ECCV 2026 论文列表

Vision-Language-Action (VLA) models have made signifi-cant strides in embodied intelligence by integrating the powerful repre-sentations of pre-trained Vision-Language Models (VLMs). However, themassive parameter scale of VLAs imposes a heavy computational burden,and these models exhibit extreme sensitivity to parameter pruning. Cur-rent paradigms often treat the resulting performance degradation as in-evitable, relying on fine-tuning or low-rank corrections to recover efficacy.We challenge this convention by questioning whether the removed param-eters are truly redundant if VLA pruning necessitates performance recov-ery to be effective, or if this paradigm masks the indiscriminate pruningof critical parameters. We revisit parameter redundancy through the lensof VLM-to-VLA adaptation, first quantifying the spatial distribution ofparameter divergence during adaptation to reveal structured patternsacross different modules. Subsequently, we introduce controlled pruningas a diagnostic probe: by comparing the direct impact of removing dif-ferent parameter subsets on VLA performance without any fine-tuning,we establish a causal link between adaptation-induced divergence signalsand functional contributions. Based on the discovered modular hetero-geneities, we design a multi-module joint pruning scheme. Evaluationson the LIBERO benchmark demonstrate that our approach reduces theparameters of OpenVLA and π0.5 by 12%–30% while maintaining approx-imately 90% of the original performance without any post-pruning recov-ery. In contrast, existing parameter pruning criteria result in total perfor-mance collapse when evaluated under the same recovery-free constraints.Our study reveals the parameter evolution mechanism in VLA adapta-tion and provides a new path for deploying efficient, robust robotic poli-cies in resource-constrained environments. Code is available at https://github.com/Niannnnnn/VLA_Parameter_Redundancy_VLM2VLA.

High-performance Self-Supervised Learning (SSL) encodersdeployed as Encoder-as-a-Service (EaaS) are highly vulnerable to ad-vanced contrastive model stealing attacks, such as RDA and ContSteal.Traditional defenses often stack complex modules or rely on heuristicnoise, struggling to balance privacy and utility. To address this, we pi-oneer the use of active perturbations for EaaS protection by proposingUniTrap, a simple yet effective single-module defense specifically tai-lored against contrastive-based stealing paradigms. Instead of mere dis-tance maximization, UniTrap explicitly optimizes global feature entropyvia a Gaussian potential function. This generates a stealthy perturba-tion that strictly enforces an isotropic, uniform distribution on the unithypersphere—acting as “one trap” to fundamentally block contrastivealignment. We theoretically demonstrate that this perfect uniformitycauses the attacker’s informative gradients to cancel out in the tangentspace, inducing gradient vanishing and permanently stalling surrogateoptimization. Extensive experiments confirm that our method achievesa state-of-the-art privacy-utility Pareto frontier across various bench-marks.

Promptable foundation segmentation models exhibit impres-sive capabilities across diverse tasks, yet their performance deterioratessubstantially on medical images due to domain and sub-domain shifts.Online test-time adaptation (OTTA) helps alleviate these shifts, but ex-isting methods remain limited in effectiveness and efficiency. To thisend, we introduce Semantic-Preserving Dual-Perturbation Adaptation(SPDA), a resource-efficient OTTA framework tailored for promptablemedical segmentation. SPDA constructs diverse and robust learning sig-nals by enforcing feature and prediction consistency across two semantic-preserving perturbations: (i) a domain-adaptive low-frequency radial-spectrum perturbation that simulates task-irrelevant intra-sub-domainvariations without out-of-domain distortions, and (ii) a prompt under-sampling perturbation that reduces prompt density while preservingtarget semantics. These consistency constraints drive the optimizationof a lightweight post-encoder adapter, facilitating efficient adaptationwith minimal computational overhead. Extensive experiments on varied2D/3D medical data demonstrate that SPDA consistently outperformsstate-of-the-art OTTA baselines under box and point prompts, whileusing substantially fewer resources. Ablation studies validate each com-ponent and indicate the suitability for real-time clinical deployment.

