Benchmark comparison of Video-ORA across seven task families Dataset-level matrix comparing Video-ORA with Qwen3.5, InternVideo3, VideoChat3, Qwen3-VL, and LLaVA-OneVision-2, including the latest ReVSI result. Primary source: OraRL paper dated 2026-08-21, Tables 1–8 and 20. Dagger values: original LLaVA-OneVision-2, VideoChat3, OneThinker, and ReVSI reports. Table 1: Benchmark comparison of Video-ORA. Best Second-best † Original-report value Task family Benchmark Video-ORA9B Video-ORA4B Qwen3.59B InternVideo38B VideoChat34B Qwen3-VL8B LLaVA-OV28B TEMPORAL GROUNDING Charades (mIoU) 61.8 56.8 50.6 53.2 56.1 48.3 52.6 ActivityNet (mIoU) 63.6 57.3 52.2 46.3 54.6 46.8 52.4 QVHighlights (mIoU) 72.5 67.5 65.8 59.2 67.0 59.4 65.7 Average 66.0 60.5 56.2 52.9 59.2 51.5 56.9 SPATIAL GROUNDING RefCOCO (R@0.5, 3 splits) 92.9 91.4 91.6 — — 89.1† — RefCOCO+ (R@0.5, 3 splits) 88.9 86.3 86.4 — — 84.0† — RefCOCOg (R@0.5, 2 splits) 90.9 89.9 88.6 — — 86.8† — Average 90.9 89.1 88.9 — — 87.9† — SEGMENTATION RefCOCO (cIoU) 79.4 76.7 75.2 — — 73.8 — RefCOCO+ (cIoU) 72.6 68.7 65.6 — — 65.1 — RefCOCOg (cIoU) 75.1 73.0 70.0 — — 70.1 — MeViS (J&F) 61.3 57.5 32.1 — — 22.9 — ReasonVOS (J&F) 63.7 63.8 21.5 — — 19.6 — Average 70.4 67.9 52.9 — — 50.3 — VISUAL TRACKING GOT-10k (AO) 78.2 67.5 46.0 — — 33.7 — SPATIAL–TEMP. GROUNDING STVG (tIoU / sIoU avg.) 35.0 32.4 27.9 — — 19.5 — VIDEO QA VideoMME 76.7 69.8 71.1 73.8 70.1† 71.4† 71.9 VideoMME-v2 32.0 26.1 29.9 27.6 28.7† 18.2† 19.9 MV-Bench 72.5 67.2 68.2 75.0 — 69.0† 66.2 MMVU (MC) 77.0 65.4 72.5 — 56.4† 65.6† — VideoHolmes 65.5 61.9 50.3 — — 40.9† — LongVideoBench 66.2 60.2 64.6 66.8 — 68.0† 66.9 MLVU-MC 77.7 73.2 76.7 77.3 — 78.1† 76.6 Average 66.8 60.5 61.9 — — 58.7† — SPATIAL INTELLIGENCE VSI-Bench 73.1 66.2 57.9 — — 57.9 70.9 ReVSI (Avg., native frames) 58.2 — 51.6† — — 49.1† — MMSI-Bench 37.9 31.9 31.7 — — 31.1 — MindCube-Tiny 57.4 48.6 41.2 — — 29.4 — Paper Avg. (3 datasets) 56.1 48.9 43.6 — — 39.5 — † Original-report value; frame, prompt, split, or decoding settings may differ. OraRL paper (Aug. 21, 2026), Tables 1–8 & 20. Paper Average excludes ReVSI. ↑ higher is better