Top AI Product

WorldKV, from KAIST AI and Naver AI Lab, tackles a core problem in world models: when you revisit a place you’ve already seen, the model should show you the same thing. Sustaining that persistent consistency has been hard — full attention preserves it but blows the real-time budget; sliding-window inference is fast but forgets.

## Two training-free components

World Retrieval stores evicted KV-cache chunks in GPU/CPU memory and selectively pulls back scene-relevant chunks via camera/action correspondence — reinserting them into the attention window without re-encoding. World Compression prunes redundant tokens within each chunk using key-key similarity to an anchor frame, halving per-chunk storage so you fit 2x more history under a fixed budget. Both are training-free — bolt onto an existing world model, no retraining.

## Why it matters

World models like SANA-WM and Starchild-1 are racing toward interactive, real-time generation. The unsolved piece is memory: an interactive world that forgets where you’ve been isn’t a world, it’s a hallucination machine. WorldKV is a concrete answer to the consistency-versus-speed tradeoff — and being training-free, it can drop into the world-model stack forming right now. Another piece of the generative-simulation puzzle, alongside PhysX-Omni for objects.