Holi-World Opens 4B Ego-Video VLM and Human-Annotated Benchmark

A new Hugging Face paper from Hao Li, Ziwei Liu and co-authors introduces HOLI-WORLD, an ecosystem that turns egocentric recordings into training data, a 4B vision-language model, and an independent benchmark in one pass.

The pieces, in the authors' own words: "HOLIWORLD-Foundry combines complementary static and dynamic perception to automatically construct world records that connect persistent scene structure and evolving object states with interaction histories and supporting observations. HOLIWORLD-Data organizes these records from approximately 100 self-collected scenes into grounded learning resources, including approximately 80K QA pairs used to train HOLIWORLD-VL-4B." The evaluation side, HOLIWORLD-Bench, is built from "independent human annotations of heldout recordings."

The paper reports experiments on ScanNet, ScanNet++, Aria Digital Twin, HOLIWORLD-Bench and EgoDynamic4D, claiming "strong performance in static perception and dynamic tracking" and that training on HOLIWORLD-Data "substantially improves HOLIWORLD-VL-4B's spatial and temporal reasoning" on the latter two. The abstract publishes no per-benchmark accuracy numbers, no comparison baselines, and no license details.

It lands into a thick week of multimodal releases: our multimodal tracker logged 87 prior stories in the last 90 days, including Tsinghua's stateful video-rule harness earlier the same day.