What happened
A new research paper introduces a way to figure out exactly why a large language model is performing worse after changes. This means developers can pinpoint the specific problem and fix it, instead of just knowing the model is broken.
Why it matters
When developers change a large language model, it often degrades in unpredictable ways. Until now, they could only tell if it was broken, not how or why. This new method lets them see the specific type of damage, like a doctor diagnosing a particular illness, which means they can apply a targeted fix.