Year 12 · Student investigation
Can restoring an internal state recover a language model’s answer preference?
Keep an individual record even when you investigate in pairs. Record the exact settings, source and data split.
Which final-block patch should recover the clean score?
What did you change? What stayed the same? What did you notice?
Compare six prompt pairs, layers and positions against no-op, zero and norm-matched random controls.
A reproducible patching report naming checkpoint, metric and scope.
This activity can show _ . It cannot tell us _ . A new case I would test is ____ .
Use fictional data. Download or print your work before leaving.