Anomalous Model Responses
Phenomenon where Large Language Models (LLMs) produce bizarre, nonsensical, or degraded outputs in response to specific, seemingly innocuous input strings. Often linked to byte-pair-encoding artifacts and LLM Glitch Tokens.
Key Characteristics
- Trigger Sensitivity: Specific token sequences cause disproportionate degradation in output quality.
- Output Degradation: Responses may become repetitive, incoherent, or semantically void.
- Tokenization Dependency: Closely tied to how byte-pair-encoding maps subwords to integer IDs.
Related Concepts
- LLM Glitch Tokens: Byte Pair Encoding and Anomalous Model Responses
- Tokenization Artifacts
- Model Collapse
References
- LLM Glitch Tokens: Byte Pair Encoding and Anomalous Model Responses (Computerphile, 2026-09-21)