Bachelor's Thesis — TUM
AI Research / Trustworthy AI · 2025
Echo: Mitigating Hallucination Potential in User Prompts
My bachelor's thesis. A shift-left approach to LLM hallucination mitigation, tackling the problem at the prompt level rather than after generation.
Context
Starting point and environment
Current hallucination research overwhelmingly focuses on LLM-sided factors: training data quality, model architecture, decoding strategies. However, the user's prompt is a controllable input surface that significantly influences hallucination risk — yet this dimension remains vastly under-researched.
Challenge
Core constraint to solve
Design a system that analyzes prompts before LLM generation to identify and mitigate hallucination-inducing patterns. Develop a novel taxonomy for user-sided hallucination risks and a quantitative metric for prompt risk assessment.
Solution
Design and implementation approach
Built a multi-agent pipeline (Analyzer, Initiator, Conversation, Preparator) with a novel taxonomy distinguishing Prompt Risk (token-level ambiguity) from Meta Risk (structural issues). Introduced Prompt Risk Density (PRD) — a weighted metric for quantifying hallucination potential. Implemented iterative human-in-the-loop refinement with structured XML outputs and color-coded risk visualization.
Outcome
Measurable impact and delivery result
Demonstrated that shift-left prompt analysis reduces downstream hallucination risk. Contributed to the research discourse on trustworthy AI in high-stakes domains (law, healthcare, finance). Proved that better prompts can bridge the accessibility gap between expensive closed-source and smaller open-source models.
Key Takeaway
“Hallucinations are a prompt problem as much as a model problem. Catching risk before generation is cheaper than fixing bad outputs later, and a clear taxonomy made the risk measurable.”
