ECCV 2026 论文列表

Multimodal Large Language Models (MLLMs) have shownimmense promise in universal multimodal retrieval, which aims to findrelevant items of various modalities for a given query. However, theirpractical application is often hindered by the substantial computationalcost incurred from processing a large number of tokens from visual inputs.In this paper, we propose Magic-MM-Embedding, a series of novel modelsthat achieve both high efficiency and state-of-the-art performance inuniversal multimodal embedding. Our approach is built on two synergisticpillars: (1) a highly efficient MLLM architecture incorporating visualtoken compression to drastically reduce inference latency and trainingtime, and (2) a multi-stage progressive training strategy designed tonot only recover but significantly boost performance. This coarse-to-finetraining paradigm begins with extensive continued training to restoremultimodal understanding and generation capabilities, progresses tolarge-scale contrastive pretraining and hard negative mining to enhancediscriminative power, and culminates in a task-aware fine-tuning stageguided by an MLLM-as-a-Judge for precise data curation. Comprehensiveexperiments show that our model outperforms existing methods by alarge margin while being more inference-efficient.

Sparse autoencoders (SAEs) have recently been proposed asinterpretable tools for concept-level manipulation, under the assumptionthat isolated features can serve as controllable intervention points. In thiswork, we systematically evaluate this assumption in the context of ob-ject erasure and steering in diffusion models. We show that while SAEsreliably detect and localize semantic concepts within diffusion modelactivations, direct intervention in their latent space frequently inducesout-of-distribution activations, resulting in severe visual artifacts. To dis-entangle detection from intervention, we use SAE activations purely assemantic detectors to identify image regions containing the target object,and replace those patch embeddings with the ones that do not containit. This detection-based replacement preserves the diffusion model’s ac-tivation statistics and produces significantly cleaner erasure results thanlatent steering. Our findings reveal a fundamental gap between conceptdetection and concept intervention in diffusion models: monosemanticor sparse features are not inherently suitable as control knobs for steer-ing. These results position SAEs as powerful interpretability tools foranalyzing generative models, but highlight important limitations whenused for direct manipulation, such as unlearning. We release the code athttps://eidoslab.github.io/PER/.

World models of embodied agents predict future observationsconditioned on an action taken by the agent. For complex embodiments,action spaces are high-dimensional and difficult to specify: for example,precisely controlling a human agent requires specifying the motion ofeach joint. This makes the world model hard to control and expensiveto plan with as search-based methods like CEM scale poorly with actiondimensionality. To address this issue, we train a lightweight policy thatmaps high-level actions to sequences of low-level joint actions. Composingthis policy with the frozen world model produces a lifted world model thatpredicts a sequence of future observations from a single high-level action.We instantiate this framework for a human-like embodiment, defining thehigh-level action space as a small set of 2D waypoints annotated on thecurrent observation frame, each specifying a near-term goal position for aleaf joint (pelvis, head, hands). Waypoints are low-dimensional, visuallyinterpretable, and easy to specify manually or to search over. We showthat the lifted world model substantially outperforms searching directlyin low-level joint space (3.8× lower mean joint error to the goal pose),while remaining more compute-efficient and generalizing to environmentsunseen by the policy.

Vision Language Models (VLMs) have achieved strong per-formance across diverse video understanding tasks. However, their view-point invariant training limits their ability to understand egocentricproperties (e.g., human-object interactions) from exocentric video ob-servations. This limitation is critical for applications such as Activi-ties of Daily Living (ADL) monitoring, where understanding egocen-tric properties is essential yet egocentric cameras are impractical to de-ploy, making it impossible to simply collect egocentric data at test time.To address this challenge, we propose Ego2ExoVLM, a VLM frame-work that learns to infer egocentric properties from exocentric videos.Our key insight is that time-synchronized ego-exo video pairs can beleveraged during training, where the egocentric viewpoint provides priv-ileged supervision (rich egocentric signal available only at training time).Ego2ExoVLM achieves this through two components: Ego2Exo Se-quence Distillation, which transfers egocentric reasoning through alanguage-level sequence distillation objective, and Ego Adaptive Vi-sual Tokens, which encourages the model to surface relevant interac-tion cues within exocentric video representations. To measure this ca-pability, we introduce Ego-in-Exo Perception, a benchmark designedto evaluate the understanding of egocentric properties from exocentricvideos. Ego2ExoVLM is evaluated on 10 tasks across Ego-in-Exo Per-ception and existing ADL benchmarks, achieving state-of-the-art resultson the ADL-X benchmark suite and outperforming strong baselines onour proposed benchmark. Code, models, and data will be released athttps://github.com/dominickrei/EgoExo4ADL.

