Model merging aims to integrate multiple task-specific fine-tuned models derived from a shared pre-trained checkpoint into a single multi-task model without additional training. Despite extensive research, task interference remains a major obstacle that often undermines the performance of merged models. In this paper, we propose ESM (Essential Subspace Merging) , a robust framework for effective model merging. We begin by performing Principal Component Analysis (PCA) on feature shifts induced by parameter updates. The resulting principal directions span an essential subspace that dominantly influences feature representations. Each task's parameter update matrix is projected onto its respective essential subspace for low-rank decomposition before merging. This methodology mitigates inter-task interference while preserving core task-specific functionality. Furthermore, we introduce a multi-level polarized scaling strategy that amplifies parameters containing critical knowledge and suppresses redundant ones, preventing essential knowledge from being overwhelmed during fusion. Extensive experiments across multiple task sets and model scales demonstrate that our method achieves state-of-the-art performance in multi-task model merging.
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
One of the major differentiators unlocked by learned codecs relative to their hard-coded traditional counterparts is their ability to be optimized directly to appeal to the human visual system. Despite this potential, a perceptual yet practical image codec is yet to be proposed.In this work, we aim to close this gap. We conduct a comprehensive study of the key modeling choices that govern the design of a practical learned image codec, jointly optimized for perceptual quality and runtime -- including within the ablations several novel techniques. We then perform performance-aware neural architecture search over millions of backbone configurations to identify models that achieve the target on-device runtime while maximizing compression performance as captured by perceptual metrics. We combine the various optimizations to construct a new codec that achieves a significantly improved tradeoff between speed and perceptual quality. Based on rigorous subjective user studies, it provides 2.3-3x bitrate savings against AV1, AV2, VVC, ECM and JPEG-AI, and 20-40% bitrate savings against the best learned codec alternatives. At the same time, on an iPhone 17 Pro Max, it encodes 12MP images as fast as 230ms, and decodes them in 150ms -- faster than most top ML-based codecs run on a V100 GPU.
Prompt learning has emerged as an effective technique for fine-tuning large-scale foundation models for downstream tasks. However, conventional prompt learning methods are prone to overfitting and can struggle with out-of-distribution generalization. To address these limitations, Bayesian prompt learning has been proposed, which frames prompt optimization as a Bayesian inference problem to enhance robustness. This paper introduces Repulsive Bayesian Prompt Learning (ReBaPL), a novel method for Bayesian prompt learning, designed to efficiently explore the complex and often multimodal posterior landscape of prompts. Our method integrates a cyclical step-size schedule with a stochastic gradient Hamiltonian Monte Carlo (SGHMC) algorithm, enabling alternating phases of exploration to discover new modes, and exploitation to refine existing modes. Furthermore, we introduce a repulsive force derived from a potential function over probability metrics (including Maximum Mean Discrepancy and Wasserstein distance) computed on the distributions of representations produced by different prompts. This representation-space repulsion diversifies exploration and prevents premature collapse to a single mode. Our approach allows for a more comprehensive characterization of the prompt posterior distribution, leading to improved generalization. In contrast to prior Bayesian prompt learning methods, our method provides a modular plug-and-play Bayesian extension of any existing prompt learning method based on maximum likelihood estimation. We demonstrate the efficacy of ReBaPL on several benchmark datasets, showing superior performance over state-of-the-art prompt learning methods.
