruning methods.
Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models
PDF ↗The recent success of reinforcement learning (RL) in large reasoning models has inspired the growing adoption of RL for post-training Multimodal Large Language Models (MLLMs) to enhance their visual reasoning capabilities. Although many studies have reported improved performance, it remains unclear whether RL training truly enables models to learn from visual information. In this work, we propose the Hallucination-as-Cue Framework, an analytical framework designed to investigate the effects of RL-based post-training on multimodal reasoning models from the perspective of model hallucination. Specifically, we introduce hallucination-inductive, modality-specific corruptions that remove or replace essential information required to derive correct answers, thereby forcing the model to reason by hallucination. By applying these corruptions during both training and evaluation, our framework provides a unique perspective for diagnosing RL training dynamics and understanding the intrinsic properties of datasets. Through extensive experiments and analyses across multiple multimodal reasoning benchmarks, we reveal that the role of model hallucination for RL-training is more significant than previously recognized. For instance, we find that RL post-training under purely hallucination-inductive settings can still significantly improve models' reasoning performance, and in some cases even outperform standard training. These findings challenge prevailing assumptions about MLLM reasoning training and motivate the development of more modality-aware RL-based training designs.
Recent advances in subject-driven video generation with large diffusion models have enabled personalized content synthesis conditioned on user-provided subjects. However, existing methods lack fine-grained temporal control over subject appearance and disappearance, which are essential for applications such as compositional video synthesis, storyboarding, and controllable animation. We propose AlcheMinT, a unified framework that introduces explicit timestamps conditioning for subject-driven video generation. Our approach introduces a novel positional encoding mechanism that unlocks the encoding of temporal intervals, associated in our case with subject identities, while seamlessly integrating with the pretrained video generation model positional embeddings. Additionally, we incorporate subject-descriptive text tokens to strengthen binding between visual identity and video captions, mitigating ambiguity during generation. Through token-wise concatenation, AlcheMinT avoids any additional cross-attention modules and incurs negligible parameter overhead. We establish a benchmark evaluating multiple subject identity preservation, video fidelity, and temporal adherence. Experimental results demonstrate that AlcheMinT achieves visual quality matching state-of-the-art video personalization methods, while, for the first time, enabling precise temporal control over multi-subject generation within videos.
Recent advances in Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have made it standard practice to reconstruct 3D scenes from multi-view images. Removing objects from such 3D representations is a fundamental editing task that requires complete and seamless inpainting of occluded regions, ensuring consistency in geometry and appearance. Although existing methods have made notable progress in improving inpainting consistency, they often neglect global lighting effects, leading to physically implausible results. Moreover, these methods struggle with view-dependent non-Lambertian surfaces, where appearance varies across viewpoints, leading to unreliable inpainting. In this paper, we present 3D Gaussian Object Removal in the Intrinsic Space (GOR-IS), a novel framework for physically consistent and visually coherent 3D object removal. Our approach decomposes the scene into intrinsic components and explicitly models light transport to maintain global lighting effects consistency. Furthermore, we introduce an intrinsic-space inpainting module that operates directly in the material and lighting domains, effectively addressing the challenges posed by non-Lambertian surfaces. Extensive experiments on both synthetic and real-world datasets demonstrate that our framework substantially improves the physical consistency and visual coherence of object removal, outperforming existing methods by 13% in perceptual similarity (LPIPS) and 2dB in peak signal-to-noise ratio (PSNR). Code is publicly available at https://applezyh.github.io/GOR-IS-project-page/
Real-world object detection must operate in evolving environments where new classes emerge, domains shift, and unseen objects must be identified as unknown--all without accessing prior data. We introduce Evolving World Object Detection (EWOD), a paradigm coupling incremental learning, domain adaptation, and unknown detection under exemplar-free constraints. To tackle EWOD, we propose EW-DETR framework that augments DETR-based detectors with three synergistic modules: Incremental LoRA Adapters for exemplar-free incremental learning under evolving domains; a Query-Norm Objectness Adapter that decouples objectness-aware features from DETR decoder queries; and Entropy-Aware Unknown Mixing for calibrated unknown detection. This framework generalises across DETR-based detectors, enabling state-of-the-art RF-DETR to operate effectively in evolving-world settings. We also introduce FOGS (Forgetting, Openness, Generalisation Score) to holistically evaluate performance across these dimensions. Extensive experiments on Pascal Series and Diverse Weather benchmarks show EW-DETR outperforms other methods, improving FOGS by 57.24%.