Integrated circuit (IC) footprint geometry labeling refers tothe process of converting pin diagrams in IC datasheets into machine-readable geometric parameters. This task is critical in printed circuitboard (PCB) design and component assembly, as accurate labeling en-sures proper IC placement and reliable connectivity. The process is chal-lenged by unstructured annotations, complex footprint arrangements,and abstract geometric diagrams, making fully automated labeling meth-ods inadequate. Traditional EDA tools require heavy manual input. Ex-isting automation methods, such as OCR or object detection, fail tocapture the implicit geometric relationships in IC diagrams, leaving thelabeling task incomplete. Recent work has shown that end-to-end largemultimodal models (LMMs) can perform IC geometry labeling. How-ever, by treating the task as a black box, it is prone to shortcut learningand lack interpretability. In this work, we introduce ICLAgent, thefirst multi-agent framework for fully automated IC footprint geometrylabeling that explicitly models the workflow of expert engineers to pro-duce more interpretable and reliable labeling outcomes. Furthermore, wepresent ICAgent-Instruct, the first dynamic planning and reasoningdataset tailored for IC footprint geometry labeling. Extensive experi-ments show that ICLAgent improves overall accuracy by 10.3% com-pared to the previous SOTA method and by 79.5% compared to man-ual annotation. Despite using only simple supervised fine-tuning on a7B model (Qwen2-VL-7B), ICLAgent surpasses general-purpose LMMssuch as GPT-5 (by 94.6%) and Gemini-3 Pro (by 15.2%).

Embodied agents must explore partially observed environ-ments while maintaining reliable long-horizon memory. Existing graph-based navigation systems improve scalability, but they often treat un-explored regions as semantically unknown, leading to inefficient frontiersearch. Although vision-language models (VLMs) can predict frontiersemantics, erroneous predictions may be embedded into memory andpropagate through downstream inferences, causing structural error ac-cumulation that confidence attenuation alone cannot resolve. These ob-servations call for a framework that can leverage semantic predictionsfor directed exploration while systematically retracting errors once newevidence contradicts them. We propose Hypothesis Graph Refinement(HGR), a framework that represents frontier predictions as revisable hy-pothesis nodes in a dependency-aware graph memory. HGR introduces(1) semantic hypothesis module, which estimates context-conditioned se-mantic distributions over frontiers and ranks exploration targets by goalrelevance, travel cost, and uncertainty, and (2) verification-driven cas-cade correction, which compares on-site observations against predictedsemantics and, upon mismatch, retracts the refuted node together withall its downstream dependents. Unlike additive map-building, this allowsthe graph to contract by pruning erroneous subgraphs, keeping mem-ory reliable throughout long episodes. We evaluate HGR on multimodallifelong navigation (GOAT-Bench) and embodied question answering (A-EQA, EM-EQA). HGR achieves 72.41% success rate and 56.22% SPL onGOAT-Bench, and shows consistent improvements on both QA bench-marks. Diagnostic analysis reveals that cascade correction eliminates ap-proximately 20% of structurally redundant hypothesis nodes and reducesrevisits to erroneous regions by 4.5×, with specular and transparent sur-faces accounting for 67% of corrected prediction errors.

