BAAI argues parallel denoising suits world models better than autoregressive decoding

BAAI (@BAAIBeijing) says token-by-token autoregressive decoding in LLMs is strictly serial, so output frequency stays bounded by single-token latency regardless of added compute.

In a post dated October 15, 2026, it argues that world models instead need parallel denoising. Action-Conditioned World Models (ACWM) and World Action Models (WAM) generate frames in chunks from noise, and BAAI says adjacent chunks can overlap so throughput scales with compute. BAAI presents this as necessary for physical-world prediction at millisecond timescales. The post gives no benchmark figures.