Can a Linear Program Do Weight-Space Programming?
Points and comments are a snapshot, not live.
A linear program achieves the same weight edit Astra used to fix reasoning loops, with 8x fewer changes.
Nathan Barry shows that Astra's 8-hour trial-and-error search to fix Qwen3.5-0.8B's thinking loops by editing two layers' down_proj weights can be replaced by a few minutes of LP solving (PDHG on GPU). The LP chooses which columns to edit and by how much, altering 16,271 weights vs Astra's 130,188, while matching or slightly beating Astra's benchmark scores (GSM8K 69.6, MATH-500 51.7, IFEval 59.7 vs Astra's 68.2, 50.9, 61.9). A heuristic direction derived entirely from model traces, with no LLM, gets about 93% of the effect. Because the edit is effectively a constant push along the loop direction, a steering vector at inference time recovers the same fix.
The results are limited to one model and one behavior (loops), rely on Astra for choosing which layers to edit, and the per-token constraints didn't matter since looping tokens aren't distinguishable from productive ones early enough.
What commenters are saying
Commenters broadly agree the article corrects hype around the original MIT CSAIL paper. The top comment notes this shows Astra was essentially doing 'normal AI engineering tasks, like running optimization algorithms over weights,' not demonstrating special insight. A split emerges: some argue the distinction is semantic since Astra did reliably arrive at the fix, while others counter that the LP's efficiency proves the original search was just inefficient. One commenter wishes for an LLM with constraint solvers 'baked in' at the embedding level rather than bolted on via tool calls.