Repairing a model by cutting into its weights.
GraftModel is a small agent that reads a model's weights, finds what it writes when it goes wrong, takes that out of a few dozen columns, and tests the result. If the model isn't better, the cut goes back.
The tree is drawn from the model itself. Each branch tip is a real neuron from Qwen3.5-0.8B, layers 10 to 13. Click one to read it.
How it works
Qwen3.5-0.8B has a bad habit. It works a problem out, then checks its answer, then checks again, and keeps going until it runs out of room. GraftModel goes after habits like that in four steps, and each one is on this page.
It reads the neurons straight from the model file, finds the direction the habit is written in, and takes it out of the 64 neurons that write it most. Then it lets the model answer real problems, and keeps the cut only if the model does better.
The method is from a team at MIT, who did this to the same model and took its MATH-500 score from 7.9% to 48.1% by changing 131,072 of 752 million weights.
Reading a neuron
Whenever a neuron fires, it adds the same 1,024 numbers to the model's running state, just scaled up or down. The model's word table lives in that same space, so you can hold a neuron's numbers up against the words and see which ones it pushes towards. You don't have to run the model at all.
Click a branch tip.
Loading
writes towards
writes away from
- Loudness
- -
- Along the direction
- -
- Drives
- -
Checked against 2,000 common words, in your browser. Middle layers rarely spell things out, so a jumble is a normal answer.
Making the cut
Drag to take the direction out of the 64 neurons marked in red. It's the real arithmetic on the real numbers, and the rest of the tree barely moves.
- Weights changed
- 0
- Share of the model
- Mean first-branch strength
- 100%
This edits a copy in your browser. The model on the server isn't touched.
Running it for real
The only real test of a cut is to let the model answer problems, and that's slow on our machines. So GraftModel first scores each cut on saved text: does the model get readier to stop where it should, without stopping too soon? Only the best few are tried for real, on the same problems with the same random seeds as before. Anything that survives then gets one exam on problems it has never seen.