Why it matters

Glitch tokens are curiosity + attack vector. Understanding shapes robustness.

Advertisement

The architecture

Under-trained tokens.

Model behavior unpredictable.

Attacker triggers.

Glitch tokens flowRare tokenunder-trainedInclude in promptattack triggerModelunpredictable outputDiscovered via clustering of embeddings; attackers scan for them
Glitch tokens.
Advertisement

How it works end to end

Discovered by Rumbelow + Watkins 2023.

Cause: token in vocab but rare in training.

Behavior: repetition, hallucination, refusals.