Humans reason about the physical world by flexibly mixing perception, language, geometry and simulation as a problem unfolds. Most AI systems still rely on specialized models, which limits their ability to be composed and jointly adapted to a single task.
$\Psi$-0.6 introduces a unified vision‑language world model that uses “streams of thought” to interleave images, text, and explicit geometry and motion representations through a shared token interface. The model can decide during inference which intermediate representation to employ and passes the generated tokens to downstream modules.
On multi‑view spatial reasoning benchmarks, $\Psi$-0.6 zero‑shot outperforms all leading open‑weight models, and gains further improvement when reasoning through geometric intermediates. It also supports counterfactual simulation, e.g., generating imagined scenes to assess whether moving an object would block a pathway.
Reinforcement learning from task feedback jointly optimizes the selection, generation, and use of intermediate representations, leading to notable gains in visual question answering that transfer to unseen benchmarks.
$\Psi$-0.6 offers a path toward adaptive physical reasoning—predicting the physical world while actively choosing the representations through which to reason. Review