Long-form narrative-to-film generation requires shot-level con-trollability and cross-clip consistency in both visual identity and charac-ter behavior—requirements that remain difficult to satisfy with currentprompt-based workflows. A core reason existing workflows remain brit-tle is the lack of a structured intermediate layer between scripts andvideo models, especially when screenplays are underspecified at key cin-ematic decision points. We introduce a structured orchestration layer forfilm-oriented script-to-video generation, implemented as a multi-agentframework that operates between scripts and off-the-shelf video genera-tors. The layer is centered on FilmDSL, a film-oriented domain-specificlanguage that makes cinematic constraints explicit, including shot andcamera directives, asset and continuity requirements, and persona cues,so that agents coordinate through a shared structured specification forplanning, generation, critique, and repair. Specifically, a generation agentconstructs asset packs and storyboard keyframes that anchor composi-tion before clip-by-clip synthesis, while a critic agent produces structuredQA signals and triggers targeted refinement without retraining the basemodel. Experiments on TV-style segments show improved controllabilityand consistency over text-only and reference-only baselines.

Large multimodal models have achieved strong reasoning oncomplex visual tasks, but their inference efficiency is often restricted bylong chains of thought. A promising solution is to pair a small draftmodel with a large target model, enabling cooperative inference employ-ing a routing signal that adaptively routes queries to either the draft ortarget model based on their difficulties for optimal efficiency and accu-racy. Yet, the remaining bottleneck is to establish a reliable query dif-ficulty signal under multimodal settings. Existing approaches designedfor language models either rely on post-hoc token probabilities, whichfall short in multimodal scenarios, or depend on supervised fine-tuning,which is a data-sensitive strategy. Both paradigms perform routing onlyafter a complete output, and ignore whether the target model can ac-tually solve the routed instances. To address this, we propose PRP,a Proactive Routing Paradigm that enables early decision-making byjointly evaluating the competence of both the draft and target models.Our Draft Rating Learning (DRL) equips the draft model with an inter-nal confidence estimator, while Joint Rating Learning (JRL) predictshow well the target model can handle a given query, thereby prioritiz-ing the allocation of samples it excels at rather than the hardest ones.These ratings enable fine-grained, instance-level Proactive Routingand substantially accelerate inference without compromising overall per-formance. Extensive experiments across multiple multimodal reasoningbenchmarks validate our effectiveness and efficiency.

Diffusion Transformers have become the dominant paradigmin generative AI, but their high computational costs severely hinderreal-time applications. Prediction-based feature caching is widely usedto accelerate diffusion transformers; however, as the number of stepsincreases, the deviation between its predictions and the reference full-compute trajectory gradually grows. An intuitive idea is to use an onlineregression model to dynamically correct this deviation, but it faces theissue of label data being unavailable during the acceleration process.This paper presents a statistical observation that the residuals betweenthe features of full computation steps using caching methods and refer-ence full-compute trajectory locally exhibit a zero-mean Gaussian dis-tribution. By treating the features of full computation steps as noisyobservations of reference features, the data acquisition problem is re-solved. Based on this observation, a plug-and-play GP-Refiner correctionframework is proposed. This method utilizes Gaussian Process Regres-sion for correction and, leveraging the properties of GPR, introducesan uncertainty-adaptive computation strategy that triggers necessaryfull-computation calibration by monitoring the posterior variance in realtime. Experiments demonstrate significant improvements across differentmodels when combined with various state-of-the-art methods. Integrat-ing the proposed framework with TaylorSeer reduces the computationalload by 19.3% while improving PSNR by 0.9 dB and reducing LPIPSfrom 0.46 to 0.29. Code is available in https://github.com/Aredstone/GP-Refiner.

Transferring articulated motion from monocular videos torigged 3D characters is challenging due to pose ambiguity in 2D obser-vations and morphological differences between source and target. Ex-isting approaches often follow a reconstruct-then-retarget paradigm, ty-ing transfer quality to intermediate 3D reconstruction and limiting ap-plicability to categories with parametric templates. We propose Mor-phGS, a framework that formulates motion retargeting as a target-drivenanalysis-by-synthesis problem, directly optimizing target morphology andpose through image-space supervision. A rig-coupled morphology pa-rameterization factorizes character identity from time-varying joint rota-tions, while dense 2D-3D correspondences and synthesized views providecomplementary structural and multi-view guidance. Experiments on syn-thetic benchmarks and real-world videos show consistent improvementsover baselines. Project page: https://xodus777.github.io/MorphGS/

