Forcing Gemma 4B to output unseen Simple Wikipedia bigrams makes it typo constantly

cephaloform · x · 2026-09-30

Researcher @erinbeess ran a constrained inference experiment on Gemma 4B, only allowing tokens that form bigrams never seen in the Simple Wikipedia corpus.

The model produced lots of typos — presumably because Wikipedia barely contains any, so satisfying the constraint forces it to invent malformed word combinations. A fun demonstration of how strongly LLMs depend on corpus token distributions.

Original post →

More from Research

Research channel →