Document parsing is a fundamental task in multimodal understanding, supporting a wide range of downstream applications such as information extraction and intelligent document analysis. Benefiting from strong semantic modeling and robust generalization, VLM-based end-to-end approaches have emerged as the mainstream paradigm in recent years. However, these models often suffer from substantial inference latency, as they must autoregressively generate long, full-page sequences when processing long-form documents. While recent hybrid methods mitigate this issue via region-level parallel decoding with VLMs, independent region decoding loses full-page context and might weaken global coherence. To address this issue, we propose Hierarchical Speculative Decoding (HSD), a two-stage local-to-global framework for document parsing. HSD first employs a lightweight pipeline drafter to predict region partitions and generate coarse drafts for each region. The first stage verifies the generated region-level drafts in parallel for efficiency, while the second stage further performs page-level verification on these refined outputs to preserve full-page coherence. Experimental results show that our HSD achieves a near-lossless 2.78× speedup with HunyuanOCR on OmniDocBench v1.5 and up to 7.04× speedup on long-document parsing tasks, demonstrating the effectiveness of our proposed method. The code is available at https://github.com/whlscut/HSD.

Real-world semi-supervised learning (SSL) often encounterssignificant challenges with long-tailed label distributions and noisy pseudo-labels, which hinder generalization and amplify confirmation bias. In thiswork, we introduce a novel framework, Gaussian Bridge Consistency(GBC), to address these challenges by constructing semantic interpo-lation paths between unlabeled samples and high-quality class anchors.Our method maintains a dynamic Prototype Atlas that stores a diverseand evolving set of labeled and pseudo-labeled exemplars per class. Foreach unlabeled instance, GBC forms a class-conditional Gaussian Fea-ture Bridge in the latent space, enabling the student model to traversea smooth trajectory from uncertain predictions to reliable class proto-types. A bridge consistency loss is applied along this path to enforcealignment with a geometrically interpolated target distribution. Further-more, we propose BridgeMix, a confidence-aware feature mixing strategythat interpolates both sample and anchor pairs to amplify cross-samplegeneralization. Extensive experiments on CIFAR10-LT and ImageNet-LT (USB benchmarks) validate the robustness and effectiveness of GBCunder realistic long-tailed SSL settings, consistently improving long tail-class performance without sacrificing scalability.

Existing tampering benchmarks largely rely on object masks, which severely misalign with the true edit signal: many pixels inside a mask are untouched or only trivially modified, while subtle yet consequential edits outside the mask are treated as natural. We reformulate VLM image tampering from coarse region labels to a pixel-grounded, meaning and language-aware task. First, we introduce a taxonomy spanning edit primitives (replace/remove/splice/inpaint/attribute/colorization, etc.) and their semantic class of tampered object, linking low-level changes to high-level understanding. Second, we release a new benchmark with perpixel tamper maps and paired category supervision to evaluate detection and classification within a unified protocol. Third, we propose a training framework and evaluation metrics that quantify pixel-level correctness with localization to assess confidence or prediction on true edit intensity, and further measure tamper meaning understanding via semantics-aware classification and natural language descriptions for the predicted regions. We also re-evaluate the existing strong segmentation/localization baselines on recent strong tamper detectors and reveal substantial overand underscoring using mask-only metrics, and expose failure modes on micro-edits and off-mask changes. Our framework advances the field from masks to pixels, meanings and language descriptions, establishing a rigorous standard for tamper localization, semantic classification and description. Code and data are available at https://github.com/VILA-Lab/PIXAR.

Generalization remains a critical bottleneck in AI-generatedimage detection. Because many modern generators are proprietary oradversarially modified, existing detectors overfit to the low-level tex-tural patterns of accessible training data, resulting in severe failureson unseen domains. Conventional regularization techniques (e.g., L1 /L2norms, Dropout) apply indiscriminate parametric constraints and fail toprovide the domain-invariant structure necessary for cross-generator ro-bustness. To address this, we propose Feature-Augmented Implicit Reg-ularization (FAIR). FAIR introduces an orthogonal, macro-structuralprior, specifically, Scene Composition Structure (SCS), during train-ing to geometrically constrain the model’s optimization trajectory. Byaugmenting the primary feature space with domain-invariant SCS fea-tures, FAIR explicitly penalizes texture-biased shortcut learning. Cru-cially, this structural prior is entirely discarded at inference, yielding asmoothed, generalized decision boundary with zero architectural or com-putational overhead. Extensive evaluations across five massive bench-marks demonstrate that integrating FAIR into state-of-the-art detectorssignificantly improves cross-generator generalization, boosting accuracyby up to 8.04% and establishing new state-of-the-art robustness in zero-shot transfer scenarios.

