Rlafuente

Two independently acting copies of the same decision model serve six soups in 512 ticks. Every move is a native game control. I trained a decision model to play cooperative Overcooked on one RTX 5080. Two independently acting copies learned to serve six soups in 512 ticks using only native controls. The policy is OpenJev , which applies the idea behind Jev : describe a situation, supply possible answers, and receive probabilities for them. OpenJev uses natural language inference to score whether each proposed answer follows from the context. Expressing actions as text lets the same classifier choose among different action spaces, making it a promising interface for learning through interaction. I’ve released the code and trained model . Figure 1. Greedy evaluations across the initial attempt, exploration restart, and two-chef continuation. Time starts at zero for each stage. Checkpoint 220 serves six soups.