ECCV 2026 Papers

Video Temporal Grounding (VTG) aims to localize specificvideo segments corresponding to natural language queries. While recentLarge Vision-Language Models (LVLMs) employ Reinforcement Learn-ing to generate Chains-of-Thought (CoT), they typically rely solely onoutcome-based supervision. Consequently, this often leads to halluci-nations, where the reasoning process becomes disconnected from thevisual content and the final prediction. Existing attempts to mitigatethis by relying on external supervision from larger models or separatereward models are computationally expensive and prone to rigid pat-terns. To address these challenges, we propose TAR (Temporal Anchor-Constrained Reasoning), a framework that introduces the temporal an-chor (T-anchor) as a transparent and auditable checkpoint mechanism.T-anchor enforces progressive refinement within the CoT, compellingthe model to continuously ground its intermediate thoughts in visual evi-dence and iteratively calibrate temporal predictions, thereby significantlyenhancing the faithfulness and autonomy of the reasoning process and fi-nal accuracy. Furthermore, we introduce a bootstrapping paradigm thatautomatically harvests high-quality CoT data using only a standard 7Bmodel, eliminating the dependency on ultra-large models. Extensive ex-periments demonstrate that TAR achieves state-of-the-art performanceand generates faithful, autonomous, and progressively refined reasoningtraces.

Recent progress in image and video synthesis has inspiredtheir use in advancing 3D scene generation. However, we observe thattext-to-image and -video approaches struggle to maintain scene- andobject-level consistency beyond a limited environment scale without apersistent, explicit geometric representation. We thus present a geometry-first approach that decouples this complex problem of large-scale 3Dscene synthesis into its structural composition, represented as a meshscaffold, and realistic appearance synthesis, which leverages powerful im-age synthesis models conditioned on the mesh scaffold. From an inputtext description, we first construct a mesh capturing the environment’sgeometry (walls, floors, etc.), and then use image synthesis, segmentationand object reconstruction to populate the mesh structure with objects inrealistic layouts. This mesh scaffold is then rendered to condition imagesynthesis, providing a structural backbone for consistent appearance gen-eration. This enables scalable, arbitrarily-sized 3D scenes of high objectrichness and diversity, combining robust 3D consistency with photoreal-istic detail. We believe this marks a significant step toward generatingtruly environment-scale, immersive 3D worlds.

Accurate survival modeling using whole slide images (WSI) is crucial for guiding cancer treatment and improving patient outcomes, yet it remains challenging due to the gigapixel resolution, small data cohorts, and limited annotations. In recent years, pathology foundation models (FM) have shown impressive performance improvements over the previous benchmarks set by the ImageNet-based backbones across a variety of WSI analysis tasks. However, the performance of individual FMs, especially in a secondary analysis task such as survival modeling, varies depending on the dataset, without a clear consensus for a superior FM. With over 25 publicly available pathology FMs so far, in this work, we address the growing challenge of model selection for survival analysis in digital pathology. We propose SurvMILKD, the first multi-teacher knowledge distillation (MKD) framework for WSIbased survival analysis. To efficiently enable MKD with high resolution of the WSIs, we integrate the weakly supervised multiple instance learning (MIL) adapters in our framework. The objective of the framework is to distill the complementary task-specific knowledge from multiple teacher FMs into the student FM, which, during inference, has collective knowledge of all FMs involved during training. Additionally, we introduce the risk-aware distillation loss to distill outcome-specific knowledge from multiple teachers into the student model. Our experiments on eight public datasets (including two external evaluation datasets) demonstrate that the proposed approach outperforms the individual FMs, their ensemble, as well as the prior MKD approaches that we benchmarked on survival analysis for the first time in literature. The codes are available at https://github.com/AIMLab-UBC/SurvMILKD.

