Show HN: Jev-Like Model Learns to Cook
Category: ai-ml
Tags: reinforcement-learning, decision-model, overcooked, nli, lora
Score: 6.0/10 (Innovation: 6, Technical: 7, Documentation: 7, Utility: 4)
This project trains a decision model using natural language inference to play cooperative Overcooked, where two independent chefs learn to serve soups using only native actions. It leverages a pre-trained OpenJev NLI classifier and applies a clipped policy gradient with LoRA fine-tuning, achieving a specific benchmark result on a single GPU. The approach is interesting for its use of text-based action scoring and environment-agnostic exploration objectives, though it is currently limited to a narrow game domain.
Target audience: AI researchers and reinforcement learning practitioners
Repository: https://rlafuente.com/posts/2026-9-26-training-a-small-decision-model-to-cook# · Python · MIT
View on Hacker News