Reasoning segmentation is a challenging vision–language taskthat performs pixel-level segmentation guided by language reasoning overimplicit textual descriptions. Existing methods typically employ largelanguage models (LLMs) to achieve such reasoning capability, yet theirsubstantial computational and memory demands limit practical deploy-ment. To overcome this, we present LReSeg, the first attempt to tacklereasoning segmentation with a small language model (SLM), achievingstrong performance with substantially reduced model size and computa-tional cost. To address the challenges of scaling reasoning segmentationto SLMs, we propose two key designs: First, unlike LLMs, SLMs havelimited capacity to provide sufficient spatial instruction cues for mask de-coding. We introduce register tokens that aggregate text-conditioned spa-tial features via the SLM’s self-attention and inject them into the visualtoken stream to enrich the mask decoder’s inputs. Second, we adopt aunified encoder architecture to eliminate the redundant visual backboneof conventional dual-encoder designs, naturally ensuring feature consis-tency between reasoning and mask decoding. LReSeg surpasses modelsof similar scale with superior efficiency, while being nearly 10× smaller(800M vs. ∼8B) and 3.3× faster than conventional 7B-scale LLM-basedmethods. Code is available at https://github.com/downdric/LReSeg.

Quantitative microstructural characterization is fundamen-tal to materials science, and electron micrographs (EMs) provide indis-pensable high-resolution insights. However, progress in deep learning-based analysis of EMs has been hampered by the scarcity of large-scale,expert-annotated public datasets. To address this issue, we introduceEM3M, a large-scale and multimodal dataset for instance-level under-standing of EMs. EM3M comprises 5,091 high-quality EMs, approxi-mately 3 million instance segmentation annotations, and image-leveltextual descriptions with disentangled attributes. The dataset is con-structed through a rigorous multi-stage curation and validation pipeline,with comprehensive statistical analyses to ensure reliability and repro-ducibility. Building upon these curated image-text pairs, we further pro-vide a text-to-image diffusion model that serves as a controllable dataaugmentation engine, demonstrating that synthetic augmentation con-sistently improves downstream segmentation performance. To establish asystematic benchmark, we evaluate representative instance segmentationmethods on EM3M. Our results reveal that conventional detection-basedand query-based methods struggle with the extreme instance densitiesand textural complexities inherent in EMs. We additionally provide anoptimized flow-based baseline to facilitate fair comparison and future re-search. EM3M1 , the generative engine2 , and an online demo3 are publiclyavailable to support future research in automated materials analysis.

Image retouching is commonly formulated as enhancing over-all visual quality through color adjustment, but in practice, it also servesto emphasize visual focus by guiding viewers’ attention toward a specificsubject or region. Achieving such focus-oriented retouching is inherentlychallenging, as it requires well-coordinated global and local adjustmentsto manipulate perceptual saliency while maintaining visual naturalness.This intricate process typically demands substantial professional exper-tise. In this study, we propose EyeControl, a multi-modal large lan-guage model (MLLM)-driven agent with a diffusion-based retouchingexecutor that enables visual focus enhancement under weak user intent.With only a few clicks or coarse strokes, EyeControl directs visual at-tention to the intended region, effectively “dotting the eye” of the image.The core idea is to explicitly link the weak user intention with the tar-get editing region and the corresponding tonal adjustment operationsduring retouching. To achieve this, the system first interprets the in-tent and image content to infer the visual focus and generate structuredintent guidance for the retouching executor. Second, the retouching ex-ecutor is encouraged to respond more strongly to the target region, ex-plicitly aligning its attention map with a designed pseudo-intent map.We also introduce an operation-consistency constraint to improve coor-dination between global and local adjustments, achieving more naturaland coherent retouching. Additionally, we contribute ControlArt-Bench,a high-quality evaluation dataset for visual focus enhancement. Exten-sive evaluations demonstrate that EyeControl yields perceptually ap-pealing results with stronger intent alignment. Code will be released athttps://github.com/DragonisCV/EyeControl.

