Tag
JuryProbe introduces an empirical diagnostic for assessing consensus risk in reference-free LLM judge panels for factuality checking, and uses a calibration-based routing policy to ground high-risk decisions with trusted references, reducing false accepts.
This paper presents HallDetect, a lightweight and reference-free framework for hallucination detection that decomposes generated content into atomic claims and verifies them with a compact entailment model. It outperforms comparably resourced baselines on multiple benchmarks and provides a claim-to-span audit trail.