Show HN: Meadow Mind – a 7B diffusion LLM plays Gym games with zero training

Category: library

Tags: diffusion-llm, reinforcement-learning, gymnasium, zero-training, control, on-device-ai

Score: 7.0/10 (Innovation: 7, Technical: 7, Documentation: 8, Utility: 6)

Meadow Mind is a zero-training decision engine that uses a 7B diffusion LLM to play Gymnasium games by interpreting a one-sentence policy and state description, making decisions in ~0.4 seconds. It replaces reinforcement learning with a novel architecture combining a perception layer, rule-driven policy, and fixed-latency decision pass, achieving impressive results on classic control tasks. Its innovative use of a diffusion LLM for real-time control and working memory sets it apart, despite niche applicability to tasks describable in a sentence.

Target audience: AI researchers, reinforcement learning engineers, game AI developers

Repository: https://github.com/Hey-Meadow/meadow-mind · Python · MIT

View on Hacker News