Can a Linear Program Do Weight-Space Programming?
The gist
Author Nathan Barry argues most of Astra's weight search was solving a linear program by trial and error.
A couple of LP solves match the edit with 8x fewer changed weights.
How it unfolded
- date unknownMIT CSAIL published "Sparks of Weight-Space Programming"; Astra could identify which game a small network was trained to play from its raw weights and write weights from scratch that beat networks trained with PPO for 100 million steps.
- date unknownAstra edited 0.016% of Qwen3.5-0.8B's weights, raising MATH-500 accuracy from 7.9% to 48.1% and IFEval from 21% to 62%.
- date unknownAstra spent about 8 hours and 28 forward passes choosing which weights to change; a couple of LP solves, about a minute each, do the same job with 8x fewer changed weights.
Key terms
- linear program
- A method to achieve the best outcome in a mathematical model whose requirements and objective are represented by linear relationships.
- PPO
- A reinforcement learning algorithm that is a policy gradient method, often used for deep RL with large policy networks.
- low-rank structure
- Approximating a matrix by one of lower rank, used for modeling and data compression with complexity constraints.
Source: Wikipedia