Linear programme matches AI weight edit with fewer changes
A write-up reports similar benchmark results after changing eight times fewer weights in Qwen3.5-0.8B, but the method has been tested on only one model.
A technical write-up says linear programming can reproduce a weight edit by GPT-6 Astra that reduced repetitive reasoning loops in Qwen3.5-0.8B. The work follows MIT CSAIL’s “Sparks of Weight-Space Programming”, which described Astra editing the model’s weights.
The author formulated the edit as a linear programme to select weight changes that push the model away from a measured “loop direction” while limiting effects on productive reasoning. The resulting edit reportedly matched Astra’s benchmark results while changing eight times fewer weights.
Using Astra’s direction, the linear programme scored 62.5 on a development set and 53.2 on MATH-500, compared with 59.4 and 50.0 for Astra’s edit. The author described the results as within noise; the MATH-500 comparison used one sample per problem. A direction estimated from the model’s own traces achieved about 93 per cent of the Astra-direction result.
The write-up also found that adding a constant steering vector at the right strength recovered nearly all of the reported improvement. The effective range was narrow: at half strength, loops returned, while at twice the strength, reasoning deteriorated.
The results are the author’s experiments, with no independent replication reported. The method was tested on one model, and the author says selecting layers without an agent and applying the approach to other behaviours remain open questions. The work addresses looping as an obstacle; it does not establish a general improvement in reasoning.

