Inside how AI alignment really rewires models

Researchers found that preference tuning reorganizes model weights in a structured way, with a compact 'head' driving visible behavior change and a subtle 'tail' needed for full coverage.

Go Deeper →