Lies, damned lies and large language models — Jodie Burchell

Two main types of LLM hallucinations exist:

• Faithfulness hallucinations - deviating from source text/context
• Factuality hallucinations - generating incorrect factual information

GPT model evolution shows increasing capabilities:

• GPT-1 (120M parameters): Basic grammar
• GPT-2: More sophisticated text completion
• GPT-3+: Ability to encode knowledge and generate coherent content

Training data quality significantly impacts hallucination rates:

• Early models relied heavily on unfiltered CommonCrawl data
• Modern approaches use filtered sources (C4, Refined Web)
• Higher quality input data generally leads to better performance

Methods to reduce hallucinations include:

• Careful prompt engineering
• Fine-tuning on specific domains
• Retrieval Augmented Generation (RAG)
• Self-refinement and collaborative refinement
• Using multiple models to cross-validate outputs

Measuring hallucination rates:

• Multiple evaluation datasets exist (TruthfulQA, HALU eval, SQuAD)
• TruthfulQA specifically tests for common misconceptions
• Current models still show significant hallucination rates (~30-40%)
• Measurement methods need to be specific to use case and domain

Large context windows help reduce inconsistencies but don’t eliminate hallucinations

Trade-offs exist between model size, performance, and hallucination rates

Critical evaluation needed when assessing model performance claims