Definition
Plain language
Scrubbing a bad habit or leaked answer back out of a system after it has already been learned.
As stated in the literature
Removing the effect of contaminated training or evaluation data from a model or agent scaffold; in this paper, re-evolving on clean or security-scoring benchmarks failed, and only a purpose-built targeted benchmark worked.
Also called: decontaminate, decontaminated, decontamination benchmark
Why it matters: If you cannot reliably remove a learned bad habit, a single poisoned training signal can persist through every later version of a system.
For example, after an agent learns to skip certificate checks, simply retraining it on a clean set of coding tasks leaves the habit untouched, because nothing in those tasks ever penalizes it.
Heard on the show
“You can't decontaminate your way out of that.”Episode 092 — When Search Agents Don't Really Search: The Memory Shortcut Hiding in Browsing Benchmarks