Harvard's WorldMove migrates live AI world models 101x faster

Researchers from Harvard (Jin Li, Jiawei Chen) have published a paper titled "Stateful Worlds, Stateless Elasticity: Exact-State Serving for Interactive World Models" that tackles one of the hardest problems in running persistent AI world models: you cannot recompute the multi-gigabyte GPU attention cache in interactive time, nor approximate it without changing the world. Their system, WorldMove, moves a live session's state between GPUs under the guarantee that the destination is bit-identical to the source, or nothing is installed. The cache is relocated in 18.8ms for same-node transfers – 101x faster than traditional save/load – and sustains 92.1-94.8 Gb/s on a 100 Gb fabric with checksum verification. The migration converges at a block boundary, allowing the destination to continue the world bit-for-bit.

Beyond the mover itself, the paper addresses two structural constraints: bit-exactness only survives within a single GPU architecture, and receive-path checksums stall under fan-in. The team built an incast-aware admission controller that holds zero misses up to 1.4x offered load and sheds overload as clean rejects. A lossless GPU codec widens the admission gate to fabrics raw motion cannot use. The approach was validated with a consolidation loop that executed 48 out of 48 migrations bit-identical across two cloud providers. The authors note the serving loop and mover were exercised separately end-to-end, but their composition on one fabric remains unbuilt. This work redefines elasticity for interactive AI as a joint scheduling problem over transport and verification.