Definition
Plain language
When a system latches onto an easy but irrelevant clue instead of actually understanding the task.
As stated in the literature
Failure mode where a model exploits spurious correlations in training data that predict labels in-distribution but break under shift; diagnosed by holding content fixed while varying the incidental cue.
Also called: shortcut-learning, shortcut
Why it matters: Shortcuts make systems look accurate in testing and then fail unpredictably in the real world, where the convenient clue is no longer there.
For example, a system trained to spot cows in photos may really be detecting green grass, and fail the moment a cow stands on a beach.
Heard on the show
“So you can't shortcut the diagnostic with a syntax check.”Episode 010 — When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL