🗂️ AI & Agents · View mindmap

Glitch Tokens

Glitch tokens refer to specific, seemingly innocuous input strings that cause Large Language Models (LLMs) to produce bizarre, nonsensical, or degraded outputs. This phenomenon highlights anomalies in how models process certain token sequences, often linked to Byte Pair Encoding artifacts or training data irregularities.

Key Characteristics

• Trigger Sensitivity: Specific character combinations or rare token sequences act as triggers.
• Output Degradation: Responses become incoherent, repetitive, or semantically void.
• BPE Connection: Often associated with how Byte Pair Encoding handles rare or unseen byte sequences, leading to unexpected model behavior.

Related Resources

• LLM Glitch Tokens: Byte Pair Encoding and Anomalous Model Responses
• Tokenization
• Model Robustness

References

• LLM Glitch Tokens: Byte Pair Encoding and Anomalous Model Responses (Computerphile, 2026-09-21)