GSExplorer keeps a persistent 3D Gaussian world state for interactive scene generation
The arXiv paper describes a generation-reconstruction loop in which a 3DGS scene guides video generation, then SLAM-based checks filter updates, with reported WorldScore gains.
In plain words
GSExplorer is a research system that builds explorable 3D scenes step by step. It keeps a persistent 3D Gaussian Splatting (3DGS) model of the world, uses it to guide video generation, and checks new content against the existing scene before accepting it. The authors report top scores on a benchmark called WorldScore.
What was announced
The framework works in a loop. At each step, the current 3DGS scene supplies multimodal rendering conditions to a video generator, while camera geometry enters through a separate Relative Ray Encoding branch. The generated video is then passed to a SLAM-based reconstruction backend, where appearance and geometric consistency checks produce a shared support mask used for selective scene updates and cross-chunk pose refinement. The team also built a data pipeline that reproduces realistic 3DGS degradation during scene extrapolation, producing 293K training pairs of degraded renderings paired with real target videos.
On the WorldScore benchmark, the authors report the best overall performance among compared methods for both single-step and three-step evaluation, including a Camera Control score of 99.07 in single-step extrapolation. On 481-frame leave-and-return rollouts, they report strong revisit consistency. These are the paper's own comparisons; the abstract does not name the competing methods or provide their scores.
What it does not show
- The evidence is the paper's abstract only; full method details, baselines, and evaluation setup are not covered.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.