Research3d ago
Weakly Supervised Hallucination Detection in Transformers
The study builds a 15,000‑sample dataset from SQuAD v2, labeling responses as grounded or hallucinated via weak supervision. It combines substring matching, sentence‑embedding similarity, and an LLM judge to distill hallucination signals into transformer activations. The result lets practitioners detect hallucinations from a model’s internal states alone, removing the need for external verification at inference.