Recent advances in large-scale vision–language models (VLMs)have greatly advanced open-vocabulary semantic segmentation (OVSS),enabling segmentation beyond predefined category constraints. However,due to the image-level pre-training objectives of VLMs, existing training-free OVSS methods often suffer from global contextual artifacts, degrad-ing their ability to capture fine-grained and spatially localized semantics.In this work, we propose a novel framework, termed CURE (Contextualdebiasing and Unbiased REfinement), to address these limitations bymitigating both explicit and implicit global biases. Technically, CUREemploys a reject-and-reconstruction module to identify explicit outliertokens and reconstruct them using local neighborhood information. Aglobal contextual debiasing module is further introduced to suppressresidual implicit bias. Finally, a spatial correlation refinement moduleis designed to enhance spatial coherence by leveraging mid-level simi-larity patterns. Extensive experiments on five datasets demonstrate thatCURE consistently improves segmentation quality. When integrated intosix baseline methods, it yields average mIoU gains of 2.8% and 7.7% withViT-B/16 and ViT-L/14 backbones, respectively.

Vector graphics are prized for their resolution independence,compact storage, and direct editability, making differentiable optimiza-tion of their parametric primitives an attractive goal. Yet classical ras-terization is discontinuous with respect to geometry, and existing reme-dies that smooth the forward pass demand increasingly elaborate heuris-tics as scene complexity grows. We trace this fragility to a gradientseesaw: design choices that improve forward geometric exactness cansystematically degrade the induced gradient signal, and vice versa. Tonavigate this tension we introduce CubicSplat, a differentiable vectorrasterizer that replaces Bézier closest-point solvers with uniform poly-line surrogates whose geometric error is bounded at O(S −2 ). The re-sulting static computation graph yields well-conditioned gradients byconstruction, while a compositing-derived visibility mechanism prunesdegenerate primitives without auxiliary regularization. On DIV2K andKodak benchmarks CubicSplat achieves state-of-the-art reconstructionquality with over 2 dB PSNR gain in the closed-fill setting, while train-ing up to 4× faster than prior methods. The code is available at https://github.com/CubicSplat/repo

Accurate yaw estimation is a bottleneck in cross-view lo-calization between ground view and Bird’s Eye View (BEV). Existingmethods couple yaw with translation and rely on height or projectionassumptions that degrade under large yaw ambiguity. We disentangleyaw from location accuracy and introduce LAYS, a radially invariantline-consensus voting method. By exploiting the radial invariance of ourformulation, we achieve sub-degree yaw precision via 3D voting over allcandidate poses, while eliminating the need for accurate location. Ourkey observation is that a ground-image column matched to BEV pixelsinduces the same yaw across all camera positions along the radial direc-tion of the pixels. LAYS matches BEV pixels to ground columns usingfeature similarity and accumulates the induced yaw votes into discrete3D bins, where correct correspondences along the radial line concentrateinto a sharp peak for the correct yaw. Experiments on Mapillary, Ford,KITTI, and VIGOR show significant gains under unknown yaw, particu-larly for normal FoV with unknown yaw (+28∼45%p), and using LAYSas a yaw prior improves downstream 3-DoF localization.

