ReWorld: An Interactive World Model with Long-Horizon Memory
ReWorld keeps old scene evidence under a fixed memory budget while still streaming interactive video.
The paper splits short-horizon control from long-horizon recall during training, then uses bounded KV cache plus a pose-indexed landmark bank at inference. Its data setup aligns multiple video sources to the same physical action scale, so movement commands mean the same thing across rendered, game, and real footage. The authors report best control fidelity and generation quality against six recent interactive world models. In minute-long out-and-back rollouts, the fixed 12-chunk cache can regenerate the starting view after sliding-window memory has lost it and full-KV attention runs out of memory. ArXiv · AI/CL/LG's note
The paper splits short-horizon control from long-horizon recall during training, then uses bounded KV cache plus a pose-indexed landmark bank at inference. Its data setup aligns multiple video sources to the same physical action scale, so movement commands mean the same thing across rendered, game, and real footage. The authors report best control fidelity and generation quality against six recent interactive world models. In minute-long out-and-back rollouts, the fixed 12-chunk cache can regenerate the starting view after sliding-window memory has lost it and full-KV attention runs out of memory. ArXiv · AI/CL/LG's note
score 6