Before trusting a weight edit, I measured the bug it fixes

Automated excerpt

With thinking on, the unedited model finished naturally on only 2 of 30 GSM8K questions at an 8,192-token budget. With the published edit applied, 10 of 30 finished and 9 of 30 were correct (vs. 1 of 30). Same 30 questions, same seeds, same 8,192-token cap, with the published edit applied.

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.