Holi-World pipeline lifts egocentric spatial QA token F1 from 18.3% to 66.4%
jiqizhixin · x · 2026-11-09
Holi-World, an arXiv preprint from Ropedia, NTU, HKUST and Peking University, links egocentric capture to spatial reasoning. Stereo RGB + IMU captures feed a 3D/4D annotation pipeline (HoliWorld-Foundry) that produces geometry- and semantics-rich spatial QA data; HoliWorld-VL fine-tuned from Qwen3.5-4B raises token F1 on EgoDynamic4D from 18.29% to 66.39% (a fine-tuning result, not zero-shot). Human-annotated HoliWorld-Bench provides independent verification.
More from Research
- TACL caps first authors at one submission every three months amid LLM-written paper flood — EhudReiter · 2026-11-09
- Exact Relative-Position Attention Kernels Deliver Up to 21x Backward Speedup at 8K Context — INSAIT-Institute · 2026-11-09
- HKUST-Led Benchmark Shows LLM Agents Drift Beyond Authorization Even Without Attackers — Chuxue Cao · 2026-11-09
- S-Agent 2.0 gives coding agents spatial investigation tools, lifting ReVSI by 20.5 points — jiqizhixin · 2026-11-09
- Papers with Code weekly trending: Inkling 2 Small and Drex top the charts — NielsRogge · 2026-11-09
- Opus 5.5 agents dig through 12,000 pages of Gauss and Riemann notebooks, find lost physics — roydanroy · 2026-11-09