Looking Beyond the Window: Global-Local Aligned CLIP for Training-free Open-Vocabulary Semantic Segmentation
PDF ↗A sliding-window inference strategy is commonly adopted in recent training-free open-vocabulary semantic segmentation methods to overcome limitation of the CLIP in processing high-resolution images. However, this approach introduces a new challenge: each window is processed independently, leading to semantic discrepancy across windows. To address this issue, we propose Global-Local Aligned CLIP (GLA-CLIP), a framework that facilitates comprehensive information exchange across windows. Rather than limiting attention to tokens within individual windows, GLA-CLIP extends key-value tokens to incorporate contextual cues from all windows. Nevertheless, we observe a window bias: outer-window tokens are less likely to be attended, since query features are produced through interactions within the inner window patches, thereby lacking semantic grounding beyond their local context. To mitigate this, we introduce a proxy anchor, constructed by aggregating tokens highly similar to the given query from all windows, which provides a unified semantic reference for measuring similarity across both inner- and outer-window patches. Furthermore, we propose a dynamic normalization scheme that adjusts attention strength according to object scale by dynamically scaling and thresholding the attention map to cope with small-object scenarios. Moreover, GLA-CLIP can be equipped on existing methods and broad their receptive field. Extensive experiments validate the effectiveness of GLA-CLIP in enhancing training-free open-vocabulary semantic segmentation performance. Codes are available at github.com/2btlFe/GLA-CLIP
Brain visual decoding aims to recognize and reconstruct perceptual visual content from brain activity, providing a promising potential for the development of brain-computer interfaces and brain-inspired intelligence. However, this task faces a fundamental challenge of information asymmetry: while natural images contain complex visual scenes with objects and backgrounds, the corresponding brain signals reflect focused attention on central objects while being contaminated by various neural noise. Previous methods that directly align visual and brain representations often overlook this inherent asymmetry, resulting in suboptimal decoding performance. To address this, we propose linguistic-prior-guided visual decoupling method, which introduces object-oriented textual descriptions as semantic guidance to explicitly decouple foreground objects from complex backgrounds in natural images. This design enables the model to automatically focus on task-relevant visual concepts while effectively filtering out irrelevant neural noise in brain signals, achieving a transition from asymmetric vision-brain alignment to semantic symmetric alignment. Extensive experiments on the THINGS-EEG and THINGS-MEG datasets demonstrate that our method achieves new state-of-the-art performance in the zero-shot brain-to-image retrieval task. The source code is available at https://github.com/TKQXX/BVSA.
Recent work has shown that eliciting Large Language Models (LLMs) to generate reasoning traces in natural language before answering the user's request can significantly improve their performance across tasks. This approach has been extended to multimodal LLMs, where the models can produce chains-of-thoughts (CoT) about the content of input images and videos. For video inputs, prior works use complex multi-step pipelines that extract and include relevant frames from videos in the CoT, or produce simpler single-stage reasoning traces at the expense of poor temporal grounding. Here, we propose the first video LLMs with single-stage reasoning that includes explicit references to relevant frames, thereby reducing temporal inconsistencies in the reasoning process. Our approach is simple, unified, and self-contained, employing a single-stage inference to handle complex video understanding tasks without relying on auxiliary modules for frame selection or caption generation. For this, we first create CoF-Data, a large dataset of diverse questions, answers, and corresponding frame-grounded reasoning traces from both natural and synthetic videos, spanning various topics and tasks. Our models, obtained by fine-tuning video LLMs on this chain-of-frames (CoF) data, generate reasoning traces that accurately identify key frames to answer given questions. In turn, this consistently improves performance across multiple video understanding benchmarks. Surprisingly, we find that synthetic data alone, despite being out-of-distribution with respect to these real-world benchmarks, provides a significant boost in model accuracy.
Inverse tone mapping (ITM) becomes significantly harder when the SDR input is produced by local tone mapping, which jointly applies global radiometric compression and spatially varying adaptations that distort dynamic range, contrast, and channel-wise color ratios. Existing ITM methods ignore this degradation structure and either regress HDR values directly or rely on a single-channel gain map, which scale luminance only and cannot restore the compressed dynamic range and wide color gamut.We introduce FastGaMer, a structured and resolution-agnostic ITM framework that explicitly mirrors this degradation process. Instead of regressing HDR values, we reconstruct a color gain map, which preserves per-channel amplification, simplifies learning, and enables proper gamut extension. Local and global degradations are inverted separately using dynamic bilateral grids and learnable 3D LUTs, followed by a lightweight neural modulator for global refinement and coherence. All high-resolution operations are network-free, yielding exceptional efficiency.To support color-GM supervision under realistic local TMO degradations, we create a dataset of over 8,000 4K SDR-GM pairs with an additional real-captured test set. FastGaMer outperforms prior lightweight ITM methods by +1.4 dB PQ-PSNR, reduces runtime by 70%, and processes 4K images in only 6.2 ms, achieving both high accuracy and real-time performance.
