AI research
A research paper titled "OA--RAG: Ontology--based Agentic RAG for Faithful LLM Reasoning" (arXiv:2610.36335), published on October 26, 2026, introduces an agentic retrieval--augmented generation framework engineered to enforce logical faithfulness and mitigate hallucinations in knowledge--critical domains such as biomedicine. Conventional RAG workflows prioritize semantic similarity over formal relation paths and rely almost entirely on the internal reasoning heuristics of large language models (LLMs), leading to unverifiable inferences. OA--RAG resolves this vulnerability by combining symbolic formal reasoning solvers directly with LLM generation, exploiting both formal ontological axioms and descriptive annotations. Each answer is coupled with verifiable ontology evidence and scored against a four--tiered grounding level—entailed, supported, related, or unrelated—to provide explicit reliability metrics. Across evaluations spanning three major ontologies and six biomedical question--answering benchmarks, OA--RAG demonstrated superior accuracy over baseline RAG and ontology--tuned pipelines, delivering the strongest performance improvements on complex, multi--hop reasoning tasks.