ECCV 2026 Papers

Computer-use agents, which leverage multimodal large lan-guage models (MLLMs) to operate computers and complete tasks, haveattracted significant attention for their utility and versatility. A majorchallenge in developing these agents is collecting large-scale, high-qualitytrajectories. The standard approach generates synthetic data through aself-improving loop: an agent is placed in a verifiable environment anditeratively fine-tuned on its successful trajectories. Despite its effective-ness, this paradigm exploits only successful trajectories and discards thefailed ones, even though failures carry rich information about a model’sweaknesses. In this work, we explore a complementary failure-driven self-improvement loop, a data-centric paradigm that turns failed trajecto-ries into agent improvements. Specifically, we employ an LLM to diag-nose failure modes, propose inference-time solutions, and generate codepatches—lightly verified by humans—that upgrade the agent. We vali-date this approach with the state-of-the-art OpenCUA-72B model on theOSWorld benchmark, improving the success rate from 42.3% to 48.9%,a gain of 6.6 percentage points, without any additional training costand with only modest inference overhead. Our results demonstrate thatfailure-driven self-improvement is a viable complement to success-basedpipelines, enabling more efficient agent improvement. Code is availableat https://github.com/snow10072740/Learning_from_Failure.

Spike camera, a type of neuromorphic sensor with high-temporal resolution, shows great promise for high-speed visual tasks.Unlike traditional cameras, spike camera continuously accumulates pho-tons and fires asynchronous spike streams. Due to unique data modality,spike streams require reconstruction methods to become perceptible to thehuman eye. However, under low-light high-speed conditions, spike streamsbecome highly sparse and noisy, making faithful pixel-wise recovery in-trinsically difficult. In this work, we propose Diff-SPK, a diffusion-basedframework for perceptual reconstruction from low-light spike streams.Diff-SPK leverages generative priors to produce visually plausible re-constructions while remaining constrained by spike-derived structuralconditions. Specifically, it first employs an Enhanced Texture f rom Inter-spike Interval (ETFI) to aggregate sparse structural information fromlow-light spike streams. Then, the encoded ETFI by a suitable encoderserves as the input of ControlNet for high-speed scenes generation. Toimprove the quality of results, we introduce an ETFI-based feature fusionmodule during the generation process.

While Multi-modal Large Language Models (MLLMs) havedemonstrated superior performance on isolated tasks such as visual ques-tion answering, their reliability remains limited in real-world scenariosthat require highly compositional visual-language reasoning. A primeexample is automated exam grading, a real-world practical task thatdemands the integration of three core capabilities: detection and local-ization, text recognition, and reasoning for correctness judgment. Cur-rent benchmarks largely focus on isolated tasks and cannot fully eval-uate MLLMs’ end-to-end ability in such complex settings. To bridgethis gap, we present GradingBench, a comprehensive benchmark basedon automated exam grading in Chinese K–12 education, which system-atically evaluates MLLMs across the entire grading pipeline. Grading-Bench comprises full-page exam papers from real educational settings,containing 3,284 sub-questions annotated with reference answers. Weevaluate MLLMs across three levels: single-question, specified-question,and full-page grading. Our experiments show that localization failureis the main bottleneck and reveal limitations in integrated multi-taskexecution. Although multi-round interaction and supervised fine-tuningbring slight improvements, the limited gains reflect fundamental weak-nesses in compositional ability. GradingBench serves as a challenging,application-driven benchmark for the community, underscoring the needfor seamless perception-cognition fusion in MLLMs to unlock their po-tential in practical domains. All code and data are available at https://github.com/ERRORSEMI/GradingBench.