Humans rarely plan whole-body interactions with objects at the level of explicit whole-body movements. High-level intentions, such as affordance, define the goal, while coordinated balance, contact, and manipulation can emerge naturally from underlying physical and motor priors. Scaling such priors is key to enabling humanoids to compose and generalize loco-manipulation skills across diverse contexts while maintaining physically coherent whole-body coordination. To this end, we introduce InterPrior, a scalable framework that learns a unified generative controller through large-scale imitation pretraining and post-training by reinforcement learning. InterPrior first distills a full-reference imitation expert into a versatile, goal-conditioned variational policy that reconstructs motion from multimodal observations and high-level intent. While the distilled policy reconstructs training behaviors, it does not generalize reliably due to the vast configuration space of large-scale human-object interactions. To address this, we apply data augmentation with physical perturbations, and then perform reinforcement learning finetuning to improve competence on unseen goals and initializations. Together, these steps consolidate the reconstructed latent skills into a valid manifold, yielding a motion prior that generalizes beyond the training data, e.g., it can incorporate new behaviors such as interactions with unseen objects. We further demonstrate its effectiveness for user-interactive control and its potential for real robot deployment.
Current approaches to zero-shot 3D-object perception typically rely on ensembles of frozen foundation models. This limits deep object understanding and cross-domain generalization, making performance inadequate for real-world deployment. The 3D-Object Perception Transformer (3PT) addresses this limitation by unifying detection, segmentation, and 6DoF pose estimation in a single framework, directly trained for 3D-object perception. Based on two large-scale trained transformers that specialize in 2D and 3D object-centric scene understanding respectively, 3PT continuously refines its object representations without depth input, enhancing 3D understanding by incorporating multi-view information. 3PT is the state-of-the-art for detection and pose estimation on the BOP benchmarks, often achieving double digit improvements, in many cases, outperforming non-zero-shot methods, and winning 7 of 11 tracks in the BOP-2025 challenge. 3PT surpasses task-specialized models for detection and pose estimation, often achieving double-digit percentage improvements on the diverse BOP-benchmarks, and in some cases outperforming non zero-shot methods. It also ranked first in 7 of 11 tracks at the BOP Challenge 2025. 3PT's high-accuracy and reliability is well-suited for practical industrial robotics applications such as bin picking and precise insertion. Project Page can be found at https://www.intrinsic.ai/publications/3pt-cvpr2026.
CROWn: A Unified Framework for Anti-Aliased Downsampling and Phase-Calibrated Fusion in 3D Medical Segmentation
PDF ↗Precise 3D medical image segmentation is a clinical cornerstone for diagnosis, therapy planning, and longitudinal monitoring. However, routine acquisition with anisotropic voxel spacing and heterogeneous reconstruction induces downsampling aliasing and cross-scale misalignment that blur boundaries, fragment topology, and undermine reliability. Existing U-shaped CNN or Transformer designs neither control alias injection at decimation nor explicitly align high-resolution evidence before decoder fusion, leading to unstable interfaces under device and protocol variability. We introduce the Coset-fibRated micrO-local co-attention Network (CROWn), a general segmentation framework that couples sampling theory with representation learning to jointly suppress aliasing and calibrate cross-scale fusion. CROWn comprises two complementary components. The Microlocal Polyphase Co-Attentive Decimator (\muPCAD) performs axis-aware polyphase analysis with pooled-subband co-attention and explicit anti-alias low-pass, routing boundary-relevant high-frequency evidence while attenuating spurious phase components during downsampling. The Octaphase Coset Fibration (OCF) anti-aliases high-resolution skips, restructures them via 3D space-to-depth into cosets, and applies phase attention with edge-gated modulation to deliver compact, phase-aligned, boundary-aware features to the decoder. Extensive evaluations across 15 publicly available datasets spanning CT, MRI, and OCT demonstrate CROWn's state-of-the-art performance against 17 recent leading methods, improves overlap and topological consistency, consistently reduces boundary errors, while maintaining controlled training and inference cost. The code is publicly available at https://github.com/IMOP-lab/CROWn.
