Linear Programming Matches Weight-Space Programming With 8x Fewer Edits

1 reports archived1 independent sourcesupdated 11/11/2026, 20:30:17

Synthesis & Latest Updates

1 Sources Cross-Validated

Blogger Nathan Barry analyzes MIT CSAIL's "weight-space programming" and finds that GPT-6 Astra's weight edit on Qwen3.5-0.8B was effectively solving a linear program by trial and error, taking about 8 hours and 28 forward passes. Two roughly one-minute LP solves achieve the same result with 8x fewer changed weights, suggesting the core mechanism can be reproduced efficiently by classical optimization and offering a direct reference for weight editing and gradient-free fine-tuning.

LATEST/Blogger Nathan Barry analyzes MIT CSAIL's "weight-space programming" and finds that GPT-6 Astra's weight edit on Qwen3.5-0.8B was effectively solving a linear program by trial and error, taking about 8 hours and 28 forward passes. Two roughly one-minute LP solves achieve the same result with 8x fewer changed weights, suggesting the core mechanism can be reproduced efficiently by classical optimization and offering a direct reference for weight editing and gradient-free fine-tuning.

Hourly heat curve

11 fully observed hours

Now

14.0

Peak

17.622:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

Hacker News AIT2Nov 11, 20:30·78 pts
Linear Programming Matches Weight-Space Programming With 8x Fewer Edits
- Astra raised MATH-500 accuracy from 7.9% to 48.1% by changing only 0.016% of Qwen3.5-0.8B weights
- The edit is essentially an LP solved by trial and error; two LP solves reproduce it with 8x fewer changed weights
- The search stage of weight-space programming can be replaced by classical optimization, reducing reliance on long reasoning runs
View Related Reports & Analysis Original Source