Generating diverse, coherent, and plausible content from par-tially given inputs remains a fundamental challenge for diffusion models.Existing approaches face clear limitations: training-based approachesoffer strong task-specific results but require costly computation, andthey generalize poorly across tasks. Training-free approaches offer betterefficiency, but they do not explicitly optimize over unobserved variables,leading to globally inconsistent results. To address these limitations, weintroduce Accelerated Likelihood Maximization (ALM), a novel training-free sampling strategy integrated into the reverse diffusion process thatsignificantly extends the applicability of diffusion models beyond simplegeneration tasks. Unlike previous methods that implicitly influence miss-ing regions through pre-generated region constraints, we directly optimizethe unobserved region during the sampling process, enabling globallycoherent and plausible generation. Furthermore, we incorporate an ac-celeration strategy that significantly improves computational efficiencywithout sacrificing performance. Experimental results demonstrate thatALM consistently outperforms state-of-the-art methods in various datadomains and tasks, establishing a powerful paradigm for versatile contentgeneration. Project website: http://hleephilip.github.io/ALM

Active learning (AL) performance is known to be budget-dependent, yet regimes are typically defined by heuristic label countsthat fail to generalize across datasets or architectures. We character-ize AL dynamics by reframing budget regimes as shifts in the domi-nant generalization mechanism. By reinterpreting PAC-style risk com-ponents as dynamic interacting terms, we prove that dominance shiftsare structurally unavoidable, creating a moving bottleneck for gener-alization. We operationalize this using measurable proxies and a seg-mented regression procedure to identify a tripartite taxonomy: data-driven, transition, and model-driven phases. Our framework explains thelong-standing observation that representativeness, coverage, and uncer-tainty strategies excel at different stages. Experiments across natural andmedical imaging show that AL efficiency depends on the alignment be-tween the strategy’s inductive bias and the active bottleneck. Moreover,self-supervised representation shift transitions earlier along the labelingtrajectory, highlighting the role of representation quality in shaping ALdynamics. Overall, this work provides a unified framework for the nextgeneration of transition-aware AL algorithms. The code is available at:https://github.com/juliamachnio/PALM.

Video diffusion models generate high-quality and diverse worlds;however, individual frames often lack 3D consistency across the outputsequence, which makes the reconstruction of 3D worlds difficult. To thisend, we propose a new method that handles these inconsistencies bynon-rigidly aligning the video frames into a globally-consistent coordi-nate frame that produces sharp and detailed pointcloud reconstructions.First, a geometric foundation model lifts each frame into a pixel-wise 3Dpointcloud, which contains unaligned surfaces due to these inconsisten-cies. We then propose a tailored non-rigid iterative frame-to-model ICPto obtain an initial alignment across all frames, followed by a global opti-mization that further sharpens the pointcloud. Finally, we leverage thispointcloud as initialization for 3D reconstruction and propose a novelinverse deformation rendering loss to create high quality and explorable3D environments from inconsistent views. We demonstrate that our 3Dscenes achieve higher quality than baselines, effectively turning videomodels into 3D-consistent world generators.

While Vision–Language Models (VLMs) demonstrate strongcapabilities, they continue to suffer from a critical limitation: insufficientfine-grained visual perception, which fundamentally limits their mul-timodal understanding. We attribute this bottleneck to text-dominantoptimization biases during pre-training, which encourage the model tooverlook fine-grained visual details, thereby limiting the capability ofmultimodal understanding. We investigate that overcoming this bottle-neck requires two key elements: (1) a unified token space paradigm thatensures stable training dynamics, and (2) a modality-aligned dense vi-sual supervision signal enriched with both structural granularity andsemantic information to capture critical visual representations. Basedon these insights, we propose VIVAS, a framework built upon the uni-fied token space paradigm, which introduces a dense-structural–semanticvision tokenizer, which expands the textual vocabulary into a unified vi-sion–language vocabulary by incorporating a visual vocabulary. Duringpretraining, VIVAS performs vision–language unified autoregressive su-pervision over both visual details and linguistic content, thereby enhanc-ing visual perception to improve multimodal understanding. Trained end-to-end on 12.4T tokens, VIVAS achieves state-of-the-art performanceacross 7 tasks and 39 multimodal benchmarks.

