New Paper Introduces Psi-0.6, a Unified Vision-Language World Model

Researchers including Khai Loong Aw, Klemen Kotar and Daniel Yamins posted a preprint describing Psi-0.6, a unified vision-language world model built to reason about physical scenes the way people do, by switching between forms of inference as a problem unfolds. The paper argues that today’s systems keep perception, language, geometry and simulation in separate specialized models, which limits how those capabilities can be combined or adapted together for a single task.

Psi-0.6 is organized around what the authors call streams of thought: a shared reasoning process that interleaves images, text and explicit representations of geometry and motion. The model generates and interprets all of these through one shared token interface, rather than passing outputs between separate tools built for each modality. The authors report that Psi-0.6 beats leading open-weight models in zero-shot tests of multi-view spatial reasoning, the task of working out how a scene looks from an angle it has not been shown, and that reasoning through geometric intermediates produces further gains on top of that.

The paper also describes a counterfactual simulation capability: Psi-0.6 can generate and interpret imagined scenes to work out, for instance, whether moving an object would block a path. Training pairs this with reinforcement learning driven by task feedback, which the authors use to jointly shape how the model selects, generates and uses its intermediate representations. That training approach, they report, raises visual question answering accuracy with gains that carry over to benchmarks the model was not trained on.

The numbers

The preprint, submitted on October 23, 2026, is a 13,260 KB PDF with eight listed authors working across perception, geometry and simulation research. The authors describe Psi-0.6 as surpassing leading open-weight models zero-shot on multi-view spatial reasoning, though the abstract does not publish a specific accuracy figure or benchmark score, leaving the exact margin for readers who pull the full document.

The approach matters beyond this one model because it is a bet against the current default of stitching together separate vision, language and simulation tools with custom glue code. If a shared token interface can hold geometry, images and text together and still beat specialized systems at spatial reasoning, that changes a concrete choice facing anyone building a system that has to reason about physical space: whether to keep composing separate models, or adopt one model trained to move between representations on its own.