Visual Autoregressive modeling (VAR) has emerged as ahighly efficient alternative to diffusion-based frameworks, achieving com-parable synthesis quality. However, as this paradigm extends to Space-time Autoregressive modeling (STAR) for video generation, scaling res-olution and frame counts leads to a "token explosion" that creates amassive computational bottleneck in the final refinement stages. To ad-dress this, we propose FastSTAR, a training-free acceleration frame-work designed for high-quality video generation. Our core method, Spa-tiotemporal Token Pruning, identifies essential tokens by integratingtwo specialized terms: (1) Spatial similarity, which evaluates structuralconvergence across hierarchical scales to skip computations in regionswhere further refinement becomes redundant, and (2) Temporal similar-ity, which identifies active motion trajectories by assessing feature-levelvariations relative to the preceding clip. Combined with a Partial Up-date mechanism, FastSTAR ensures that only non-converged regions arerefined, maintaining fluid motion while bypassing redundant computa-tions. Experimental results on InfinityStar demonstrate that FastSTARachieves up to a 2.01× speedup with a PSNR of 28.29 and less than 1%performance degradation, proving a superior efficiency-quality trade-offfor STAR-based video synthesis.

Text-to-image diffusion models power everyday creative tasks,but they still reproduce the demographic biases in their training data. Oncommon prompts such as “a photo of a nurse,” “a photo of a CEO”, theyskew their outputs toward one gender, driven by the statistics of trainingdata rather than anything in the text. Existing debiasing methods showpromise in narrow settings but require retraining, batch-level control, orprompt-specific tuning, limiting their scalability. We propose EquiSteer,a training-free method that works per sample by steering cross-attention(CA) activations at inference time. For each target attribute, EquiSteerprecomputes steering vectors from contrastive prompts. Then at gener-ation time, a prompt-aware gate leaves attribute-specific prompts un-touched, while for neutral ones it clears existing attribute signals fromthe CA activations and injects a target attribute. Across SD-1.5, SD-2.1,SDXL, and SANA, EquiSteer reduces the average parity gap by up to87%, with minimal effect on image quality and text-image alignment.Code is available at https://github.com/Atmyre/EquiSteer.

Multimodal Large Language Models (MLLMs) have demon-strated remarkable success in various Remote Sensing (RS) tasks. How-ever, their ability to comprehend negation remains underexplored, limit-ing deployment in real-world applications where models must explicitlyidentify what is false or absent, e.g., emergency responders need to locatenon-flooded routes for evacuation. To comprehensively study this limi-tation, we introduce RS-Neg, the first benchmark to evaluate negationunderstanding across region-level to scene-level tasks. Specifically, we de-sign an automated data generation pipeline for RS imagery, using LLMsto synthesize diverse negation queries, and introduce a dynamic visualfocus module for verification. Our evaluation reveals that advanced RSMLLMs struggle with negation, exhibiting hallucinations and substantialperformance degradation. To close this gap, we propose NeFo, a noveltest-time adaptation method that explicitly incorporates the logical roleof negation into the model optimization. Remarkably, using about 5%unlabeled test samples, NeFo significantly improves the negation under-standing of models and shows strong generalization to unseen tasks.

Coordinating navigation and manipulation with robust per-formance is essential for embodied AI in complex indoor environments.However, as tasks extend over long horizons, existing methods oftenstruggle due to catastrophic forgetting, spatial inconsistency, and rigidexecution. To address these issues, we propose ESCAPE (EpisodicSpatial memory Coupled with an Adaptive Policy for Execution), op-erating through a tightly coupled perception-grounding-execution work-flow. For robust perception, ESCAPE features a Spatio-Temporal FusionMapping module to autoregressively construct a depth-estimation-free,persistent 3D spatial memory, and a Memory-Driven Target Groundingmodule for precise interaction mask generation. To achieve flexible ac-tion, our Adaptive Execution Policy dynamically orchestrates proactiveglobal navigation and reactive local manipulation to seize opportunis-tic targets. ESCAPE achieves state-of-the-art performance on the AL-FRED benchmark, reaching 65.09% and 60.79% success rates in test seenand unseen environments with step-by-step instructions. By reducing re-dundant exploration, our ESCAPE substantially improves path-length-weighted metrics and remains robust (61.24%/56.04%) even without de-tailed guidance for long-horizon tasks.

