Why do LLMs hallucinate?
summary
Explains hallucination as two concrete failure modes: missing facts inside a model’s implicit knowledge graph or wrong/noisy source text, and cites RLHF used during GPT-4 training as a mitigation. Recommends RAG/grounded generation implemented by ingesting docs, doing chunking and embeddings, storing vectors, retrieving by similarity and sending retrieved facts plus a prompt to the LLM. Ships an open-source hallucination evaluation model on Hugging Face, shows a news Q&A demo with citations, multilingual output, and argues fine‑tuning is usually costlier and less effective.
AI-written summary · about
chapters (14)
- 0:00Introduction and journey
- 1:44What are hallucinations
- 3:09Why LLMs hallucinate
- 6:18Measuring hallucinations
- 8:04Solutions to hallucinations
- 11:10What is RAG
- 13:17Challenges in RAG
- 17:08RAG demo and example
- 22:31Fine-tuning vs RAG
- 25:25NLP chunking explained
- 27:13Hallucination evaluation model
- 28:38Future of LLMs and RAG
- 30:26Why smaller models hallucinate
- 31:54Q&A session
related talks
39mLLMs and Hallucinations: Understand, Measure, Control
37mAI Hallucination Explained 🤯 Why AI Gets Things Wrong
41mTruth Checker: Generative Large Language Models and Hallucinations
26mReducing Hallucinations and Evaluating LLMs for Production
41mEnhancing LLMs with Graph Technology
52mGive Your LLMs a Left Brain
28mWhat Are AI Hallucinations?
01h 06mWhen calibration goes awry: hallucination in language models