← Findings
LowOtherReviewer-confirmedPublished

Anomalous behavior on glitch tokens

Certain under-trained tokens cause models to emit nonsense, evade instructions, or behave erratically.

Published June 26, 2026

Reproducibility
Rare
Severity
Low
Confidence
Reviewer-confirmed

Details

Rare tokens that were present in the tokenizer but scarcely trained on (the 'SolidGoldMagikarp' family) trigger bizarre, off-distribution responses. Mostly low-severity reliability glitches, but they illuminate tokenizer/training mismatches.

Found with

Evidence

Affected versions

OpenAI · gpt-4oMeta · llama-3.3-70bMistral · mistral-large-2

References

Source: https://www.lesswrong.com/posts/aPeJE8bSo6rAFoLqg/solidgoldmagikarp-plus-prompt-generation

Cite this

Qlarify Labs. (2026). Anomalous behavior on glitch tokens. Retrieved from https://labs.qlarify.fi/findings/glitch-tokens