We study dynamic Gaussian Splatting from monocular videos.While recent advancements in dynamic Gaussian Splatting offer a promis-ing foundation for modeling dynamic scenes, they often overfit to thetraining views and fail under occlusion or complex scene motion due tothe lack of reliable regularization signals in under-constrained regions.We propose Semantic Motion Graph (SMG), a novel approach that mod-els the Gaussian motion as the semantic motion. Our key insight is thatthe real-world scene motion is often structured by semantic coherence:regions that are spatially close and semantically related tend to exhibitconsistent dynamics. To leverage this prior, we build SMG to modelstructured motion of the scene. The Gaussian motion is driven by the mo-tion of SMG nodes. We further observe that the uncertainty of Gaussianmotion arises from both unreliable off-the-shelf priors and weakly con-strained regions during optimization. SMG addresses this by using reli-able graph nodes to guide the motion of nearby unreliable nodes. To eval-uate dynamic Gaussian Splatting under challenging real-world scenarios,we introduce a new multiview dataset collected under an ego-exo setup.Extensive experiments demonstrate that SMG achieves state-of-the-artperformance on monocular dynamic Gaussian Splatting across challeng-ing real-world benchmarks. Project: https://smg-gaussian.github.io/.

Sign language video generation demands precise hand andfacial articulation, yet modern video diffusion models, trained predom-inantly on spoken-language video, produce artifacts that render signingunintelligible. We propose SignRefine, a sign language video genera-tion model that produces comprehensible signing from 2D keypoint con-ditioning alone, generalizing across appearances and visual conditions.Our approach builds on a pretrained video diffusion transformer and in-troduces local adapters with spatial grounding to selectively refine handand face regions, steering the strong base model’s prior toward accu-rate articulation. To enable this work and support broader sign languageresearch, we present NVSign, a large-scale dataset of video content na-tively produced in sign language, offering diverse signer appearances,environments, and natural conversational settings. Trained on this data,our model shows up to 30% improvement in hand pose precision metricsover the strongest baseline and is preferred by sign language users forvisual quality and comprehensibility in more than 80% of comparisons.

Large Vision-Language Models (LVLMs) excel at multimodaltasks but remain prone to object hallucinations. Prior training-free reme-dies often uniformly strengthen visual signals, which may also amplifyirrelevant regions and introduce spurious evidence, harming fluency. Wepropose Context-aware Attention Intervention (CAI), a training-free inference-time mechanism that enforces a see only when needed prin-ciple via two-axis selectivity: where to look and when to intervene. Ateach decoding step, CAI derives token-specific visual relevance fromearly-layer representations to localize semantically aligned regions, andapplies a conservative, entropy- and depth-gated attention tilt only foruncertainty-spiking tokens in deeper layers where visual grounding de-grades, leaving confident tokens and irrelevant regions largely unchanged.This targeted intervention strengthens visual grounding while preserv-ing linguistic fluency, and it yields consistent improvements even with-out contrastive decoding, which remains optional as an auxiliary bias-suppression module. Extensive experiments across multiple LVLM back-bones and benchmarks show that CAI achieves state-of-the-art halluci-nation mitigation, and our analysis characterizes CAI as a KL-minimalattention reweighting with bounded interference under inactive gates orsmall tilts. Code is available at https://github.com/Iris1946/CAI.

We propose triangular consistency as a first-principled con-straint for optical flow, which is agnostic to network architecture, super-vision type, and dataset, and applies to both image-pair and multi-framesettings. This simple but powerful constraint is to compose two flows toinduce a third flow and enforce consistency among the three. The com-posed flows may arise from (i) image pairs, yielding cycle consistency; (ii)multiple video frames, producing longer-range motion through temporalchaining; or (iii) image pairs combined with controlled synthetic transfor-mations, which becomes data augmentation. This triangular consistencyintroduces negligible computational overhead and requires no additionalannotations. Since it is derived directly from the geometry of opticalflow, it does not rely on model-specific assumptions and serves as a “uni-versal” plug-and-play component for optical flow training. Experimentsshow consistent improvement across supervised, unsupervised, and trans-fer learning settings. Code: https://github.com/lsuvision/tri-flow.

