Ψ-0.6: Adaptive Streams of Thought in a Unified Vision-Language World Model
Ψ-0.6 can choose geometric representations and imagined scenes as intermediate steps in physical reasoning.
The model generates and interprets images, text, geometry and motion through a shared token interface. The authors report that it surpasses leading open-weight models on multi-view spatial reasoning without task-specific training, with further gains from geometric intermediates. It can also simulate counterfactual scenes, such as checking whether moving an object would block a pathway. Reinforcement learning improves its choice and use of intermediate representations, raising visual question-answering accuracy with gains that transfer to unseen benchmarks.
HF Daily Papers' note
The model generates and interprets images, text, geometry and motion through a shared token interface. The authors report that it surpasses leading open-weight models on multi-view spatial reasoning without task-specific training, with further gains from geometric intermediates. It can also simulate counterfactual scenes, such as checking whether moving an object would block a pathway. Reinforcement learning improves its choice and use of intermediate representations, raising visual question-answering accuracy with gains that transfer to unseen benchmarks.
HF Daily Papers' note
score 6