Hallucination Self-Play: Bootstrapping Reinforced Detector via Evolved Generator
Hallucination Self-Play (HSP) is a framework where a detector and a generator iteratively improve each other without human supervision. The generator creates plausible hallucinations, which the detector finds, and this feedback loop allows both models to co-evolve. This process progressively enhances the detector's ability to catch hallucinations, aiming for high performance against advanced AI models.