Precise control of 3D facial expressions from text is crucialfor virtual avatars, animation, and human–computer interaction, yet ex-isting text-to-3D methods jointly generate identity, expression, and tex-ture, making fine-grained expression control difficult. We instead for-mulate text-driven expression synthesis as a regression problem in thedisentangled parameter space of a 3D Morphable Model (3DMM). Thissetting, however, requires paired data linking detailed language to pre-cise expression parameters, which is missing from existing resources. Tofill this gap, we introduce Txt2Emote, a benchmark of diverse 3D facialexpressions with fine-grained textual annotations obtained from GPT-4o and a high-fidelity face tracker, providing both explicit descriptionsdetailing facial features and implicit descriptions referencing the situa-tional context behind the expression. Leveraging this dataset, we presentEmoteGPT, a text-to-3D expression framework based on a Multi-ModalLarge Language Model (MLLM) with a dedicated <Expr> token to se-mantically ground expression representations, which are then decodedinto 3DMM parameters. We further improves EmoteGPT by augmentingtraining with large-scale image-to-3DMM data, surpasses state-of-the-arttext-to-3D face synthesis methods on emotion recognition metrics andin perceived expressiveness. Integrated into avatar pipelines, our methodenables photorealistic and stylized 3D avatars, as well as expressive 3D-consistent 2D face synthesis from textual input.

MeanFlow training commonly enforces the meanflow identitywith Jacobian-vector products (JVP). This is computationally expensiveand hard to scale, since the JVP path often weakens fused-kernel andcompiler-level optimization benefits. We revisit JVP-free MeanFlow froman error-driven perspective. Rather than a drop-in JVP replacement, finitedifferencing exposes controllable error structure that directly informs al-gorithm design. We identify two dominant error channels: step-truncationerror and trajectory-velocity error.Guided by this decomposition, we introduce a JVP-free differential Mean-Flow framework for from-scratch MeanFlow-style one-step image gen-eration. For step-truncation error, we use stochastic signed one-sideddifferencing, which needs only one additional forward pass and recoverssecond-order truncation behavior in expectation. For trajectory-velocityerror, we propose DuoFlow, which predicts mean and instantaneous veloc-ities on the same sampled state to improve trajectory consistency in dif-ferential updates. We further introduce progressive self-bootstrapping todrive trajectory velocity toward a fixed-point-like self-consistent tendencyduring training. As a fast derivative-evaluation scheme, our one-forwarddifferential estimator is between 7.17× and 34.83× faster than JVP ondirectional-derivative micro-benchmarks for batch sizes from 1 to 64, andreduces peak memory by a factor between 1.71× and 2.68×. On ImageNet256×256 from-scratch training, it improves FID by 30% over a matchedJVP-based MeanFlow baseline. Code will be available at here.

3D Gaussian Splatting (3DGS) has emerged as a promis-ing paradigm for reconstructing streamable free-viewpoint video (FVV)from multi-view videos. However, 3DGS-based FVVs typically lack userinteraction and editing capabilities, which diminishes the immersive ex-perience. Recent research has integrated language features from CLIPinto 3DGS via distillation, enabling open-vocabulary queries and sup-porting many downstream applications. Nevertheless, the stringent re-quirements of FVV, low frame size and high FPS, make current languageGaussian representations unsuitable for language-embedded FVV. In thispaper, we propose DLGStream, a novel language-embedded FVV repre-sentation that streams time-varying language features alongside Gaus-sian attributes to support 4D environment interaction, scene editing, andspatial intelligence. Specifically, we propose a dual-opacity dynamic lan-guage Gaussian representation, which maintains two opacity attributesfor color and language features to deal with performance degradationthat occurs when colors and features are jointly optimized. Further-more, we introduce an interpolation-based deformation field to reducetemporal redundancy. This deformation field can also be used for 4Dframe interpolation, boosting FVV sequences from low to high FPS. Ex-perimental results demonstrate that DLGStream achieves superior per-formance in both on open-vocabulary segmentation and reconstructionquality with an average frame size of merely 43 KB. The code is availableon https://github.com/kkkzh/DLGStream.