Scaling monocular 3D Gaussian Splatting (3DGS) SLAM to kilometer-level outdoor environments poses two tightly coupled challenges: fragile long-term pose tracking and excessive memory overhead during large-scale mapping. In this paper, we propose KiloGS-SLAM, a highly efficient and robust monocular 3DGS-SLAM system that jointly addresses both bottlenecks. Since high-fidelity scene reconstruction fundamentally relies on drift-free camera poses, we first introduce a motionadaptive hybrid tracking module. This module features a conditiontriggered three-tier solving pipeline. It dynamically switches between Essential matrix and PnP models to handle geometric degeneracies. An ondemand foundation model can also be activated to rescue the trajectory from catastrophic drift. To ensure the system can sustain these long trajectories without memory exhaustion, we subsequently design a lifecyclemanaged Gaussian mapping strategy. By integrating probabilistic initialization with chunk-based multi-view densification and pruning, this full-pipeline optimization effectively reduces primitive redundancy while preserving high-frequency details. Together, the robust tracking guarantees the geometric foundation required for accurate mapping, while the memory-efficient lifecycle-managed mapping enables large-scale operation. Extensive experiments across three challenging outdoor datasets demonstrate that our approach achieves state-of-the-art tracking accuracy and rendering quality, successfully scaling to sequences of over 10,000 frames on a single GPU. Project Page: https://3dagentworld. github.io/KiloGS-SLAM/

Multimodal large language models (MLLMs) have demon-strated powerful capabilities in general spatial understanding and rea-soning. However, their fine-grained spatial understanding and reasoningcapabilities in complex urban scenarios have not received significant at-tention in the fields of both research and industry. To fill this gap, wefocus primarily on road markings as a typical example of fine-grainedspatial elements under urban scenarios, given the essential role of theintegrated road traffic network they form within cities. Around roadmarkings and urban traffic systems, we propose RoadBench, a system-atic benchmark that comprehensively evaluates MLLMs’ fine-grainedspatial understanding and reasoning capabilities using Bird’s-Eye View(BEV) and First-Person View (FPV) image inputs. This benchmarkcomprises eight tasks consisting of 3,040 strictly manually verified testcases, constructed from 2,137 unique BEV images and 721 unique FPVimages collected from five Chinese cities with relatively consistent trafficconventions. These tasks form a systematic evaluation framework thatbridges understanding at local spatial scopes to global reasoning. Theynot only test MLLMs’ capabilities in recognition, joint understanding,and reasoning but also assess their ability to integrate image informationwith domain knowledge. After evaluating 20 mainstream MLLMs, weconfirm that RoadBench is a challenging benchmark for MLLMs whilerevealing significant shortcomings in existing MLLMs’ fine-grained spa-tial understanding and reasoning capabilities within urban scenarios. Incertain tasks, their performance even falls short of simple rule-basedor random selection baselines. These findings, along with RoadBenchitself, will contribute to the comprehensive advancement of spatial un-derstanding capabilities for MLLMs. The benchmark code is availableat https://github.com/tsinghua-fib-lab/RoadBench, and the supple-mentary material provides example data, prompts, evaluation scripts,and raw evaluation results.Road MarkingsBEV Image # of Lanes: 4Road MarkingsFPV Image # of Lanes: 4Fig. 1: Examples of road markings in BEV and FPV images. Road markings such aslane dividers and turning arrows provide fine-grained visual cues for lane counting andlane designation recognition.

