Reproduce “HalluGuard: Evidence-Grounded Small Reasoning Models to Mitigate Hallucinations in Retrieval-Augmented Generation” on OpenResearch

About this paper

Loris Bergeron, Ioana Buhnila, Jérôme François, Radu State

HalluGuard, a 4-billion parameter Small Reasoning Model (SRM), classifies document-claim pairs as grounded or hallucinated and provides transparent, evidence-grounded justifications. It achieves an average balanced accuracy of 75.7% on the LLM-AggreFact benchmark and 84.0% on RAGTruth, matching or outperforming larger specialized models while using significantly fewer parameters.

Key insights

• Small Reasoning Models (SRMs) can achieve competitive, and in some cases superior, performance for complex tasks like hallucination detection compared to much larger LLMs, demonstrating efficiency advantages.
• Explicit reasoning traces (via 'think mode') and advanced preference alignment techniques like ORPO are critical for robust hallucination mitigation and producing high-quality, transparent justifications.
• High-quality synthetic preference datasets, developed through multi-stage curation and LLM-based consensus filtering, effectively enable the distillation of complex reasoning capabilities into smaller models.

Summary by alphaXiv.