Holi-World Opens 4B Egocentric Vision-Language Model and 80K-QA Benchmark Built From Human-Ego Video

Hao Li, Ziwei Liu and co-authors open Holi-World, an ecosystem that automatically builds learning resources from egocentric video. The release bundles HOLIWORLD-Foundry (static+dynamic perception stack), HOLIWORLD-Data (~100 self-collected scenes, ~80K QA pairs), the HOLIWORLD-VL-4B model and a human-annotated benchmark, evaluated on ScanNet, ScanNet++ and Aria Digital Twin for static perception and dynamic tracking.