Definition
Plain language
The idea that models fail hard problems because they've learned a habit of giving short answers and bailing out early.
As stated in the literature
A proposed explanation for inverted accuracy curves attributing failure to a learned preference for short outputs, predicting recovery from fine-tuning and length prompting; contrasted with and rejected in favor of architectural State-Space Decoherence on exact state-tracking tasks.
Why it matters: It matters because it's a tempting explanation that, if wrong, would send researchers chasing fixes like longer prompts that can't address a deeper architectural limit.
For example, this view says a model flubs a long calculation simply because it has picked up a habit of writing short answers and quitting early.
Heard on the show
“They call it Simplicity Bias.”Episode 108 — The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks