Aether AI Releases Causal World Model CausalWM, Tops TriWorldBench Leaderboard

Many existing world models take a direct path: input the current frame and action instructions, and directly generate future video. The report points out that this approach is very effective when the data scale is large enough, but motion, contact, and spatial change are all compressed into the model's interior, making it hard for outsiders to judge whether the model understands the physical process or has found some kind of shortcut in the training data. Once the scene becomes complex and the prediction horizon lengthens, a single mistaken contact judgment in the middle can make subsequent frames drift further and further off.

CausalWM's response is to explicitly predict part of the physical change before generating the future video. The team selected Optical Flow (which describes where objects move and by how much), Depth (the distance between objects and the camera), and Pointmap (which adds a three-dimensional spatial position to pixels) as intermediate variables. The report emphasizes that the key is not the use of these three types of variables—they have long been widely used in video prediction and visual understanding—but rather organizing them into an ordered reasoning chain.

Taking pushing a cup as an example, the model first judges what is moving in the frame and where it is moving (Physical Motion, expressed with Optical Flow), then infers how the three-dimensional spatial relationships change (Geometry, expressed with Depth and Pointmap), and only then generates the future frame. This forms the chain Observation → Physical Motion → Geometry → Future Observation. Each step is both a prediction target and the condition for the next step of reasoning.

The first mechanism given in the report is the stage-ordered attention mask. It assigns reading permissions to different reasoning stages: earlier steps can only see information that has already occurred and cannot read ahead into later Pointmaps or future frames. The report explains that otherwise the model might appear to complete the intermediate predictions while in fact having reverse-engineered the earlier variables from the final answer, and the causal chain would lose its meaning. Therefore, whether in training or inference, the model must advance in the same direction: Motion → Geometry → Future.

The second mechanism is the training pipeline itself. The report says Aether AI built a mixed dataset of about 30,000 hours of embodied video, covering robotic manipulation, first-person video, multi-view robotic data, and other physical interaction scenarios, with training split into three steps: Pixel-level Pre-training → Causal CoT Mid-training → Multi-objective RL Post-training.

The first step uses LTX-2.3-22B as the base and learns language-conditioned future prediction on large-scale video; to address the problem that a large amount of video lacks robot action annotations and that different robots differ greatly in joints and control methods, the team introduced CD-LAM to extract a unified latent action from video and trained a multi-view version. The second step generates intermediate variables step by step in the order Flow → Pointmap → Future RGB. The third step uses multi-objective reinforcement learning to score multiple sampled future outcomes along dimensions such as physical consistency, visual quality, and task completion, and in turn optimizes the intermediate Causal CoT path.

The report says that on the newly updated robotic world model benchmark TriWorldBench, CausalWM topped the Leaderboard, achieving Top-1 with a TWB-Score of 66.04. The benchmark examines world models' predictive ability on robotic tasks, especially the three-view consistency of the head camera, left wrist camera, and right wrist camera, and contains six evaluation dimensions and 19 specific metrics. The report says that it entered the top ranks or took the lead on multiple metrics including three-view consistency, task alignment, motion quality, perspective plausibility, and image quality.

But the report also makes clear: it did not sweep all individual first places. A leading total score is more like an indication that, under this evaluation suite, CausalWM can handle multi-view consistency, task understanding, physical interaction, and video generation quality well at the same time.

The report also mentions that the team tested under different settings such as PAI-Bench and ranked first in the latest ranking. PAI-Bench's cases all come from real-world collection such as dashcams and industrial cameras, spanning autonomous driving, robotic manipulation, industrial scenarios, human activities, physical common sense, and other domains; the evaluation introduces a Domain Score and lets a multimodal large model act as judge to interrogate physical details item by item. The judgment the report draws from this is: if the same method can achieve relatively good results under multiple task settings, it at least provides a signal at the methodological level—explicitly modeling the intermediate Motion and Geometry may indeed help predict changes in the physical world more stably.

The report proposes an additional capability: during training, CausalWM continuously receives visual variables such as optical flow, depth, and Pointmap and uses them to generate the next-stage result, and over time it learns a more general operation—as long as new conditions can be represented as visual information, they have a chance to be placed into the context and participate in future generation.

This leaves an entry point for control. The report gives an example: if researchers have already planned a robotic arm's joint motion trajectory in a simulator, using the robot's URDF structural information and forward kinematics, this string of joint data can be rendered into an action trajectory in the frame, then encoded into visual tokens and fed into CausalWM; after a small amount of fine-tuning, the model can recognize this condition that was not originally in the Causal CoT and generate a future video consistent with the specified trajectory. Manually drawn object paths, geometric constraints, and other visual signals can also be input through the same interface.

The report's positioning of this is: this approaches the core of the causal problem—ordinary prediction asks "what is most likely to appear next," while adding a trajectory is equivalent to actively changing a condition and then observing how the future changes, and the model begins to handle intervention-flavored questions such as "if the robotic arm moves this way, how will the cup move." But the report also clearly acknowledges that this is still far from counterfactual reasoning in the strict sense.

Three types of information need to be distinguished. The first is confirmed release facts: Aether AI released the first version of the causal world model CausalWM, with a parameter scale of 16B, adopting the Causal Chain-of-Thought paradigm, and publicly released the code, model weights, project homepage, and paper link. The second is the team's statements and evaluation results relayed by the report, including the 30,000 hours of data, the three-step training pipeline, TriWorldBench's 66.04 score and Top-1, and first place on PAI-Bench; these all come from a single source's relay and are treated in this article as source_claim, without independent third-party verification. The third is inference, such as the methodological-level signal that "explicitly modeling intermediate variables may help predict more stably," which the report itself also expresses only as "at least provides a signal."

The report also mentions the team's earlier RSIAgent: it lets an agent enter unfamiliar software, actively explore and repeatedly verify to figure out the relationship between operations and results, and then distill the experience into Memory; the report says that it did not update model parameters, yet pushed open-source models such as Kimi-K3 and GLM-5.3 to surpass closed-source models such as GPT-6 Astra on OSWorld 2.0 and Agents' Last Exam. These numbers and conclusions likewise come from the report's relay, lack first-party sources, and can only be viewed as source_claim.

What the report finally gives is the team's own positioning of its route: CausalWM is only a first-version attempt in the direction of causal world models, and there is still a long way to go before more complete causal discovery and counterfactual reasoning. The report cites Turing Award winner Judea Pearl's three-level division (association, intervention, counterfactual), saying that most existing mainstream models remain at the first level, "association." This framing statement belongs to the report's citation and commentary and does not constitute independent verification of CausalWM's capabilities.