Offline feature extraction with pretrained encoders followedby multiple instance learning (MIL) aggregation is still the dominantparadigm for whole slide image (WSI) analysis, but the domain gap be-tween natural and histopathology images limits representation quality.End-to-end optimization can reduce this gap by training the encoder withslide-level objectives, but it often suffers from (i) cost-driven fine-grainedand morphological information loss caused by low-magnification small in-puts and random patch sampling, and (ii) sparse supervision from slidelabels alone. We propose PyraE2E, an end-to-end framework that turnsthe intrinsic multi-resolution WSI pyramid into a dense self-supervisionsignal via super-resolution (SR). For each sampled low-resolution (LR)patch at low magnification, we use its spatially aligned high-resolution(HR) patch at higher magnification as a reconstruction target, providingcross-scale pixel-level supervision while keeping computation boundedby operating on LR inputs. PyraE2E combines an embedded Cluster-Score Sampling module to select informative LR regions with a sharedGlobal–Local Partitioned (GLP) encoder whose HR-informed featuresare jointly optimized by an SR reconstruction head and a slide-level pre-diction head. Joint training couples dense cross-scale supervision withslide-level objectives, improving end-to-end representations and down-stream WSI prediction under controlled computational cost.

Point-Based unsigned distance fields (UDFs) learning oftenaggregates Euclidean neighborhoods as unstructured point sets. Eachquery is typically supervised by a single distance value. This offers lit-tle relational guidance for separating nearby surfaces, reducing featurediscriminability for distance regression. We propose PrimitiveUDF, aprimitive-based UDF framework with two coupled components: projection-induced neighborhood reorganization and hybrid primitive construction.The former uses spherical projection and a sphere-to-plane mapping toobtain a stable local chart for assessing surface-relative relations. The lat-ter grows triangle primitives with continuity and non-crossing preferences,and prunes direction-consistent segments for thin or sparse structures.Primitives are then encoded and aggregated by an attention-based re-gressor for query-conditioned UDF prediction. Experimental results onsynthetic and real-world datasets demonstrate around 10% improvementon scene-level benchmarks and consistent cross-scene gains, with balancedrobustness under noise and sparsity conditions. Code: PrimitiveUDF.

Spatiotemporal prediction of complex dynamical systems de-pends not only on historical observations but also on evolving exter-nal control parameters. Crucially, for ionospheric Total Electron Con-tent, existing concatenation-based conditioning fails to capture region-dependent responses driven by space weather.We propose PMGC-SimVP, a spatiotemporal prediction framework withParametric Multi-scale Gated Convolution. PMGC injects external con-trol parameters into local dynamics via latitudinally structured mod-ulation, enabling explicit modeling of non-stationary evolution and re-gional heterogeneity. Built upon a purely convolutional backbone anda low-rank temporal adapter, PMGC-SimVP remains computationallye_x001E_cient.We evaluate PMGC-SimVP on global TEC forecasting, where dynam-ics are strongly driven by space weather and exhibit pronounced lati-tudinal variability. The proposed method achieves consistent improve-ments over recent state-of-the-art models, including Transformer-basedand frequency-domain approaches, across overall evaluation settings andduring geomagnetic disturbances. Experiments on generic video predic-tion benchmarks further verify the robustness of the proposed backboneeven without parameter injection.

Driven by the growing capacity and training scale, Text-to-Video (T2V) generation models have recently achieved substantialprogress in video quality, length, and instruction-following capability.However, whether these models can understand physics and generatephysically plausible videos remains a question. While Vision-LanguageModels (VLMs) have been widely used as general-purpose evaluators invarious applications, they struggle to identify the physically impossiblecontent from generated videos. To investigate this issue, we construct aPID (Physical Implausibility Detection) dataset, which consists of a testsplit of 500 manually annotated videos and a train split of 2,588 pairedvideos, where each implausible video is generated by carefully rewritingthe caption of its corresponding real-world video to induce T2V modelsproducing physically implausible content. With the constructed dataset,we introduce a lightweight fine-tuning approach, enabling VLMs to notonly detect physically implausible events but also generate textual expla-nations on the violated physical principles. Taking the fine-tuned VLMas a physical plausibility detector and explainer, namely PhyDetEx,we benchmark a series of state-of-the-art T2V models to assess their ad-herence to physical laws. Our findings show that although recent T2Vmodels have made notable progress toward generating physically plau-sible content, understanding and adhering to physical laws remains achallenging issue, especially for open-source models. Our dataset, train-ing code, and checkpoints are available at https://github.com/Zeqing-Wang/PhyDetEx.

