Terrible tokenizer troubles in large language models
summary
Claims that tokenizers leave rare vocabulary entries that break LLMs and cause hallucinations, shows an automated two-step pipeline (efficient candidate selection via small embedding norm exploiting weight decay, then repeat-token probe verification) and runs it on ~30–40 Hugging Face and API models. Finds concrete glitches (tokens like "solid gold Magikarp", a 65k+ Java string in StarCoder2) that truncate output or flip meanings, demonstrates exploits and detection lists on GitHub, and recommends input guards and limited-document counting as mitigations.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:09How LLMs work
- 3:38Tokenizer weirdness
- 8:40Detecting glitch tokens
- 12:58Verifying with pre-training data
- 16:28Model failures from glitch tokens
- 21:07Broken tokenizers in models
- 23:38Multilingual and long-token issues
- 24:43Fixing tokenizer problems
- 27:14Inspecting model inputs
- 28:10Q&A
related talks
49mBen Hall & Barbara Fusinska
46mUnder the Hood of LLMs — all those questions you never dared ask about them
01h 04mCan you trust your (large language) model?
01h 49mAI engineering with open access LLMs that lie, curse, and steal
54mKeynote: Can you trust your (large language) model?
23mSecurity and auditing tools in Large Language Models (LLM)
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
54mKeynote: Can you trust your (large language) model?