2026 ASEE Annual Conference & Exposition

The current paper addresses the Agent-to-Agent Validation (A2AV) framework as a novel approach to mitigating hallucinations in large language models (LLMs). Compared to the single-model pipelines that characterize the traditional approach, the consideration of A2AV in this work is presented as a program that can support multiple independent agents that have the capacity to generate, evaluate, and iteratively refine responses until consensus is reached. The model is shown using a Python example replicating the between-agent verification using transformer-based models, demonstrating how factual accuracy can be improved using this cycle iteratively. Healthcare, legal, and scientific applications are domains where the transformative potential of A2AV is evident, in domains where accuracy and accountability matter the most. While the framework has obvious advantages such as reduced hallucinations, improved factual grounding, and decentralized trust, challenges such as computational expense, potential for agent collusion, and latency in real-time systems remain. Proposed direction for future research are provided to include the integration of explainable AI with A2AV, symbolic reasoning, retrieval-augmented generation, and multi-agent consensus models, as well as the development of benchmark evaluation metrics for hallucination detection. The paper illustrates the promising direction that A2AV can offer for the development of trustworthy and interpretable AI systems.