While the most fundamental pretraining paradigm typically trains modality-specific models on their respective datasets, the Platonic Representation Hypothesis that representations eventually align across modalities as data and model scale suggests an intriguing possibility: large language models (LLMs) could be pretrained on visual corpora to reach parity with text-pretrained models, thereby expanding data sources to break the text-scaling bottlenecks, and leveraging richer visual cues for more comprehensive corpus understanding. This paper makes the first attempt to demonstrate the feasibility of this implication by introducing Masked Autoregressive Pretraining for Learning language intelligencE (MAPLE), a novel visual pretraining paradigm for LLMs that leverages raw document images to improve language intelligence. MAPLE is universal to integrate masked auto-regressive models with various LLM backbones, where the LLMs are incentivized to generate latent hypotheses for the masked regions based on the unmasked regions. We verify MAPLE in the domain of math reasoning with multiple LLM backbones and show that MAPLE consistently surpasses text-only pretraining relatively by at most 40.2% on average accuracy across four math reasoning benchmarks. Further analyses show that visually pretrained LLMs learn a shared latent space that aligns document visuals with text and exploits layout and structural cues, supporting visual pretraining as a feasible and scalable route to stronger language models.
This paper tackles the challenge of recovering 4D dynamic scenes from videos captured by as few as four portable cameras. Learning to model scene dynamics for temporally consistent novel-view rendering is a foundational task in computer graphics, where previous works often require dense multi-view captures using camera arrays of dozens or even hundreds of views. We propose 4C4D, a novel framework that enables high-fidelity 4D Gaussian Splatting from video captures of extremely sparse cameras. Our key insight lies that the geometric learning under sparse settings is substantially more difficult than modeling appearance. Driven by this observation, we introduce a Neural Decaying Function on Gaussian opacities for enhancing the geometric modeling capability of 4D Gaussians. This design mitigates the inherent imbalance between geometry and appearance modeling in 4DGS by encouraging the 4DGS gradients to focus more on geometric learning. Extensive experiments across sparse-view datasets with varying camera overlaps show that 4C4D achieves superior performance over prior art.
Recent photo-realistic 3D talking head via 3D Gaussian Splatting still has significant shortcoming in emotional expression manipulation, especially for fine-grained and expansive dynamics emotional editing using multi-modal control. This paper introduces a new editable 3D Gaussian talking head, i.e. EmoDiffTalk. Our key idea is a novel Emotion-aware Gaussian Diffusion, which includes an action unit (AU) prompt Gaussian diffusion process for fine-grained facial animator, and moreover an accurate text-to-AU emotion controller to provide accurate and expansive dynamic emotional editing using text input. Experiments on public EmoTalk3D and RenderMe-360 datasets demonstrate superior emotional subtlety, lip-sync fidelity, and controllability of our EmoDiffTalk over previous works, establishing a principled pathway toward high-quality, diffusion-driven, multimodal editable 3D talking-head synthesis. To our best knowledge, our EmoDiffTalk is one of the first few 3D Gaussian Splatting talking-head generation framework, especially supporting continuous, multimodal emotional editing within the AU-based expression space.Please visit our website for more details and information:https://liuchang883.github.io/EmoDiffTalk/
MaskDiME: Adaptive Masked Diffusion for Precise and Efficient Visual Counterfactual Explanations
PDF ↗Visual counterfactual explanations aim to reveal the minimal semantic modifications that can alter a model's prediction, providing causal and interpretable insights into deep neural networks. However, existing diffusion-based counterfactual generation methods are often computationally expensive, slow to sample, and imprecise in localizing the modified regions. To address these limitations, we propose MaskDiME, a simple, fast, yet effective diffusion framework that unifies semantic consistency and spatial precision through localized sampling. Our approach adaptively focuses on decision-relevant regions to achieve localized and semantically consistent counterfactual generation while preserving high image fidelity. Our training-free framework, MaskDiME, performs inference over 30x faster than the baseline and achieves comparable or state-of-the-art performance across five benchmark datasets spanning diverse visual domains, establishing a practical and generalizable solution for efficient counterfactual explanation. Our code is available at https://github.com/clguo/MaskDiME.
