MIT CSAIL replaces weight search with LP edits for Qwen3.5-0.8B

The author explores whether solving a linear program can replace the trial-and-error weight search used in MIT CSAIL's recent weight-space programming technique for language models. By framing the weight edit as a linear program, the author shows that a GPU solver can match the accuracy of the original agent-driven edits while modifying significantly fewer weights. The results demonstrate that optimization techniques can efficiently compute weight updates to correct model behaviors such as thinking loops. Open article →