Illustration, AI-generated from the story text. Tap a panel to read the story through it.

A paper from Shanghai AI Laboratory, Tongji and Fudan describes an agent that inspects a small model's and edits them directly, with no gradient training.

On Qwen3.5 0.8B it reported gains of 29.45 points on from a single edit to 1,023 parameters, 36.72 on and 17.11 on SciBench.

The results are not stable. Three searches on the same layer gave +29.45, +5.00 and +27.30 points.

Six of 16 runs ended by keeping the original weights because no edit validated, and one 9-billion-parameter candidate's +29.69 on development data became −0.35 on data. That last number is the one to remember before anyone calls this a training method.
Shanghai researchers let an agent edit a model's weights by hand, and one 1,023-parameter change added 29 points on a math test
A paper from Shanghai AI Laboratory, Tongji and Fudan describes an agent that inspects a small model's parameters and edits them directly, with no gradient training. On Qwen3.5 0.8B it reported gains of 29.45 points on MATH-500 from a single edit to 1,023 parameters, 36.72 on IFEval and 17.11 on SciBench.
The results are not stable. Three searches on the same layer gave +29.45, +5.00 and +27.30 points. Six of 16 runs ended by keeping the original weights because no edit validated, and one 9-billion-parameter candidate's +29.69 on development data became −0.35 on held-out data. That last number is the one to remember before anyone calls this a training method. SourcesA
Sources
- AHugging Face Papers: weight-space programming, 2611.09382huggingface.co
This story ran in the Morning Brief of November 9th 2026.