Definition
Plain language
When training a model on extra data actually makes it worse at the task you care about.
As stated in the literature
A degradation in target-task performance caused by training on data whose distribution mismatches the target; e.g. fine-tuning a capable agent on easy, noisy single-app demonstrations erodes the cross-application reasoning the benchmark demands.
Why it matters: It shows that more training data isn't always better, and the wrong data can actively undo skills you wanted to keep.
For example, fine-tuning a strong multi-app agent on a pile of easy, sloppy single-app demos can leave it worse at the complex tasks it used to handle.
Heard on the show
“But the cleanest version of their "we prevent negative transfer" story leans on information you wouldn't have in production.”Episode 169 — Why Better Bug Reports Can Make AI Coding Agents Worse