VideoJev unifies video understanding and interaction with 100x faster QA inference
MCG-NJU · hf · 2026-10-14
VideoJev casts video understanding and interaction as structured decision tasks: pretrained MLLMs swap autoregressive decoding for a shared head that scores temporal intervals and textual options against video evidence, unifying temporal grounding and video QA. Cached and joint inference reuse visual computation across queries. VideoJev-4B and 35B-A3B improve mean mIoU over their backbones by 6.7 and 10.8 points on three grounding benchmarks with 2.8-3.1x measured speedups, and match Thinking-mode backbones on four video QA benchmarks with over 100x faster inference.
Related event: NJU MCG Releases VideoJev: 100x Faster Video Understanding(3 posts)→
More from Multimodal
- Runway integrates with Claude Motion to turn explainers into videos — tlakomy · 2026-10-15
- Scale CEO spotlights Muse Wonder: untrained for 3D yet beats Gemini 4 in physics visualization tests — alexandr_wang · 2026-10-15
- From a single image to a fully rigged, procedurally animated character — shekitup · 2026-10-15
- Seedance 2.5's Creative Upscale lets you reroll drafts at 480p, upscale only the keeper to 4K — anthara_ai · 2026-10-15
- Argon generates a llama typing at a computer using Fourier epicycles — osanseviero · 2026-10-15
- Prompting empty-room storytelling in Midjourney: two cups, one pulled-out chair — tisch_eins · 2026-10-15