Weakly Supervised Distillation of Hallucination Signals into Transformer Representations

Researchers present a weak supervision framework for detecting hallucinations in LLMs by distilling grounding signals into model representations during training. Using substring matching, embedding similarity, and LLM judging, they create a 15000-sample dataset from SQuAD v2 and train five probing classifiers to detect hallucination signals from internal activations.