JA EN
Glossary › dual-anchored

GLOSSARY

dual-anchored

appears in 2 paper titles

Definition

A design that pins something to two reference points instead of one — for instance constraining a model to stay close to both its original behavior and a human-provided reference, so that optimizing against either alone cannot drift unchecked. It is a descriptive coinage rather than a standard term, so it appears mostly inside method names. When you meet it, the thing to pin down is what the two anchors actually are and what happens when they disagree.

Explainers using this term