GSAR: Typed Grounding for Multi-Agent LLM Hallucination Recovery

The increasing deployment of autonomous multi-agent LLM systems for critical tasks, such as incident investigation and diagnostic reporting, underscores an urgent need for enhanced trustworthiness, particularly in mitigating hallucinations. Current groundedness evaluators often fall short by treating all supporting evidence as interchangeable and providing only a single, undifferentiated signal, which offers limited actionable control over downstream agent behavior. This lack of nuance impedes the reliable operation of AI in high-stakes environments.

GSAR, a novel grounding-evaluation and replanning framework, addresses these limitations through several core innovations. It first partitions claims into a four-way typology—grounded, ungrounded, contradicted, and complementary—thereby giving explicit recognition to alternative perspectives. Crucially, GSAR assigns evidence-type-specific weights to reflect the epistemic strength of different data sources, moving beyond a simplistic binary assessment. This granular approach culminates in an asymmetric contradiction-penalized weighted groundedness score, providing a more robust and informative metric. This score is then coupled to a three-tier decision function (proceed, regenerate, replan), which drives a bounded-iteration outer loop under an explicit compute budget, enabling dynamic and controlled recovery from ungrounded claims.

Evaluated on the FEVER dataset with gold Wikipedia evidence and across four independently trained LLM judges (gpt-5.4, claude-sonnet-4-6, claude-opus-4-7, gemini-2.5-pro), GSAR demonstrated consistent improvements, with all ablations reproducing in the same direction across every judge. This rigorous validation, including a head-to-head comparison against Vectara HHEM-2.1-Open, establishes GSAR as a pioneering framework that couples evidence-typed scoring with tiered recovery under an explicit compute budget. The implications are significant for enhancing the reliability of multi-agent LLM systems, fostering greater trust in their outputs, and enabling their responsible deployment in sensitive applications where factual accuracy and robust error recovery are paramount.