RealtimeWAM: One-Step Asynchronous World Action Models
RealtimeWAM targets the two latency costs left in efficient world action models: repeated action denoising and waiting between video and action experts.
The paper proposes one-step action generation using Teacher-Anchored Consistency Distillation, which supervises the student against the frozen teacher’s multi-step rollout endpoint. It also introduces Cross-Expert Wavefront Pipelining, sharing the video KV cache block by block so the action expert can begin before the video expert fully finishes. Across LIBERO, LIBERO-Plus, and RoboTwin, the authors report under 1% performance drop with end-to-end speedups of about 25x on H100. Code and checkpoints are stated as available. HF Daily Papers' note
The paper proposes one-step action generation using Teacher-Anchored Consistency Distillation, which supervises the student against the frozen teacher’s multi-step rollout endpoint. It also introduces Cross-Expert Wavefront Pipelining, sharing the video KV cache block by block so the action expert can begin before the video expert fully finishes. Across LIBERO, LIBERO-Plus, and RoboTwin, the authors report under 1% performance drop with end-to-end speedups of about 25x on H100. Code and checkpoints are stated as available. HF Daily Papers' note
score 5