Despite the remarkable capabilities of Large Vision-LanguageModels (LVLMs), they remain highly susceptible to hallucinations. Exist-ing inference-time mitigation strategies largely rely on visual attention asa diagnostic proxy. However, we demonstrate that this approach is insuf-ficient in certain cases due to the “anchoring trap,” where parameterizedlinguistic priors override correct visual grounding. In this paper, we pio-neer a representation geometry perspective, identifying hidden states asa complementary diagnostic signal beyond visual attention. We modelthe evolution of hidden states across layers and generation steps as con-strained sequences governed by a static visual subspace and a dynamiclinguistic subspace within a high-dimensional Euclidean space. We revealthat hallucinations inherently stem from abrupt structural deviations,specifically, the spontaneous departure of latent representations from thevisual subspace and their subsequent drift into the linguistic domain.To quantify this phenomenon, we propose the Geo-Semantic TurbulenceIndex (GSTI) to adaptively monitor visual misalignment and semanticvolatility. Building upon GSTI, we introduce Orthogonal Residual-BasedIntervention (ORBIT), an architecture-agnostic, plug-and-play mecha-nism that recalibrates divergent representation sequences by employinga closed-loop intervention encompassing residual extraction, visual re-anchoring, and norm-preserving injection of suppressed visual evidence.Extensive experiments across diverse LVLM architectures demonstratethat ORBIT significantly mitigates hallucinations and enhances generalvision-centric performance with exceptional computational efficiency.

Temporal Action Localization (TAL) typically relies on seg-ment annotations or offline access to full videos, limiting scalability andonline use. We introduce Point-Supervised Online TAL (POTAL), whichlocalizes actions in streaming videos using only one temporal point perinstance. To solve POTAL, we propose OnPoint, an offline-to-onlinemulti-level distillation framework that transfers knowledge from a point-supervised offline teacher to an online student via (i) pseudo-segmentinstance distillation, (ii) class-activation sequence distillation, and (iii)anticipatory window-level distillation. We further improve robustness byincorporating the original point labels into student training and by re-fining anchor decoding with actionness-guided attention calibration. Ex-periments on five datasets show OnPoint consistently outperforms strongbaselines, establishing a solid foundation for POTAL. †

Test-time scaling is a paradigm where large models use ad-ditional compute at inference to achieve better performance, withoutchanging model weights. While it has been widely studied for Large Lan-guage Models (LLMs), its applicability to Large Vision-Language Mod-els (LVLMs) remains less explored and analyzed, with limited analysis ofwhether, when, and to what extent these approaches transfer to LVLMs.In this work, we ask a simple but fundamental question: can conven-tional test-time scaling methods developed for LLMs be directly appliedto LVLMs? We present the first comprehensive study of test-time scal-ing for LVLMs, spanning multiple models and model sizes, nine test-timescaling methods, and six diverse benchmarks. Our main findings is that1) different from previous findings, small, well-performing models benefitthe most from test-time scaling, enabling performance improvements ofup to around 30%, reaching large models performance, and often out-performing them, 2) LVLMs lose focus when given more compute thannecessary, and 3) Visual information is encoded early in the reasoningchain, after which the chain is dominated by text-only reasoning andthe contribution of image tokens drops significantly. Finally, we also pro-vide a global and fine-grained analysis on the quality and informationsufficiency of the reasoning chains produced. Overall, our findings andanalysis provide practical guidance and insights into LVLMs and theirdeployment in research and industry.

