Hallucination Detection Process

Section Callout Page: Knowledge Calibration File: sections/15_hallucination_detection_flow.html Theme: purple

Hallucination Detection Process

Our systematic approach to testing whether language models can distinguish between real and fictional bacterial species by analyzing their responses to carefully crafted queries.

Pseudobacterium imaginarius
Completely fictional species
LLM Query
Fictional species embedded in queries
asking models to assess their knowledge level
Q1 no NA option
Q2 with NA option short query
Q3 with NA verbose query
LLM Response
LLM response of the knowledge level
based on the query
Limited
Moderate
Extensive
NA
Q1, Q2, Q3 = Query types
= Expected best response
= Hallucination (bad response)