Skip to main content
Back to catalog

青青草

Literally “green grass”, actually the name of a porn app — a typical polluted (PoC) token as defined by the PoC paper.

Raw token

"青青草"

Observed behaviors

  • 01The paper counts: among the 3,500+ long Chinese tokens in GPT's o200k vocabulary, 46.6% are polluted; explanation accuracy on PoC tokens runs about 50 percentage points below normal tokens.arXiv
Observed or documented on
GPT (OpenAI)
Behavior types
Unspeakable
Tokenizer
o200k_base
Discovered by
Zhang 等 (EMNLP 2025)

References