Recent years have witnessed remarkable progress in world models, which primarily aim to capture the spatiotemporal correlations between an agent's actions and the evolving environment. However, existing approaches often suffer from tight runtime coupling or depend on offline reward signals, resulting in substantial inference overhead or hindering end-to-end optimization. To overcome these limitations, we introduce WPT, a World-to-Policy Transfer training paradigm that enables online distillation under the guidance of an end-to-end world model. Specifically, we develop a trainable reward model that infuses world knowledge into a teacher policy by aligning candidate trajectories with the future dynamics predicted by the world model. Subsequently, we propose policy distillation and world reward distillation to transfer the teacher's reasoning ability into a lightweight student policy, enhancing planning performance while preserving real-time deployability. Extensive experiments on both open-loop and closed-loop benchmarks show that WPT achieves state-of-the-art performance with a simple policy architecture: it attains a 0.11 collision rate (open-loop) and achieves a 79.23 driving score (closed-loop), surpassing both world-model-based and imitation-learning methods in accuracy and safety. Moreover, the student sustains up to 4.9xfaster inference, while retaining most of the gains.
Normalizing Flows (NFs) have been established as a principled framework for generative modeling. Standard NFs consist of a forward process and a reverse process: the forward process maps data to noise, while the reverse process generates samples by inverting it. Typical NF forward transformations are constrained by explicit invertibility, ensuring that the reverse process can serve as their exact analytic inverse. Recent developments in TARFlow and its variants have revitalized NF methods by combining Transformers and autoregressive flows, but have also exposed causal decoding as a major bottleneck. In this work, we introduce Bidirectional Normalizing Flow (BiFlow), a framework that removes the need for an exact analytic inverse. BiFlow learns a reverse model that approximates the underlying noise-to-data inverse mapping, enabling more flexible loss functions and architectures. Experiments on ImageNet demonstrate that BiFlow, compared to its causal decoding counterpart, improves generation quality while accelerating sampling by up to two orders of magnitude. BiFlow yields state-of-the-art results among NF-based methods and competitive performance among single-evaluation ("1-NFE") methods. Following recent encouraging progress on NFs, we hope our work will draw further attention to this classical paradigm.
Despite remarkable advancements in multimodal large language models (MLLMs), their fine-grained visual understanding is constrained by a primary reliance on sparse textual supervision. Existing efforts to introduce visual supervision typically do so during post-training, when visual representations have already been largely fixed, causing such signals to act mainly as auxiliary constraints rather than as a primary force for shaping perceptual features. In this paper, we aim to fundamentally reshape the model's perceptual backbone by incorporating vision supervision directly into the pre-training stage. We observe that pixel-level image patches and textual tokens naturally coexist in a shared, raw high-dimensional space characterized by an inherent input symmetry. Leveraging this insight, we propose UVU, a novel vision-language unified autoregressive framework that eschews vector quantization. It uniquely employs continuous visual encoding for lossless representation of visual inputs and proposes a large-scale iterative hierarchical clustering algorithm to construct a pixel-level visual codebook, thereby extending the vocabulary for unified supervision and enabling autoregressive generation of pixel-level image tokens alongside textual tokens. UVU effectively synergizes pixel-level visual perception with semantic-level visual understanding, internalizing visual reconstruction capabilities and unlocking the facilitative role of visual supervision in enhancing understanding in the pre-training stage. Extensive experiments across multiple tasks demonstrate that MLLMs are capable of achieving superior multimodal understanding performance under the supervised learning paradigm of UVU.