Continuous visual autoregressive (AR) models have demon-strated promising performance in image generation, but their inherentlysequential nature results in slow inference speed. Speculative decoding, asuccessful acceleration technique for large language models (LLMs), haseffectively accelerated discrete visual AR models. However, the absenceof an analogous theory for continuous distributions precludes its use inaccelerating continuous AR models. To fill this gap, this work presentscontinuous speculative decoding, and addresses challenges from: 1) lowacceptance rate, caused by inconsistent output distribution modeled bytarget and draft models, and 2) modified distribution without analyticexpression, caused by a complex integral. For challenge 1), we addresslow acceptance rates through an approximated criterion, a novel denois-ing trajectory alignment strategy based on reparameterization proximity,and token pre-filling. For challenge 2), we introduce acceptance-rejectionContinuous speculative decodingDiscrete speculative decoding Proposal distribution VerificationProposal distribution VerificationDraft model Target model Draft model Target modelY Inconsistency existsYModified distribution NModified distributionResample Modify & Modify & NNormalize NormalizeNo analytic expression Hard to integrateFig. 2: Comparison between discrete and continuous speculative decoding. Discretesituation offers the convenience of directly computing probabilities and simply samplingfrom modified distributions. In contrast, continuous situation faces challenges in theinconsistency of output distributions, leading to low acceptance criterion as well as lowacceptance rate, and the modified distributions without analytic expression, caused bycomplex integral.sampling algorithm with an appropriate upper bound, thereby avoid-ing explicitly calculating the integral. Furthermore, our denoising trajec-tory alignment is also reused in acceptance-rejection sampling, effectivelyavoiding repetitive diffusion model inference. Extensive experiments onvarious models at 256×256 and 512×512 resolutions demonstrate thatour approach achieves over 2× wall-time speedup while preserving theimage generation quality.

Egocentric vision systems capture human behavior from visi-ble cues, but overlook physiological indicators of autonomic states suchas stress, engagement, and attention. Heart rate variability (HRV) is awidely used noninvasive marker of autonomic regulation under stress.HRV reflects small timing differences between successive heartbeats andhas so far been out of reach for egocentric platforms, where motion andnoise in gaze video mask exactly this fine-grained timing. We proposeEgoHRV, a method that estimates HRV as well as heart rate (HR) fromthe gaze cameras that are already integrated into egocentric headsets. Ourpipeline combines a 3D backbone with a novel low–high decompositionmodule that extracts the blood volume pulse (BVP) signal from gazevideo. Our cross-domain pretraining aligns the frequency-domain repre-sentations of contact-based and camera-derived signals. This alignmentgives EgoHRV the temporal precision to recover HRV from the subtlefluctuations in gaze video. EgoHRV achieves state-of-the-art accuracy forHR and HRV estimation from egocentric video, and its uncertainty-awaredesign improves downstream behavioral modeling. Integrating our HRVestimates and confidence measures into EgoExo4D’s proficiency estimatorraises accuracy by 17.8%. Beyond skill, continuous HRV estimation alsoopens egocentric systems to stress- and arousal-aware estimation tasks.Code: https://github.com/eth-siplab/EgoHRV

Open-set object detection leverages language prompts toguide the perception of categories outside the training set. However,for categories that are difficult to describe or cannot be effectively ab-stracted semantically, models exhibit reduced receptiveness to guidanceprovided by textual descriptions. Recent approaches have explored vi-sual prompting, but they often lack inherent generalization capabilitiesand suffer from limitations in semantic, spatial, and contextual aware-ness during prompt construction. Additionally, isolated prompt path-ways hinder unified optimization. Inspired by the general process of hu-man object search, we designed Ceptor, a unified detector guided bydiverse prompts for open-set object detection. Specifically, we proposeInfusion Strategies to infuse our framework with the generalizationcapabilities of a vision-language model (VLM), leveraging its alignmentproperties and generalizable feature space to construct diverse promptsfor different object types and varying category distributions. Further-more, we introduce a novel prompt generation method, PRE, whicheffectively preserves generalization, contextual information, and spatialaccuracy. Harmonization strategies are also incorporated to ensurecoordinated optimization across various prompts. Ceptor achieves strongperformance with various prompts across multiple datasets, demonstrat-ing the effectiveness of our methodology. Models and code are releasedat https://github.com/jinyanglii/Ceptor.