Triangular Review: A Multi-Agent Framework for Mitigating Hallucinations in Large Language Models

Large language models remain vulnerable to hallucinations and sycophancy when faced with misleading premises. In this paper, we propose Triangular Review, a heterogeneous multi-agent framework that decomposes question answering into three specialized roles: risk analysis, answer generation, and independent fact-checking. Unlike existing self-reflection or homogeneous voting methods, our architecture introduces a separate reviewer model that independently generates a reference answer and adjudicates the generated answer against it. The reviewer also performs review-and-write correction, directly producing a corrected answer instead of relying on the generator to revise itself. We evaluate the framework on 690 questions spanning five datasets, including an adversarial stress test designed to induce premise-consistent errors. Results show a 98.55% system pass rate and a 96.96% strict answer accuracy, with a 90.0% interception rate on adversarial stress tests and only 2 false alarms. These results demonstrate that heterogeneous triangular collaboration can substantially improve the trustworthiness of LLM systems without sacrificing standard benchmark performance.