Jev-Like Model Learns to Cook

OpenJev turns descriptions into game controls by scoring how well each proposed action follows from the situation. In this cooperative cooking experiment, a text description of the Overcooked kitchen and recent outcomes is paired with the statement "Your next action should be" each of six buttons (four directions, stay and interact). A natural language inference classifier rates how well each statement follows, and those ratings are normalized into probabilities. Game rewards train the choices.

The trained model is OpenJev, a separate implementation inspired by TypeSafe AI's Jev interface, which takes unstructured state in and returns probabilities over predefined outputs. Two chefs share the same model but choose independently from their own observations. No planner assigns jobs or executes cooking sequences, so preparation and delivery must emerge from individual button presses.

The author reports six soups in 512 game steps, using the highest-probability action at each step. That was one evaluation of the best saved model on one seed and layout; nearby versions scored zero to six soups, and a new layout yielded zero. The reported result supports learning this kitchen's task, with transfer still unresolved. Videos and charts were not inspected.