Dynamic Gated Pathways wire every ViT layer to every LLM layer, boosting MLLMs over LLaVA-1.5
HKUST · hf · 2026-11-09
Researchers from Northwestern Polytechnical University and HKUST tackle the under-explored 'where to connect' question in MLLMs: how hierarchical visual features should be wired into an LLM. Their Dynamic Gated Pathways connect all ViT layers to all LLM layers, letting each LLM layer dynamically gate its preferred visual granularity, yielding consistent multimodal gains over LLaVA-1.5 baselines.
More from Multimodal
- Hyper3D Rodin's Agentic Mode turns reference images into editable, animatable 3D models — manishkhosiya · 2026-11-09
- Dreamina launches AI short-drama studio with in-house script, image and video models — aziz4ai · 2026-11-09
- AI-generated Ghost Rider images: flame biker riding across the sun and leaping between twin towers — egeberkina · 2026-11-09
- Neon Noir Kinetic Overdrive: a motion-blur image prompt template with fill-in slots — LudovicCreator · 2026-11-09
- AI-generated "Two of Fucks" tarot card meme goes viral — MickeySteamboat · 2026-11-09
- AI-made Japanese horror anime "THE SECRET MASKS" tests AI storytelling — IntentionSubject2843 · 2026-11-09