autoresearch-mlx

autoresearch-mlx is an Apple Silicon–optimized implementation of the autoresearch framework that enables autonomous AI research loops to run natively on MLX without requiring PyTorch or CUDA dependencies. It maintains the core autoresearch structure, where an AI agent iteratively edits a training script, executes experiments under a fixed time budget, and evaluates results based on a defined metric such as validation bits per byte. The system is tailored for Apple hardware, leveraging unified memory and MLX capabilities to achieve efficient training on Mac devices. It includes a minimal and focused project structure consisting of data preparation utilities, a modifiable training file, and a program specification that governs the agent’s behavior. The framework logs experiment results and supports continuous iteration, enabling long-running optimization cycles that can reveal hardware-specific performance patterns.