A team of researchers published ProvenanceGuard, a post-generation verification layer for Model Context Protocol (MCP) agents that checks whether each claim in an LLM answer is supported by the specif…
An independent evaluation of TypeSafe's jev-1.13.0 judge model on 495 claims from the LLM-AggreFact benchmark found it matched a human oracle on all 500 repeated decisions in LangChain's earlier test,…
Researchers introduced GraphLoom, a reliability-calibrated multimodal knowledge-graph retrieval-augmented generation framework that improves answer quality and evidence faithfulness by routing compact…
A new study auditing eight automatic attribution metrics across multiple datasets finds that no single metric consistently performs best, with rankings inverting across datasets (Kendall tau = -0.64).…