We propose Discrete Diffusion Bridges (DDB), a novel frame-work designed to resolve the fundamental spatiotemporal misalignmentof standard discrete diffusion in image translation and generation. Bycorrupting data into a pure mask state via a random schedule, the con-ventional forward process induces a twofold misalignment: spatially, thispure-mask destination entirely discards the rich structural priors of thesource image; temporally, the random masking order inherently contra-dicts the “easy-first, hard-last” decoding mechanism used during infer-ence. To address this, DDB constructs a direct and efficient trajectorybetween domains. Spatially, we introduce a hybrid absorption mechanismthat redefines the absorbing state to a stochastic mixture of mask andsource tokens, effectively injecting source prior as spatial anchors into thelatent space. Temporally, we design an information-guided noise schedulethat quantifies semantic variation to prioritize the corruption of high-information regions at earlier timesteps. This ensures the model learnsto resolve difficult semantic changes using robust context from invariantregions. Extensive experiments validate the versatility and robustnessof our framework across diverse generative paradigms. DDB effectivelybalances edit alignment with structural fidelity across both text-guidedsemantic manipulation and pure structural image translation, while in-herently complementing text-to-image generation and guaranteeing ro-bust high-quality decoding under extremely low sampling steps. Codeand models are available at https://github.com/HKU-HealthAI/DDB.

Recent progress in 3D Gaussian Splatting (3DGS) has en-abled dense visual SLAM with pinhole cameras, yet most pipelines arenot designed for panoramic imagery. We present Cube-Splat, the firstpanoramic GS-SLAM framework that factorizes each 360◦ frame into acubemap of four fixed-orientation virtual pinhole views sharing a singleoptical center. By designating the front face as the primary pose state, weaccumulate gradients from all faces via an adjoint mapping, thereby en-abling multi-face observations to coherently update a single state whilestrictly preserving cross-view geometric consistency. Concurrently, ourmapping module densifies and optimizes anisotropic Gaussians using ag-gregated cubemap rays for high-fidelity, dense reconstruction. Further-more, to rigorously evaluate panoramic SLAM under diverse and chal-lenging conditions, we introduce SynPano, a highly scalable, photoreal-istic synthetic dataset featuring parameterized complex trajectories andmulti-modal ground truth. Extensive evaluations on two public bench-marks (PALVIO and OmniBlender) and our SynPano dataset, collec-tively encompassing both indoor and outdoor scenes, demonstrate thatCube-Splat achieves state-of-the-art (SOTA) performance in tracking ac-curacy and reconstruction fidelity. Both the source code and the SynPanodataset are available at https://github.com/guoxf304/CubeSplat.

Existing image forgery detection (IFD) methods either ex-ploit low-level, semantics-agnostic artifacts or rely on multimodal largelanguage models (MLLMs) with high-level semantic knowledge. Althoughnaturally complementary, these two information streams are highly hetero-geneous in both paradigm and reasoning, making it difficult for existingmethods to unify them or effectively model their cross-level interactions.To address this gap, we propose ForenAgent, a multi-round interactiveIFD framework that enables MLLMs to autonomously invoke, execute,and iteratively refine Python-based low-level tools around the detectionobjective, thereby achieving more flexible and interpretable forgery analy-sis. ForenAgent adopts a two-stage training pipeline with Cold Start andReinforcement Fine-Tuning to progressively improve tool interaction andreasoning adaptability. We design a human-inspired dynamic reasoningloop with global perception, local focusing, iterative probing, and holisticadjudication, and implement it as both a data-sampling strategy anda task-aligned process reward. To support training and evaluation, webuilt FABench, a heterogeneous agent-forensics dataset with 100k imagesand about 200k agent-interaction question-answer pairs. Experimentsshow that ForenAgent exhibits emergent tool-use competence and re-flective reasoning on challenging IFD tasks when assisted by low-leveltools, charting a promising route toward general-purpose IFD. The codeis available at https://github.com/zfr00/ForenAgent.

