Massive language fashions (LLMs) usually generate hallucinations — unsupported content material that undermines reliability. Whereas most prior works body hallucination detection as a binary process, many real-world functions require figuring out hallucinated spans, which is a multi-step determination making course of. This naturally raises the query of whether or not specific reasoning might help the advanced process of detecting hallucination spans. To reply this query, we first consider pretrained fashions with and with out Chain-of-Thought (CoT) reasoning, and present that CoT reasoning has the potential to generate no less than one appropriate reply when sampled a number of occasions. Motivated by this, we suggest RL4HS, a reinforcement studying framework that incentivizes reasoning with a span-level reward perform. RL4HS builds on Group Relative Coverage Optimization and introduces Class-Conscious Coverage Optimization to mitigate reward imbalance challenge. Experiments on the RAGTruth benchmark (summarization, query answering, data-to-text) present that RL4HS surpasses pretrained reasoning fashions and supervised fine-tuning, demonstrating the need of reinforcement studying with span-level rewards for detecting hallucination spans.
- † Nationwide Taiwan College, Taiwan