VES-RFT: Rewarding Visual Evidence Sensitivity to Mitigate Hallucinations in Large Vision-Language Models
PDF ↗Vision-Language Models (VLMs) often over-rely on linguistic priors even when images are provided, leading to object hallucinations. We revisit object-wise hallucination from the perspective of how visual evidence shapes the model's uncertainty. For each input, we measure decision uncertainty with and without the image, and define a Visual Evidence Sensitivity (VES) signal as the image-attributable change in entropy. Building on this signal, we introduce Visual Evidence Sensitivity Reinforcement Fine-Tuning (VES-RFT), a training-time reinforcement fine-tuning method that explicitly rewards reliance on correct visual evidence. We pair this continuous, annotation-free signal with a verifiable reward that enforces factual object correctness by automatically checking generated object mentions against the image, yielding a computable objective without human annotations. We optimize the dual objective using critic-free GRPO with KL regularization, requiring only parallel image and no-image passes during training while preserving single-pass inference. Across multiple VLM families and benchmarks, VES-RFT consistently suppresses object hallucinations and improves robustness under ambiguity. Specifically, on LLaVA-7B, VES-RFT reduces 12.8 and 1.8 on CHAIR_S/CHAIR_I of MS-COCO, and increases POPE accuracy by 4.92%. Extensive experiments indicate that turning uncertainty into a learnable reward, paired with verifiable correctness signals, provides a scalable mechanism for training-time hallucination mitigation and stronger visual grounding.
UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement
PDF ↗Image quality assessment (IQA) and image restoration are fundamental problems in low-level vision. Although IQA and restoration are closely connected conceptually, most existing work treats them in isolation. Recent advances in unified multimodal understanding-generation models demonstrate promising results and indicate that stronger understanding can improve generative performance. This motivates a single model that unifies IQA and restoration and explicitly studies how IQA can guide restoration, a setting that remains largely underexplored yet highly valuable. In this paper, we propose UARE, to our knowledge the first Unified vision-language model for image quality Assessment, Restoration, and Enhancement. Built on pretrained unified understanding and generation models, we introduce a two-stage training framework. First, a progressive, easy-to-hard schedule expands from single-type distortions to higher-order mixed degradations, enabling UARE to handle multiple degradations. Second, we perform unified fine-tuning of quality understanding and restoration with interleaved text-image data, aligning IQA signals with restoration objectives. Through multi-task co-training, UARE leverages IQA to boost restoration and enhancement performance. Extensive experiments across IQA, restoration, and enhancement tasks demonstrate the effectiveness of UARE. The code and models will be made available.
On Token's Dilemma: Dynamic MoE with Drift-Aware Token Assignment for Continual Learning of Large Vision Language Models
PDF ↗Multimodal Continual Instruction Tuning aims to continually enhance Large Vision Language Models (LVLMs) by learning from new data without forgetting previously acquired knowledge. Mixture of Experts (MoE) architectures naturally facilitate this by incrementally adding new experts and expanding routers while keeping the existing ones frozen. However, despite expert isolation, MoE-based continual learners still suffer from forgetting due to routing-drift: old-task tokens become mistakenly attracted to newly added experts, degrading performance on prior tasks. We analyze the failure mode at the token level and reveal the token's dilemma: ambiguous and old tokens in new-task data offer minimal learning benefit yet induce forgetting when routed to new experts, due to their ambiguous routing assignment during training. Motivated by this, we propose LLaVA-DyMoE, a dynamic MoE framework that incrementally expands the MoE with drift-aware token assignment. We characterize token types via their routing score distributions and apply targeted regularization. Specifically, a token-level assignment guidance steers ambiguous and old tokens away from new experts to preserve established routing patterns and alleviate routing-drift, while complementary routing score regularizations enforce expert-group separation and promote new-expert specialization. Extensive experiments demonstrate that our LLaVA-DyMoE effectively mitigates routing-drift-induced forgetting, achieving over a 7% gain in mean final accuracy and a 12% reduction in forgetting compared to baselines. The project page is zhaoc5.github.io/DyMoE.