Multi-modal Large Language Models (MLLMs) have signifi-cantly advanced video reasoning, yet Video Question Answering (VideoQA)remains challenging due to its demand for temporal causal reasoningand evidence-grounded answer generation. Prevailing end-to-end MLLMframeworks lack explicit structured reasoning between visual percep-tion and answer derivation, causing severe hallucinations and poor in-terpretability. Existing methods also fail to address three core gaps:faithful visual clue extraction, utility-aware clue filtering, and end-to-end clue-answer alignment. Inspired by hierarchical human visual cogni-tion, we propose ClueNet, a clue-aware video reasoning framework with atwo-stage supervised fine-tuning paradigm without extensive base modelmodifications. Decoupled supervision aligns clue extraction and chain-based reasoning, while inference supervision with an adaptive clue filterrefines high-order reasoning, alongside lightweight modules for efficientinference. Experiments on NExT-QA, STAR, and MVBench show thatClueNet outperforms state-of-the-art methods by ≥ 1.1%, with strongbenchmark generalization, hallucination mitigation, and improved infer-ence efficiency. This work bridges the perception-to-generation gap inMLLM video understanding, providing an interpretable, faithful reason-ing paradigm for real-world VideoQA applications.

Multimodal Large Language Models (MLLMs) have recentlydemonstrated strong progress in visual–linguistic understanding, yet theirperformance on text-centric video reasoning remains highly sensitive toinput quality. Real-world user-provided videos frequently contain motionblur, compression artifacts, noise, and low-resolution text, substantiallyimpairing reliable text reading and downstream reasoning. WhetherMLLMs can robustly read and reason over in-the-wild text under di-verse quality conditions remains an unanswered fundamental question.We introduce ClearText-Video (CTVid), a large-scale, scene-text-awarebenchmark for studying text-centric video understanding under controlledquality variation. CTVid contains 4,639 real-world text-rich egocentricvideos, 550K+ frames, 1.6M human-verified scene-text annotations, and220K+ spatial/temporal question–answer pairs in Chinese and English.For each high-quality video, CTVid provides content-matched Degraded-Quality and Restored-Quality variants, supporting two task families:Text-Centric Video Restoration and Multi-Video Quality VideoQA. Weevaluate 18 representative restoration methods and 16 state-of-the-artMLLMs on CTVid. The results show that visual enhancement does notguarantee textual fidelity or downstream reasoning gains: blur is moredamaging than low resolution, restored videos can alter the textual evi-dence used by MLLMs, and OCR-only pipelines remain far below directmultimodal reasoning. CTVid exposes the gap between video restorationand text-grounded understanding, providing a rigorous foundation forrestoration-aware, quality-robust text-centric video systems. Benchmark:https://github.com/jinlong17/CTVid-Bench.

The rise of home-deployed embodied AI systems is driving agrowing need for fast, metric 3D reconstruction of residential spaces tosupport navigation, interaction, and long-horizon task execution. How-ever, the commonly used pinhole-camera 3D reconstruction pipelinesstruggle to model large indoor residences efficiently due to their limitedfield of view, to which achieving full coverage across multiple rooms oftenrequires thousands of images and incurs drift from long chains of incre-mental alignment. In this work, we present CasaMaestro (Spanish wordsmeaning “house” and “master”), a feedforward model that can take onlytwenty to fifty sparse multi-view indoor panoramas as input and directlypredicts metric depth along with camera poses, allowing fast point-cloudDA3-AUC30 Pi3-AUC30 VGGT-AUC30Yaw Step 90° Yaw Step 45° Yaw Step 15° DA3-time Pi3-time VGGT-timeMulti Viewpoints Single Viewpoint0.9 8000.8 7000.7 6000.65000.54000.43000.30.2 2000.1 1000 0 (s)AUC Yaw step 90 Yaw step 45 Yaw step 15 TimesFig. 2: Illustration of existing problems. Left visualization shows pinhole modelseither face limited FoV in sparse capture or accumulative error in dense sequence. Rightfigure shows pose accuracy and processing time under different input density.reconstruction of the entire house with full coverage. CasaMaestro isthe first model that supports house-scale reconstruction with multi-viewpanoramas. Experiments show that CasaMaestro can robustly providehigh quality results in both real-world and synthetic scenes, which canserve as a strong foundation for acquiring house-scale 3D indoor assetsto be applied in close-loop simulation.

