Machine Learning Algorithms & Natural Language Processing

Jun 3, 2026 · Artificial Intelligence

Can Multimodal Models Ditch Frame Sampling? LLaVA‑OneVision‑2.0’s Codec‑Stream

LLaVA‑OneVision‑2.0 replaces uniform frame sampling with a codec‑stream visual unit, integrates a OneVision‑Encoder that tokenizes video as state‑plus‑incremental evidence, and demonstrates consistent gains on 18 video, 11 spatial‑reasoning and 4 tracking benchmarks while open‑sourcing its model, data and code.