Concept map: Hallucination and unsupported output

Taught and used in

LLM architecture and adaptation, Generative AI

Insights 21

• Supply volatile facts at request time because parametric knowledge is frozen at training
• Slice LLM accuracy by entity popularity because long-tail facts fail first
• Use majority voting over sampled answers for verifiable reasoning tasks, and price it
• Triage hallucinations as contradictions or unsupported additions because the fixes differ
• Score long answers claim by claim instead of with one overall verdict
• Use disagreement across samples as a hallucination flag when no reference exists
• Replace "check your answer" prompts with an external verifier
• Give the model a scored way to abstain, or your eval rewards confident guessing

Projects 8

• Flagship: passenger-rights document Q&A with a retrieval baseline, one controlled improvement and an honest generation mode
• Hybrid lexical/latent retrieval and lightweight reranking ablation over the academy's own knowledge base
• Chunking experiment on long technical papers: evidence recall at a fixed context budget
• LLM-as-judge reliability study: agreement with human pairwise labels, position bias and length bias
• Intent classification for banking support: LLM zero/few-shot vs a TF-IDF baseline on cost, accuracy and failure modes
• Summary faithfulness evaluation for legislative bills: human claim labels vs automatic checks