Standard-definition (SD) maps are widely used in autonomous driv-ing, but accurate ego-centric SD map retrieval typically relies on costly RTK-GNSS joint localization. In contrast, low-cost GNSS introduces meter-level 3-DoF pose offsets with high variance, yielding misaligned local maps that candegrade downstream planning and control. Prior visual re-localization methodsmostly rely on dense rasterized BEV representations, which are computation-heavy and easily distracted by scene content irrelevant to road geometry. We pro-pose VectorReLoc, the first sparse re-localization framework that directly alignsvectorized SD maps with online-constructed vectorized visual maps to estimatethe 3-DoF pose offset. VectorReLoc further introduces a feature alignment objec-tive that structures the embedding space for accurate offset regression, togetherwith a reliability predictor to identify potentially unreliable corrections and mit-igate silent failures. To enable realistic training and evaluation, we additionallypresent a large-scale dataset with paired RTK-retrieved and GNSS-retrieved SDmaps, providing real pose-offset labels rather than simulated noise used in priorwork. Experiments on public benchmarks and the proposed dataset demonstrateimproved accuracy, robustness, and efficiency over previous approaches.

Vision-language models (VLMs) can describe urban scenesin rich detail, yet consistently fail to produce reliable human preferencelabels in domain-specific tasks such as safety assessment and aestheticevaluation. The standard fix, fine-tuning or RLHF, requires large-scaleannotations and model retraining. We ask a different question: can afrozen VLM be aligned with human preferences without modifying anyweights? Our key insight is that VLMs are strong concept extractorsbut poor decision calibrators. We propose a three-stage post-hoc pipelinethat exploits this asymmetry: (i) interpretable evaluation dimensionsare automatically mined from consensus exemplars; (ii) an Observer–Debater–Judge chain extracts robust concept scores from the frozenVLM; and (iii) locally-weighted ridge regression on a hybrid manifoldcalibrates these scores to human ratings. Applied as UrbanAlign onPlace Pulse 2.0, the framework reaches 70.8% accuracy (κ=0.41) acrosssix perception categories, outperforming all baselines by +9.6 pp andzero-shot VLM by +14.1 pp, with full interpretability and zero weightmodification.

Autoregressive video diffusion models have emerged as apromising approach for long video generation, achieving strong perfor-mance in streaming settings. However, existing methods are restrictedto forward temporal generation, whereas practical video creation oftenrequires flexible generation order, e.g., conditioning on future context toextend backward, or on both past and future context for inbetween gener-ation. We bridge this gap by training a single autoregressive model thatsupports generation in arbitrary temporal directions. A key technicalchallenge arises from the Causal 3D VAE widely used in video diffusionmodels, where latents are encoded strictly conditioned on past context.While suited for forward generation, this causal structure causes inter-block discontinuities when generation proceeds backward. To addressthis, we introduce blockwise anchor latents, a set of auxiliary latentsthat aim to restore the missing past context at block boundaries dur-ing backward generation. Built on this design, we propose UniTemp, abidirectional distillation framework that trains an autoregressive studentmodel for any-direction video generation. At inference time, UniTempconditions on arbitrary past and/or future frames, improving control-lability for both bidirectional and inbetween generation. Through ex-tensive experiments, we demonstrate that UniTemp maintains compet-itive performance on short and long video generation in comparison toforward-only methods, while enabling diverse creation workflows such asbidirectional video extension, inbetween generation, looping video gen-eration, scene transition, and visual story generation. Project website:https://lzhangbj.github.io/projects/unitemp/.

Existing jailbreak defense frameworks for Large Vision-LanguageModels (LVLMs) often suffer from a safety–utility tradeoff, where strength-ening safety inadvertently degrades performance on general visual-groundedreasoning tasks. In this work, we investigate whether safety and utilityare inherently antagonistic objectives. We focus on a modality-inducedbias direction consistently observed across datasets, which arises fromsuboptimal coupling between the LLM backbone and visual encoders.We further demonstrate that this direction undermines performance onboth tasks. Leveraging this insight, we propose TBOP (Two Birds, OneProjection), an efficient inference-time jailbreak defense that projectscross-modal features onto the null space of the identified bias direction toremove the corresponding components. Requiring only a single forwardpass, our method effectively breaks the conventional tradeoff, simultane-ously improving both safety and utility across diverse benchmarks.

