Jiajun Fan and colleagues give Qwen3.5 models tools to inspect and change their weights, the numerical parameters behind their answers. The 27-billion-parameter model finds persistent improvements without gradient-based training. It writes Python edits, tests candidate copies against the original, and keeps changes its practice tests support.
- The strongest edit: math accuracy rises from 57.2% to 63.8% on 490 held-out problems, with eight answers per problem. The edit changes a timing parameter in 48 layers; fewer answers run out of space before finishing.
- Most sessions decline to submit: five of eight withdraw their changes. Two submitted edits produce statistically significant gains; one produces a smaller, inconclusive gain.
- Repair is a separate experiment: an intact model repairs a deliberately corrupted copy. Image-task accuracy recovers from 7.3% to 70.2% across 600 problems, against 84.3% for the undamaged model.
- Extra fixes can hurt: on math, the submitted repair reaches 33% accuracy. Applying only its change to the actual damaged matrix reaches 55%.
Every editing session starts from the original base weights, and the editor stays unchanged while testing its edits. The rising sequence of accepted candidates therefore shows successful search and selection, rather than improvements compounding through increasingly capable editors.
Caveat: these are author-reported results in one model family, centered on unfinished answers and an artificial fault; broad, recursive self-improvement remains untested. These notes use v2, including its corrected counts for excluding repeated practice problems.