Aether AI has released the first version of its causal world model, CausalWM, a 16B-parameter system built around what the team calls Causal Chain-of-Thought. Instead of jumping straight from a current frame to a future frame, the model explicitly predicts intermediate physical variables such as optical flow, depth, and pointmaps, then feeds those results back into the context before generating future observations. The goal is not only to predict what the future looks like, but to model how that future unfolds step by step. According to the article, CausalWM ranked No. 1 on the latest TriWorldBench leaderboard with a TWB-Score of 66.04, a benchmark focused on robotic world-model prediction and consistency across head, left-wrist, and right-wrist camera views. The team also reported leading results on PAI-Bench and other embodied world-model benchmarks. Aether AI said the method uses a stage-ordered attention mask to keep the reasoning chain causal during both training and inference, and trains the model in three stages: pixel-level pre-training, Causal CoT mid-training, and multi-objective RL post-training. The release includes a paper, code, model weights, and a project page. The article frames CausalWM as Aether AI’s first step toward what it calls real-world causal intelligence, extending the team’s earlier RSIAgent work from software environments into the physical world.