MIT CSAIL demonstrates LLM weight-space programming
MIT CSAIL researchers introduced Weight-Space Programming, a method that lets an LLM reason about and directly edit neural-network weights without gradients or a training loop. The authors report that GPT-6 Astra improved Qwen3.5-0.8B by editing 0.016% of its weights, raising MATH-500 from 7.9% to 48.1% and tripling IFEval under the same sampling settings. They also report that Astra created small game-playing networks that outperformed networks trained with PPO for 100 million steps.
The work presents a distinct route to model optimization and interpretability in which an LLM modifies weights directly rather than relying on conventional gradient-based optimization.
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
Authors
Some source content is omitted. Continue reading on the original site.
#MIT CSAIL#Weight-Space Programming#GPT-6 Astra#Qwen3.5-0.8B#MATH-500#IFEval#PPO