KAIST's ME-World Jointly Denoises Multi-Agent Ego Streams

Most egocentric world models predict first-person observations for a single agent. KAIST CVLab's ME-World drops that assumption. The paper, from Dahyun Chung, Siyoon Jin and six co-authors at KAIST AI, formulates multi-agent egocentric world modeling as synchronized ego-stream generation, with three requirements the authors spell out: "cross-view action consistency, shared-environment consistency, and consistent propagation of interaction-induced state updates."

The architecture "jointly denoises multiple ego streams in a shared token sequence, conditions each stream on all agents' target-view poses, and grounds generation with shared environment memory," the abstract says. The authors also introduce three new metrics for shared-world consistency, covering environment, state-update and identity.

Training and evaluation span real two-person head-mounted camera cooking footage (the CoMind dataset) and synthetic data built from Inter-X and InterHuman motion applied to VRM avatars in Blender scenes. The project page reports 221-frame autoregressive sequences, and scaling to two or three agents without model modification. The abstract's summary line: "ME-World improves shared-world consistency, action control, identity preservation, and video quality over existing methods."

The abstract publishes no per-metric numbers against the listed baselines, which include GEN3C, EgoSim, MultiWorld and Solaris. ME-World lands amid a run of KAIST work in our tracker, including a separate KAIST benchmark result covered a day earlier.