Linear programming reproduces MIT weight-editing fix with 99% less time and 8x fewer weights
IgorCarron · x · 2026-11-10
In MIT CSAIL's "Sparks of Weight-Space Programming," GPT-6 Astra fixed Qwen's thinking loops by editing its weights directly. nathanrs replaced Astra's trial-and-error tuning with linear programming, constructing a weight edit matching its results while taking 99% less time (minutes vs 8 hours) and changing 8x fewer weights. Blog post and repo are available.
More from Research
- Chris McIntosh named Lead AI Scientist at UHN's Peter Munk Cardiac Care Centre — VectorInst · 2026-11-10
- Vector team wins MICCAI 2026 pediatric brain tumor segmentation challenge — VectorInst · 2026-11-10
- Benjamin Haibe-Kains wins community award for open science in AI drug discovery — VectorInst · 2026-11-10
- Nicolas Papernot wins 2026 NSERC McDonald Fellowship for trustworthy AI foundations — VectorInst · 2026-11-10
- Kieran Campbell wins CSMB New Investigator Award for ML in single-cell and cancer genomics — VectorInst · 2026-11-10
- Medical foundation model study for prostate cancer wins MICCAI IJCARS Best Paper Award — VectorInst · 2026-11-10