Weight-Space LP Cuts Edit Search to Minutes

The compelling takeaway is that some “weight-space programming” may be ordinary optimization hiding behind an agentic workflow.

• –L1-regularized constraints select sparse weight columns that suppress reasoning loops while preserving productive activations.
• –Using Astra’s direction, the LP reaches comparable GSM8K, MATH-500, and IFEval results with 16,291 changed weights versus Astra’s 130,188.
• –The edit behaves much like a residual-stream steering vector, suggesting inference-time interventions may sometimes replace checkpoint surgery.
• –The result remains narrow: it uses Astra’s chosen layers and direction, targets one small model, and fixes a relatively simple looping behavior.
• –The released code makes this a useful baseline for testing whether refusal, sycophancy, or other behaviors can be edited with the same approach. [Code](https://github.com/nathanrs/weight-space-lp)