Does weight-space programming need to touch the hidden layers? Editing only three output-layer rows of Qwen3.5-0.8B — the rows scoring the `` stop token and two spellings of `Wait` — reached 55.2% (276/500) on MATH-500, compared with 52.2% (261/500) for the published Astra edit that modifies feed-forward layers 15 and 19, according to a 17 November 2026 blog experiment. The author cautions the runs used different engines, with an exploratory paired bootstrap interval of −0.4 to +6.4 points and a descriptive McNemar p of 0.110, so the result does not establish superiority or equivalence. The output-head edit averaged 15,340 tokens versus about 21,638 retokenized reference tokens on the 20 questions with published reference text, 29.1% fewer on that subset, while the unedited model scored 7.0% (35/500). An output-head experiment · 17 November 2026 Does weight-space programming need to touch the hidden layers? I tried leaving Qwen’s internal layers alone and editing only three rows that score its next token. The model reached 55.2% on MATH-500. Then I looked at the runs that still would not stop. In Sparks of Weight-Space Programming https://rikiyat.github.io/blog/weight-space-programming/ , Astra improved Qwen3.5-0.8B by editing two internal feed-forward layers. Nathan’s follow-up https://nathan.rs/posts/weight-space-lp/ showed how a linear program could find a related edit. My experiment asks a narrower question: how much of the result can we recover at the model’s final output layer? A reasoning model already has a token for ending its scratchpad: