University of Tehran Researchers Unveil New Framework to Combat AI ‘Hallucinations’

A comprehensive study led by the University of Tehran offers a roadmap for making Artificial Intelligence more trustworthy and reliable in sensitive fields.

In a significant breakthrough addressing one of the most critical flaws in modern Artificial Intelligence, researchers at the University of Tehran have published a comprehensive analysis detailing advanced methods to eliminate ‘hallucinations’ in Large Language Models (LLMs).

The study, spearheaded by Dr. Shahrokh Asadi, a faculty member at the University of Tehran’s Farabi College and head of the research team, presents a sophisticated framework that transforms Generative AI from an ‘approximate’ probability engine into a ‘traceable, evidence-based’ tool. The report is based on a rigorous review of over 300 scientific references.

The 4-Step Framework for Reliable AI

The team introduced a structured, four-stage classification model designed to unlock the complexities of modern AI architectures: Indexing, Retrieval, Fusion, and Generation. This framework allows experts to decode the most intricate AI systems and evaluate their effectiveness in real-world operational environments.

From Vector to Graph: A Leap in Reasoning

According to Dr. Asadi, a central highlight of the research is tracing the evolutionary path of Retrieval-Augmented Generation (RAG) technology. While older models relied on ‘Vector RAG’ which simply matches surface-level word similarities, the new wave utilizes Graph RAG.

‘Unlike older models, graph-based architectures provide multi-step reasoning capabilities,’ Dr. Asadi explained. ‘This is absolutely vital for answering complex, analytical questions.’

The Rise of ‘Agentic’ and Multi-Modal AI

The study further analyzes emerging patterns that are currently pushing the boundaries of AI. Dr. Asadi highlights the pivotal role of Agentic RAG and hybrid multi-modal systems. These cutting-edge approaches enable AI to simultaneously leverage text, images, and complex databases to extract truth, rather than just generating plausible-sounding text.

Toward ‘Explainable’ and Standardized AI

The research has major implications for benchmarking and stability in AI. By providing precise evaluation criteria, it allows researchers and engineers to assess and standardize the reliability of their systems.

‘This shifts the path from simple data retrieval to a reasoning system based on clear evidence and secure AI usage for large-scale decision-making,’ Dr. Asadi noted.

The study, conducted in collaboration with student researcher Amir Abbas Kamalipour, serves as a blueprint for building a new generation of ‘explainable’ and ‘auditable’ AI systems.