AetherAI's CausalWM tops TriWorldBench with causal chain-of-thought world modeling

机器之心 · wechat · 2026-09-20

AetherAI released CausalWM, a 16B causal world model introducing Causal Chain-of-Thought: instead of jumping straight to future frames, it explicitly reasons Observation → Motion (optical flow) → Geometry (depth/pointmap) → Future, with stage-ordered attention masks preventing answer leakage. Trained on 30k hours of embodied video in three stages (pixel pretraining on LTX-2.3-22B, Causal CoT mid-training, multi-objective RL), it tops TriWorldBench (TWB-Score 66.04) and PAI-Bench. The causal chain also doubles as a control interface for intervention-style queries. Paper, code, and weights are open.

Original post →

More from Embodied

Embodied channel →