ECCV 2026 Papers

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.

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.

Recent training-free Neural Architecture Search (NAS) meth-ods have introduced zero-cost proxies (ZCPs) to automate architecturedesign without expensive training or expert intervention. However, exist-ing ZCPs either completely ignore training dynamics or rely on static orinitialization-time characterizations of optimization that fail to capturefine-grained correlations with performance. Moreover, they overlook theheterogeneous importance of different model components, which limitstheir ability to generalize across datasets and search spaces. To mitigatethese limitations, we propose an Effective Rank Score called ER-Scoreas a novel ZCP to quantify stability in the training dynamics ofover-parameterized networks. Experiments show that ER-Score con-sistently outperforms existing ZCPs across multi-scale tasks and diversearchitectures, including convolutional neural networks (CNNs) and Vi-sion Transformers (ViTs). We further introduce topology-weighted strate-gies for CNNs that incorporate topological information via operation-wise feature map aggregation, resulting in a topology-weighted vari-ant termed TER-Score. Extensive experiments demonstrate that TER-Score achieves state-of-the-art ranking consistency and the best pop-ulation convergence when integrated with evolutionary search on theNAS-Bench-301 benchmark. Finally, our method achieves the lowesttest errors of 2.41% and 23.55% on the DARTS search space for theCIFAR-10 and ImageNet-1k datasets, respectively. Code is released athttps://github.com/Thiswycf/TER-Score.

Thermal infrared (TIR) imaging is indispensable for all-weatherperception and energy management, yet existing 3D thermal modelingmethods remain primarily confined to static environmental snapshots.These approaches typically struggle to capture complex temporal tem-perature variations and lack the capacity to simulate scenes under unseenenvironmental conditions. To bridge this gap, we propose ThermoGS,a unified framework that couples 4D Gaussian Splatting (4DGS) withfundamental thermodynamic principles to achieve physically-consistentthermal field emulation. Unlike previous data-driven methods that di-rectly regress radiance, ThermoGS decouples intrinsic material proper-ties, including emissivity (e), convective heat transfer coefficient (h), vol-umetric heat capacity (c), and absorptivity (α)—from transient temper-ature states. By integrating Physics-Informed Neural Networks (PINNs),our framework ensures that the predicted spatiotemporal thermal distri-bution strictly adheres to energy balance equations, enabling generalizedsimulation across arbitrary viewpoints and time periods. Furthermore,we introduce the Thermo Dataset, a comprehensive, drone-captured col-lection featuring synchronized multi-modal data and diverse environmen-tal parameters. Extensive experiments demonstrate that ThermoGS sig-nificantly outperforms comparative methods in thermal simulation accu-racy, maintaining a predicted temperature error within 1 degree Celsius.The code is available at https://github.com/NPU-CVPG/ThermoGS.

Vision-Language-Action (VLA) models have emerged as apromising paradigm for end-to-end autonomous driving. However, exist-ing VLAs’ traffic relies heavily on text-centric visual question answeringand chain-of-thought reasoning data, which emphasizes linguistic rea-soning rather than action-grounded planning. As a result, the learnedrepresentations capture semantic knowledge but lack spatial dependenciescrucial for reliable trajectory prediction. We propose DriveTeach-VLA,a framework that explicitly teaches VLAs what to see and where tolook. Driving-aware Vision Distillation (DVD) injects driving-specificperceptual priors into the vision encoder, while 2D Trajectory-GuidedPrompts (2D-TGP) provide spatial conditioning aligned with feasibledriving trajectories. Together they form a vision-guided learning pipeline:what to see (DVD pretraining) → where to look (TGP-guided SFT) →how to act (TGP-guided GRPO). DriveTeach-VLA achieves the state-of-the-art performance on NAVSIM and nuScenes. Our code is available at:https://github.com/ShivaTeam/DriveTeach-VLA.

Video Large Language Models (Video LLMs) aim to under-stand and reason over dynamic visual content. However, videos inher-ently exhibit low task-relevant information density because only a smallfraction of frames and regions provide meaningful cues. As a result, theuseful signals are sparse across space and time. We present STVFocus, atraining-free framework that unifies spatio-temporal focusing to capturesalient contexts and build informative video representations. STVFocusconsists of three components. First, Hierarchical Temporal Focus appliesa coarse-to-fine frame selection strategy that captures global temporalstructure while selectively emphasizing informative moments without ex-cessive query bias. Second, Query-informed Spatial Focus uses spatialsaliency to highlight query-relevant intra-frame regions. Finally, Tem-poral Reallocation reinvests the saved spatial capacity into additionalhigh-value frames, enabling richer temporal coverage. By jointly leverag-ing temporal and spatial saliency, STVFocus produces compact yet se-mantically enriched video representations. Experiments on VideoMME,LongVideoBench, and MLVU show consistent gains with meaningfulmargins, confirming the effectiveness of unified spatio-temporal focus-ing.

