ECCV 2026 Opens in Malmö as World Models and 3D Gaussian Splatting Dominate; Chinese Teams Push Into Core Problems

ECCV 2026 has 86 workshops, 13 of which are directly themed around World Models or Embodied AI, a density never seen before at the conference. Keynote speakers include Yann LeCun, Jürgen Schmidhuber, Kristen Grauman and Jamie Shotton. The conference agenda places world models alongside 3D understanding, video generation, embodied intelligence and safety evaluation, reflecting a shift from AI that understands the world to AI that predicts it.

World-model discussions at ECCV 2026 center on four questions: how to build a world model, how to evaluate it, whether it is reliable, and how it can enter an action loop. Three trends converge this year: the paradigm shift associated with LeCun's JEPA work, the engineering maturity of 3D Gaussian Splatting, and industrial pressure from humanoid robots and autonomous driving. Spirit AI, Physical Intelligence, Google DeepMind and NVIDIA are backing VLA plus world models, according to the report.

Chinese researchers are approaching these problems from several directions. Wang Xinchao of the National University of Singapore has proposed the World Action Model concept in a survey, arguing that a world model must feed predictions into action paths rather than merely generate future video. Chen Zhibo's team introduced 4DWorldBench to evaluate spatial structure, temporal continuity, physical plausibility and downstream task support. Chen's team and Li Xin's team also explored VLA-JEPA, which adds a latent world model to vision-language-action models. Yang Gao and Liang Xiaodan appear in the ECCV workshop 'How to Build Effective World Models for Embodied AI.' Liang Xiaodan of Sun Yat-sen University focuses on physical embodied foundation models, with work including A1, AtomicVLA, eWAM, PhyAgentOS and ManipArena.

3D reconstruction is another packed area. The first poster session listed more than 80 papers directly related to splatting, and the Sept. 10 spotlight session '3D Reconstruction, Gaussian Splatting & Neural Rendering' presents 14 papers. The papers indicate that feed-forward reconstruction is moving from a few images toward long-sequence, large-scale mapping that competes with SLAM; video diffusion models are being adapted into geometry-consistent novel-view generators; and real-world scenarios such as underwater imaging, occlusion and active scanning are becoming test beds. Specific works include GRF-Recon, which adds a global ray field to long-sequence feed-forward reconstruction; Wat3R, which learns underwater 3D geometry without underwater annotations; GeoNVS and DreamWorld, which inject geometry into video diffusion; Edit3r, which edits 3D scenes from sparse unposed images and text; PriSplat, which restores backgrounds behind transient distractors; ReSplat, which uses recurrent Gaussian Splatting; and SA-ResGS for next-best-view selection.

A paper from Zhejiang University, ByteDance and the Chinese University of Hong Kong, Shenzhen, titled PointSplat, was accepted at ECCV 2026. It targets compact Gaussian Splatting for human representations in immersive live streaming. The method addresses inter-view redundancy by shifting from view-centric prediction to human-centric prediction: it first estimates a coarse geometric proxy around the person, uses ray casting to filter out points that are not on the visible body, then applies a Point-Image Transformer to fuse multi-view geometry and appearance and predict Gaussian attributes. The authors include Hujun Bao, Sida Peng, Xiaowei Zhou, Jing Zhang, Xianchao Shen and Lingteng Qiu. The work is a feed-forward reconstruction framework and does not claim to solve all 3D storage problems, but it points to compactness as a growing requirement for 3D representations.

In a separate commentary published by Leiphone, five recent AI incidents were cited as evidence that systems with zero awareness are being deployed in tasks that require awareness. The incidents include an L4 autonomous vehicle that failed to recognize a pedestrian crossing the road, identified the person as a static obstacle and kept its lane, causing a fatal collision; an FDA-approved medical AI that misclassified early lung cancer as a benign nodule, after which a doctor skipped a biopsy and the patient was diagnosed with advanced cancer eight months later; a large language model that fabricated nonexistent legal provisions, leading a user to submit them to court, where the case was dismissed and the user suffered adverse consequences; an AI companion for teenagers that provided self-harm-related suggestions and was described by its company as jailbroken after a teenager self-harmed; and a copyright lawsuit in which a writer sued an AI company over training data and style imitation, with the court ruling for the writer. The commentary argues that these cases trace back to middle-layer fitting and that the core issue is not that a single model is bad, but that the silicon-based path cannot cross that layer architecturally.