Domain adaptation addresses the distributional shifts be-tween source and target domains. Given increasing data privacy con-straints, Source-Free Domain Adaptation (SFDA) has attracted growinginterest, as it precludes access to raw source data during adaptation.While existing SFDA methods predominantly rely on source-like datageneration or heuristic pseudo-labeling, their connection to target errorgeneralization is often unclear. In this work, we first present a theoreticalanalysis by establishing an upper bound on the target error that over-comes explicit dependence on the classical joint-error term. Guided bythis analysis, we propose an optimization principle based on AdversarialProxy Perturbation (APP), serving as a practical bridge between thederived bound and empirical training. We further extend the frameworkto multi-source scenarios and enable the integration of zero-shot priorsfrom vision-language models (e.g., CLIP) through mutual distillation.Extensive experiments on standard benchmarks demonstrate consistentimprovements over state-of-the-art methods in both unimodal and multi-modal settings, validating the proposed theoretical framework in practice.

Digital twins replicate the appearance and physical behav-ior of real-world objects for interactive simulation in XR, robotics, andgaming. While neural representations such as 3D Gaussians enable high-fidelity twins reconstructed from images, existing systems simulate onlyone instance at a time and barely reach real-time performance, makinglarge-scale rollout evaluation for robot learning, model predictive control,and multi-user XR impractical. We present Boba, the first batched simula-tor for physics-based Gaussian digital twins. Our key idea is to separate thestatic twin template from the dynamic simulation state and co-design thephysics, deformation, and rendering/visualization pipelines for batchedexecution. This structure enables shared template computation acrossinstances while minimizing memory traffic and synchronization overhead.Guided by this principle, we introduce compact surrogate spring–massmodels and a set of system-level optimizations—including mixed-precisionGaussian skinning, memory-efficient execution, and shared-memory-awarebatching—that significantly reduce compute and memory costs. On anXR-configured Jetson Orin, Boba’s local configuration achieves over 10×single-instance speedup over the prior state-of-the-art PhysTwin pipeline.Its distributed configuration, split between the Orin and a server, reaches25.1 ms end-to-end latency, 13.2× lower than the same PhysTwin base-line on Orin, while lowering incremental power by 22.2% relative toBoba’s local configuration. For batched execution, Boba scales to 3,310FPS on RTX 4090, a 16.2× system-level aggregate-throughput improve-ment over Boba’s optimized single-instance configuration on the sameGPU, and accelerates predictive control for deformable object manipula-tion, achieving speedups from 26× to an estimated > 2,000× relative toPhysTwin-based planning. These results show that Boba can supportpower-constrained XR applications and scalable robot planning.

Layout understanding, or the interpretation of element orga-nization, is essential for document analysis, user interface (UI) creation,and graphic design. While recent vision-language models (VLMs) ex-cel at interpreting atomic layouts composed of independent elements,they struggle with compositional layouts that require reasoning over vi-sually entangled elements within hierarchical multi-layer structures. Inthis paper, we introduce a new task, compositional layout understanding,and present CoDeLayout, a VQA dataset of ∼20K real-world multi-layerlayouts annotated with compositional element pairs and design intent.Through empirical analysis on CoDeLayout, we identify two key chal-lenges for existing VLMs: semantic drift between textual metadata andvisual content, and structural ambiguity in hierarchical inter-elementrelationships. To address these challenges, we propose MASON, a post-training paradigm that integrates multimodal alignment (MA) and struc-tural perception (SP). MA enhances element interpretation by ground-ing metadata-defined elements to their visual counterparts, mitigatingsemantic drift, while SP models layer-aware inter-element spatial rela-tionships to improve hierarchical understanding and reduce structuralambiguity. Experiments reveal substantial gaps in existing VLMs: eventhe strongest baseline, GPT-o3, achieves only 79.68% accuracy, whereasQwen2.5-VL 7B with MASON reaches 91.66%. Notably, MASON sur-passes full-data Direct Finetune using only 30% of the training data andscales better with additional data.