Image registration is an ill-posed dense vision task, wheremultiple solutions achieve similar loss values, motivating probabilisticinference. Variational inference has previously been employed to capturethese distributions, however restrictive assumptions about the posteriorform can lead to poor characterisation, overconfidence and low-qualitysamples. More flexible posteriors are typically bottlenecked by the com-plexity of high-dimensional covariance matrices required for dense 3Dimage registration.In this work, we present a memory and computationally e!cient in-ference method, Structured SIR, that enables expressive, multi-modal,characterisation of uncertainty with high quality samples. We proposethe use of a Sampled Importance Resampling (SIR) algorithm witha novel memory-e!cient high-dimensional covariance parameterisationas the sum of a low-rank covariance and a sparse, spatially structuredCholesky precision factor. This structure enables capturing complex spa-tial correlations while remaining computationally tractable.We evaluate the e!cacy of this approach in 3D dense image registrationof brain MRI data, which is a very high-dimensional problem. We demon-strate that our proposed method produces uncertainty estimates that aresignificantly better calibrated than those produced by variational meth-ods, achieving equivalent or better accuracy. Crucially, we show thatthe model yields highly structured multi-modal posterior distributions,enable e"ective and e!cient uncertainty quantification.

Deep learning models exhibit fairness concerns when pre-dictions are inadvertently influenced by sensitive attributes. However,existing attempts to make Path-Specific Counterfactual Fairness opti-mizable rely on estimating marginal potential outcome probabilities—anapproach that fundamentally requires high-dimensional conditional den-sity estimation and breaks down in modalities such as medical images,where the curse of dimensionality renders reliable estimation infeasible.To address this limitation, we reduce the problem of enforcing Path-Specific Counterfactual Fairness to a causal conditional independenceconstraint and prove that satisfying this constraint is sufficient to elim-inate the unfair causal effect. This reduction replaces intractable coun-terfactual estimation with a discriminative optimization objective thatremains scalable in high-dimensional settings.

Vision-language models (VLMs) such as CLIP enable zero-shot classification by comparing image features with text prompts in ashared embedding space. A fundamental property underlying this ca-pability is the global comparability of logits across arbitrary candidateclasses. However, VLMs are often adapted to fine-grained domains usingtechniques such as LoRA. While this improves in-domain accuracy, out-of-domain accuracy degrades. This leads to a higly fragmented modelecosystem, with thousands of specialized models. Multi-Expert-Domain(MED) classification seeks to address this problem, by merging LoRAstrained independently on specialized domains. However, due to the inde-pendent training, the various domain experts no longer produce globallycalibrated logits. As a result, when evaluating over the union of multipledomain-specific class sets, heterogeneous logit scales induce cross-domaininterference and artificially high confidence for out-of-domain classes,inducing prediction errors. In this work, we identify domain supervi-sion and cross-domain logit miscalibration as the key issue to scalablemulti-domain zero-shot recognition. We propose a mixture-of-expertsMED architecture, MED-DSLC, combining domain supervised training anddomain-wise logit scaling, to explicitly restore global logit comparability.MED-DSLC is a lightweight solution for MED classification, which is shownto preserve within-domain discrimination while reducing cross-domainlogit interference with minimal data. Extensive experiments across di-verse fine-grained benchmarks demonstrate that it substantially improvesmean accuracy (+15%), cross-domain robustness, and scalability in thesize of MED classification problem. Our results show that restoring output-level calibration is essential under highly data imbalanced settings forachieving a truly zero-shot VLM under multi-domain specialization. Codeis publicly available at MED-DSLC.

Multimodal Large Language Models (MLLMs) have achievedremarkable success in visual understanding but remain constrained invisual generation due to the fundamental feature discrepancy betweensemantic perception and pixel-level reconstruction. Bridging this gap re-quires overcoming two core challenges: endowing semantic encoders withhigh-fidelity reconstruction capabilities, and effectively aligning gener-ative models with semantic spaces without relying on external teach-ers. To this end, we propose a novel unified multimodal framework fea-turing Semantic-Pixel self-alignment and Adaptive Routing (SPAR).First, to reconcile semantic perception with pixel-level reconstruction, weintroduce an asymmetric dual-stream unified tokenizer. A lightweightsemantic stream anchors discriminative features, while a Transformer-augmented pixel stream recovers fine-grained visual details into a uni-fied compact latent space. Second, to eliminate external dependencies, wepropose a self-aligned generation paradigm that natively leverages thisoptimized tokenizer as an internal alignment teacher for the diffusionmodel. Furthermore, to facilitate flexible multimodal interaction withinthis unified space, we introduce Dynamic Token Routing, which enableseach token to adaptively aggregate multi-layer MLLM features based onits distinct semantic demands. Extensive experiments demonstrate thatSPAR establishes the state-of-the-art for unified architectures, achiev-ing exceptional generation and reconstruction quality while preservingfoundational visual understanding capabilities.

Visual analogy learning enables image editing via demonstration rather than textual description, allowing users to specify complex transformations difficult to articulate in words. Given a triplet {a, a′, b}, the goal is to generate b′ such that a : a′ :: b : b′. Recent methods adapt text-to-image models with a single Low-Rank Adaptation (LoRA) module, but they face a fundamental limitation: attempting to capture the diverse space of visual transformations within a fixed module constrains generalization. Inspired by recent work showing that LoRAs in constrained domains span meaningful, interpolatable semantic spaces, we propose LoRWeB, which specializes the model for each analogy task in a single inference pass. LoRWeB dynamically composes learned transformation primitives, informally, choosing a point in a “space of LoRAs”. We introduce two key components: (1) a learnable basis of LoRAs to span the space of different visual transformations, and (2) a lightweight encoder that dynamically weighs these basis LoRAs given the input analogy pair. Comprehensive evaluations demonstrate state-of-the-art performance and significantly improved generalization to unseen transformations. Our findings suggest LoRA basis decompositions are a promising direction for flexible visual manipulation tasks. See our website for code.