A paper written on Probe that have high accuracy in detecting AI Hallucinations of models.
https://www.hallucination-probes.com/
- Uses specific dataset LongFact++ that is a collection of prompts designed to elicit detailed responses with specific Fact (entities, dates, locations, etc)
- Annotate output completions with LLM with web searchto determine if evidence is factually grounded
- Train proves to predict whether each token is part of a hallucinated entity. Linear probes read hidden states from each intermediate layer and output halucination probabilities for each token
