Over the past few years, the development of world models has predominantly followed the Scaling trajectory. Nevertheless, in real-world applications, humanoid robots frequently encounter action failures, which has led the industry to pivot towards causal large models. Recently, Aether AI introduced the first-ever causal world model, CausalWM, which incorporates a causal chain of thought for future prediction. This model has secured the top spot in the overall TriWorldBench evaluation and has showcased exceptional performance across various tests in R-Bench and PAI-Bench.

At present, embodied intelligence predominantly follows two main paths: VLA (Vision-Language-Action) and world models. Among these, world models are required to transition from a focus on image generation to a deeper understanding of the causal relationships between actions, environments, and outcomes. By extracting physical variables such as optical flow and depth, CausalWM constructs a stage-ordered reasoning chain. It then makes step-by-step predictions in accordance with the sequence of physical changes, while also implementing stage-ordered attention masks to prevent information leakage.

The training process of CausalWM involves approximately 30,000 hours of mixed data across three stages: pixel-level pre-training, Causal CoT (Chain of Thought) mid-stage training, and multi-objective reinforcement learning. CausalWM not only boosts performance in embodied intelligence tasks but also leverages its physical variables as control signals to facilitate the model's adaptation to new scenarios.

Aether AI perceives CausalWM as a crucial component of the causal world model pathway and envisions that future causal intelligence could extend its reach into the realms of biomedical and scientific discovery. The core objective is to empower models to comprehend the causes of change and accurately predict futures following interventions.