Federated Learning (FL) emerged as a promising distributedmachine learning paradigm. However, extending FL to the class incre-mental learning scenarios introduces unique challenges: 1) Capacity con-flict and catastrophic forgetting from the shared model overloading, 2)Heterogeneity from Non-Independent and Identically Distributed (Non-IID) data, and 3) Synchronized class misalignment. In this paper, we pro-pose Fisher-Routed MiXture of Experts for Federated Class-IncrementalLearning (FedFMX), a novel framework to address these challenges viaadaptive expert specialization across clients. The crucial insight is toroute each sample to an expert subset that jointly optimizes knowledgeacquisition and retention. Specifically, we introduce a Fisher-Routed Ex-pert Scoring (FRES) module to estimate expert importance via Fisher-based stability cost and gradient-based plasticity gain. Then, we designan Adaptive Expert Selection (AES) module by quantifying marginalcontributions for adaptive expert subset determination. Finally, by therouting-aware regularization (RAR), we achieve load balance and effi-cient FL training. We theoretically prove the O(T −1 ) convergence rate.Extensive experiments on multiple benchmarks compared with state-of-the-art methods demonstrate the superiority of FedFMX.

We present TriO, a multi-modal unsupervised world modelthat predicts 4D occupancy, obstacle segmentation, flow and LiDAR.In contrast to prior work, TriO utilizes three distinct sensor modali-ties (camera, LiDAR, and RADAR) as both inputs and sources of self-supervision, eliminating the need for additional human annotations. Thanksto its novel supervision, the model is able to segment any occupancyfrom the drivable surface, overcoming the limitations of existing open-set methods in handling long-tail objects. TriO achieves state-of-the-artresults in multiple 3D and 4D tasks, including occupancy, flow, and Li-DAR prediction, as well as zero-shot road obstacle segmentation acrossmultiple datasets such as Argoverse 2, and Spotting the Unexpected.Fig. 1: We present TriO, an unsupervised tri-modal occupancy world model designedto perceive and forecast anything that could be considered path-blocking.

Real-world object detection operates under ambiguous su-pervision, where unlabeled regions may correspond to missing annota-tions of known objects or genuinely unknown categories. These chal-lenges have been addressed separately in Sparsely Annotated Object De-tection (SAOD) and Open-World Object Detection (OWOD). In prac-tice, their co-occurrence remains an open problem. To address this prob-lem, we introduce Sparsely Annotated Open-World Object Detection(SA-OWOD), a new task that jointly considers sparse supervision andthe presence of unseen categories. We propose Dual-Perspective ObjectDiscovery (DPOD), a unified framework that jointly models unlabeledknown and unknown instances via two complementary mechanisms. TheKnown Target Recovery Module (KTRM) recovers supervision for unla-beled known instances and explicitly regularizes the feature space to sep-arate known and unknown representations. Complementarily, the Dual-Disagreement Target Generator (DDTG) identifies reliable unknown can-didates through cross-view semantic inconsistency. By integrating thesemodules, DPOD resolves contradictory supervision signals caused byambiguous unlabeled regions. As a result, it prevents misclassificationbetween known and unknown objects and stabilizes the decision bound-aries. Experimental results on sparsely annotated open-world bench-marks demonstrate that the proposed method outperforms existing open-world detection methods, particularly in detecting unknown objects. Thecode is publicly available at: https://github.com/HelloHeeju/SA-OWOD

The reliance of mobile GUI agents on Multimodal LargeLanguage Models (MLLMs) introduces a severe privacy vulnerability:screenshots containing Personally Identifiable Information (PII) are of-ten sent to untrusted, third-party routers. These routers can exploit theirown MLLMs to mine this data, violating user privacy. Existing privacyperturbations fail the critical dual challenge of this scenario: protect-ing PII from the router’s MLLM while simultaneously preserving taskutility for the agent’s MLLM. To address this gap, we propose the Dual-Task Adversarial Protector (DualTAP), a novel framework that,for the first time, explicitly decouples these conflicting objectives. Dual-TAP trains a lightweight generator using two key innovations: (i) a con-trastive attention module that precisely identifies and targets only thePII-sensitive regions, and (ii) a dual-task adversarial objective that si-multaneously minimizes a task-preservation loss (to maintain agent util-ity) and a privacy-interference loss (to suppress PII leakage). To facilitatethis study, we introduce PrivScreen, a new dataset of annotated mobilescreenshots designed specifically for this dual-task evaluation. Compre-hensive experiments on six diverse MLLMs (e.g., GPT-5) demonstrateDualTAP’s state-of-the-art protection. It reduces the average privacyleakage rate to 31.7 percentage points (a 2.6× relative improvement)while, critically, maintaining an 80.8% task success rate—a negligibledrop from the 83.6% unprotected baseline. DualTAP presents the firstviable solution to the privacy-utility trade-off in mobile MLLM agents.Our code is available at https://github.com/fyzhang1/DualTAP.