Research
LLM Hallucination Mitigation
Mohamed Nejjar's research on making Large Language Models more reliable, tackling hallucination at the prompt level.
Why This Research Matters
Large Language Models (LLMs) are transforming enterprise workflows, from legal document analysis to medical triage to customer service automation. But they have a critical flaw: hallucination. LLMs confidently generate plausible-sounding outputs that are factually wrong, internally inconsistent, or entirely fabricated.
Most hallucination research focuses on the model side: better training data, improved architectures, constrained decoding strategies. Mohamed Nejjar's research takes a fundamentally different approach: a shift-left strategy that tackles hallucination risk at the prompt level, before generation even begins.
Echo: Mitigating Hallucination Potential in User Prompts Through AI-Guided Iterative Refinement
Technical University of Munich, School of Software Engineering and AI
The core insight:every LLM output has two actors, the model and the user. The user's prompt is a controllable input surface that significantly influences hallucination risk, yet this dimension remains vastly under-researched.
Key Research Contributions
Novel Hallucination Taxonomy
Distinguishes between Prompt Risk (token-level ambiguity, vagueness, presuppositions) and Meta Risk (structural issues like multi-hop complexity, scope overload, conflicting constraints).
Prompt Risk Density (PRD)
A novel quantitative metric for measuring hallucination potential before generation. Weighted scoring across risk categories, normalized by prompt complexity.
Multi-Agent Pipeline
Four specialized agents (Analyzer, Initiator, Conversation, Preparator) that collaborate to analyze, diagnose, and iteratively refine prompts with human-in-the-loop validation.
High-Stakes Domain Impact
Demonstrated applicability in law, healthcare, and finance, where LLM hallucinations carry real consequences. Better prompts bridge the gap between expensive closed-source and smaller open-source models.
LLMs for Science: Usage for Code Generation and Data Analysis
Journal of Software: Evolution and Process (Wiley) · 55+ citations on Scopus
Co-authored a peer-reviewed paper examining the use of Large Language Models in research and data analysis. The work has been cited over 180 times, contributing to the academic discourse on responsible AI adoption in scientific workflows. This research established the empirical foundation for Mohamed Nejjar's subsequent work on LLM reliability and hallucination mitigation.
Research Philosophy
Mohamed Nejjar's research sits at the intersection of AI reliability and practical enterprise deployment. The goal is not to publish papers in isolation, but to produce research that directly improves how organizations deploy Large Language Models in production: safely, reliably, and at scale.
This philosophy is informed by hands-on experience building AI systems at BCG Platinion, Allianz SE, and Fraunhofer, environments where LLM reliability is not an academic curiosity but an operational requirement.
Frequently Asked Questions
What is Prompt Risk Density (PRD)?
Prompt Risk Density is a quantitative metric introduced in Mohamed Nejjar's Echo research for measuring an LLM prompt's hallucination potential before generation. It scores a prompt across weighted risk categories, such as token-level ambiguity and structural complexity, normalized by prompt length, giving a single density score that flags risky prompts pre-emptively.
How does Echo differ from model-side hallucination mitigation?
Most hallucination research targets the model: better training data, architectures, or decoding strategies. Echo takes a shift-left approach instead: it analyzes and refines the user's prompt before generation even begins, using a multi-agent pipeline (Analyzer, Initiator, Conversation, Preparator) with human-in-the-loop validation to reduce hallucination risk at the input stage.
What is the difference between Prompt Risk and Meta Risk in Echo's taxonomy?
Prompt Risk covers token-level issues in a prompt: ambiguity, vagueness, and unstated presuppositions. Meta Risk covers structural issues: multi-hop complexity, scope overload, and conflicting constraints. Echo's taxonomy separates the two so each can be diagnosed and mitigated independently.
Is this hallucination mitigation research peer-reviewed?
Echo is Mohamed Nejjar's Bachelor's thesis at the Technical University of Munich (2025). It builds on his earlier peer-reviewed, co-authored paper 'LLMs for Science: Usage for Code Generation and Data Analysis,' published in the Journal of Software: Evolution and Process (Wiley) and cited 180+ times on Scopus.