Multimodal large language models (MLLMs) have achieved impressiveprogress in vision-language reasoning, yet their ability to understand temporallyunfolding narratives in videos remains largely underexplored. Narrative under-standing requires more than recognizing isolated events: models must maintaincoherent representations of who is doing what, when, and where across scene tran-sitions and temporal gaps. We introduce NARRATIVE T RACK, the first benchmarkto evaluate narrative understanding in MLLMs through fine-grained entity-centricreasoning. Unlike existing benchmarks limited to short clips or coarse scene-levelsemantics, we decompose videos into constituent entities and evaluate modelsusing a Compositional Reasoning Progression (CRP), a structured frameworkthat progressively increases narrative complexity across three dimensions: en-tity existence, entity changes, and entity ambiguity. This progression requiresmodels to move beyond local perception to reasoning about entities’ temporalpersistence, state changes, and fine-grained perceptual disambiguation. To enablescalable benchmark construction, we develop a fully automated entity-centricpipeline that extracts temporally grounded entity representations and providesthe foundation for CRP. Evaluations of state-of-the-art MLLMs reveal that ex-isting models struggle to maintain coherent entity representations under visualtransitions and temporal dynamics. Open-source general-purpose MLLMs exhibitstrong perceptual grounding but weak temporal continuity, while video-specializedMLLMs capture temporal context yet frequently hallucinate entities’ contexts.These findings uncover a fundamental trade-off between perceptual grounding andtemporal reasoning, indicating that narrative understanding emerges only fromtheir integration. NARRATIVE T RACK provides the first systematic framework todiagnose and advance temporally grounded narrative comprehension in MLLMs.

Zero-shot video moment retrieval (ZMR) aims to overcomethe limitations of traditional approaches that require large-scale datasetsannotated with text and its relevant temporal spans. Despite advances inpre-trained vision–language models (VLMs) and multimodal large lan-guage models (MLLMs), existing ZMR methods still heavily depend onquery-to-video content similarity, making them vulnerable to modal-ity and language-style gaps. These gaps lead to unreliable span pro-posals and unstable moment retrieval results. To address this issue, wepropose Self-Similarity-based Moment proposal and Scoring (Self-SiMS)that instead exploits intrinsic relationships within videos, enabling ro-bust span generation and scoring. By deriving self-similarity only fromthe video content, we circumvent the noisy and mismatched patternsof query–frame or query–caption similarities, thereby mitigating bothmodality and language-style gaps. Furthermore, we introduce a query-aware MLLM-based reasoning stage to further sharpen alignment be-tween text and video. Extensive experiments demonstrate that Self-SiMSachieves the state-of-the-art performance across ZMR benchmarks.

Unified multimodal models have demonstrated remarkableversatility in multimodal understanding and generation. However, theystruggle with image generation under complex compositional instructions,a setting frequently encountered in real-world applications. To addressthis, we present MANGO, a novel model that unleashes image generationcapability at both its performance floor and ceiling. (1) To raise theperformance floor, we introduce Functionality-Oriented Transformers,an architectural design that assigns each Transformer branch to a dedi-cated function, such as image generation or image understanding. Thisfunction-disentangled design alleviates cross-functional interference inher-ent in mainstream modality-oriented architectures, thereby unleashingthe model’s image generation capacity. (2) To elevate the performanceceiling, we propose Paint-CoT, a framework that enables image generationthrough a human-like artistic chain of thought, consisting of planning,acting, reflection, and correction. Leveraging the model’s multimodalunderstanding and reasoning capabilities, this reasoning-guided processfurther unleashes image generation capacity. Further, to overcome the dif-ficulty and high cost of collecting multi-step aligned chain-of-thought datafor end-to-end training, we develop a multi-task joint training paradigmthat decomposes the full reasoning-generation pipeline into manageablesubtasks. This strategy maximizes the utilization of existing or readilyconstructed supervision, enabling e!ective Paint-CoT-based image gener-ation without requiring fully aligned multi-step annotations. Extensiveexperiments demonstrate that MANGO consistently achieves superiorperformance across diverse image generation benchmarks, delivering sub-stantial improvements in complex image generation scenarios.