Large language models (LLMs) write smoothly but still sometimes make believable yet incorrect or unsupported claims so called hallucinations. To address this, we propose MACV (Multi-Agent Cross-Verification), a modular system that has several cooperating components; a Primary Response Generator (PRG) that drafts answers; a Fact Checking Agent (FCA) that checks claims against external sources like Wikipedia or PubMed; a Domain-Specific Validator (DSV) that inspects specialist content using structured datasets and ontologies; an Adversarial Tester (AT) that looks for contradictions and weak reasoning; and finally a Consensus Mechanism that combines agent outputs into a confidence-tagged final response. We explain the architecture, agent interfaces, consensus strategies, and how we evaluate the system. Tests on benchmarks—SQuAD, SciQ, and FinQA datasets show MACV meaningfully lowers hallucination rates compared with single-LLM and RAG baselines while adding modest compute overhead. We also examine failure modes, scalability trade-offs, and practical deployment issues, and conclude that heterogeneous, cross-checking agents are a promising path toward more reliable LLM behavior in realworld settings.
Paper
Full text
MACV: A Specialized Multi-Agent and Consensus Framework for Reliable LLM Outputs
Semantic Scholar · 2026
Abstract
Large language models (LLMs) write smoothly but still sometimes make believable yet incorrect or unsupported claims so called hallucinations. To address this, we propose MACV (Multi-Agent Cross-Verification), a modular system that has several cooperating components; a Primary Response Generator (PRG) that drafts answers; a Fact Checking Agent (FCA) that checks claims against external sources like Wikipedia or PubMed; a Domain-Specific Validator (DSV) that inspects specialist content using structured datasets and ontologies; an Adversarial Tester (AT) that looks for contradictions and weak reasoning; and finally a Consensus Mechanism that combines agent outputs into a confidence-tagged final response. We explain the architecture, agent interfaces, consensus strategies, and how we evaluate the system. Tests on benchmarks—SQuAD, SciQ, and FinQA datasets show MACV meaningfully lowers hallucination rates compared with single-LLM and RAG baselines while adding modest compute overhead. We also examine failure modes, scalability trade-offs, and practical deployment issues, and conclude that heterogeneous, cross-checking agents are a promising path toward more reliable LLM behavior in realworld settings.