$Ψ$-0.6: Adaptive Streams of Thought in a Unified Vision-Language World Model

[arXiv] Khai Loong Aw, Klemen Kotar, Baihan Zhang and others propose Ψ-0.6, a unified vision-language world model that can adaptively combine perception, language, geometry and simulation within a shared reasoning process, which the authors call "streams of thought." These streams interleave images, text, and explicit representations of geometry and motion, which the model generates and interprets through a shared token interface. The paper states that the model surpasses leading open-weight models zero-shot on multi-view spatial reasoning, and that reasoning via geometric intermediate representations can further improve it; it also supports counterfactual simulation, such as judging whether a moving object will block a path. Reinforcement learning based on task feedback jointly optimizes the selection, generation and use of intermediate representations, improving visual question answering accuracy, and the gains transfer to unseen benchmarks. The authors say it points toward a path for adaptive physical reasoning.