What Else Changes When You Fix a Model?

A model can learn from a new corpus while losing old behavior, receive privacy-protecting noise while losing task accuracy, or refuse a harmful prompt while also refusing a harmless neighbor. Success at the intended target leaves a nearby question unanswered.

The studies suggest a bounded way to investigate: locate what changes, choose a comparator, and inspect the closest relevant boundary. That boundary is different for training updates, privacy calibration, and prompt repair. Neither a plausible mechanism nor a reassuring broad score settles the other studies’ questions.

Imagine adapting a language model to a new corpus while still needing it to handle earlier material. The new examples may say little about some words the model previously used.

Yet the model’s output machinery can change even for words absent as targets in those examples. If old behavior worsens, simply knowing that the model learned the new material does not identify which part of adaptation deserves attention.

Now consider a training procedure that adds noise to limit what someone can infer about whether a particular record was used. A larger noise setting may be reassuring for that purpose, but it can affect the model’s accuracy on its task.

Here the practical choice is not just whether to add noise: it is how to compare ways of setting its amount while keeping the privacy question precisely defined.

Or suppose a model has produced harmful assistance, and a targeted change makes it refuse that prompt. A harmless request with similar wording might also be refused.

A broad test of the model’s general capabilities could look nearly unchanged while that nearby mistake remains. The question across these situations is how to test a targeted change against the boundary most likely to reveal what its stated success leaves uncertain.