Ψ-0.6
A unified vision-language world model that supports invoking representations such as text, depth maps, motion information, and camera poses during inference and adaptively selects reasoning paths through reinforcement learning.
Timeline (last 90 days)
The research team at NeuroAI Lab released the Ψ-0.6 model.
Ψ-0.6 extends reasoning from a single language modality to combinations of multiple representational forms, including text, depth maps, motion information, camera poses, and imagined generated scene images.
Ψ-0.6 acquires through reinforcement learning a scheduling policy that adaptively selects among different representations, and the model learns during training to judge which reasoning path suits the current task.
Ψ-0.6 surpasses leading open-weight models in a zero-shot manner on the MindCube complex spatial reasoning benchmark.
Ψ-0.6 outperforms other world models in predicting object motion on FlowBench.
Ψ-0.6 is trained on simulated room data yet can transfer to real robots' tabletop manipulation scenarios.
The same Ψ-0.6 model simultaneously serves in five roles: perceiver, vision-language model, world model, reasoner, and policy network.
Ψ-0.6's paper, code, and project homepage have all been made public.
Current public information about Ψ-0.6 consists mainly of evaluations self-reported by the authors, and third-party reproduction and broader benchmark comparisons remain to be added.
All Facts (9)
Current public information about Ψ-0.6 consists mainly of evaluations self-reported by the authors, and third-party reproduction and broader benchmark comparisons remain to be added.
50%UnverifiedΨ-0.6 extends reasoning from a single language modality to combinations of multiple representational forms, including text, depth maps, motion information, camera poses, and imagined generated scene images.
50%UnverifiedΨ-0.6 acquires through reinforcement learning a scheduling policy that adaptively selects among different representations, and the model learns during training to judge which reasoning path suits the current task.
50%UnverifiedΨ-0.6 surpasses leading open-weight models in a zero-shot manner on the MindCube complex spatial reasoning benchmark.
50%UnverifiedΨ-0.6 outperforms other world models in predicting object motion on FlowBench.
50%UnverifiedΨ-0.6 is trained on simulated room data yet can transfer to real robots' tabletop manipulation scenarios.
50%UnverifiedThe same Ψ-0.6 model simultaneously serves in five roles: perceiver, vision-language model, world model, reasoner, and policy network.
50%UnverifiedΨ-0.6's paper, code, and project homepage have all been made public.
50%UnverifiedThe research team at NeuroAI Lab released the Ψ-0.6 model.
50%