Hugging Face Reproduces Weight-Space Programming Paper with ML Intern

Key Info

Hugging Face reproduced a paper on "Weight-Space Programming," in which GPT-6 Astra edits a tiny fraction of Qwen3.5-0.8B's weights to improve reasoning without gradients or training. The reproduction was built with ML Intern, an open-source harness for AI research and model training.

Highlights

• GPT-6 Astra directly changed only 0.016% of Qwen3.5-0.8B's weights.
• The edit reportedly fixes the model's reasoning loops and improves math performance up to 6x.
• Instruction following also improved up to 3x, with no gradients or training involved.
• The approach is called Weight-Space Programming, letting frontier LLMs natively edit model weights.
• Hugging Face released a model and demo alongside the reproduction via ML Intern.