ReWorld is an interactive world model that follows user actions, remembers previously shown places, and streams in real time. It separates control and memory during training and bounds them at inference. Mixed per-head attention windows confine most heads to recent past while a small set of global heads attends over entire history; random head routing and random chunk dropping make sparse histories in-distribution. At inference, a bounded KV cache backed by a pose-indexed landmark bank retrieves landmarks nearest the current pose. A metric-scale-aligned data engine places eight sources (Unreal-rendered fly-throughs, game roaming, real-world footage) on one physical action scale, and palindrome trajectories supply revisit evidence. Distribution-matching distillation confined to a LoRA adapter compresses sampling to four steps.
ReWorld is an interactive world model with long-horizon memory. It separates control and memory during training and bounds them at inference. Mixed per-head attention windows confine most heads to recent past while a small set of global heads attends over entire history; random head routing and random chunk dropping make sparse histories in-distribution. At inference, a bounded KV cache backed by a pose-indexed landmark bank retrieves landmarks nearest the current pose. A metric-scale-aligned data engine places eight sources on one physical action scale, and palindrome trajectories supply revisit evidence. Distribution-matching distillation confined to a LoRA adapter compresses sampling to four steps.
The architecture uses mixed per-head attention windows: most heads attend to recent context, while a small set of global heads attend over the entire history. Random head routing prevents either capability from binding to particular heads, and random chunk dropping makes sparse histories in-distribution. At inference, a bounded KV cache is backed by a pose-indexed landmark bank, retrieving landmarks nearest the current pose. Training data is aligned to a metric-scale action space across eight sources, with palindrome trajectories providing revisit evidence. Distillation via distribution matching into a LoRA adapter reduces sampling to four steps.
This work addresses a core tension in interactive world models: control requires short horizons, while memory requires unbounded context. The approach of separating these during training and bounding them at inference could influence future designs for real-time interactive AI systems, such as embodied agents, game AI, and streaming world models. The use of a metric-scale-aligned data engine across diverse sources suggests a trend toward unified action spaces for training world models.
ReWorld could enable more efficient and capable interactive world models for applications such as gaming, simulation, robotics, and virtual environments. The bounded memory and four-step sampling reduce computational cost, potentially making real-time interactive AI more feasible for commercial deployment.
Observable next signals include: whether ReWorld is released as open source or integrated into a product; whether the landmark bank retrieval approach is adopted by other world model researchers; and whether the four-step distillation method is applied to other generative models. Further validation on real-world interactive tasks and scaling to longer horizons would be key indicators of progress.