When calibration goes awry: hallucination in language models
summary
Claim: pretraining’s likelihood objective naturally produces pretraining→calibration which, under sparse factual coverage, forces hallucination; tools: information‑theoretic lower bounds and a reduction from generation to a binary classifier; quantitative diagnostic: hallucination rate follows a Good‑Turing rare‑fact rate (fraction seen once) and can be large for obituary‑style data but small for paper titles; constructive takeaway: thresholding LM probabilities yields a threshold classifier from LM and post‑training tradeoffs reduce hallucination at the cost of calibration.
AI-written summary · about
chapters (11)
- 0:00Introduction and context
- 2:11Defining AI hallucinations
- 3:28Nature of AI mistakes
- 5:16Pre-training mechanisms
- 8:37Statistical basis for errors
- 14:18Calibration and generalization
- 25:56Mathematical bounds on errors
- 32:39Learnability and classification
- 44:57Addressing hallucinations
- 53:42Human-inspired architectural insights
- 56:09Meta-principles and conclusion
related talks
32mWhy do LLMs hallucinate?
41mTruth Checker: Generative Large Language Models and Hallucinations
37mAI Hallucination Explained 🤯 Why AI Gets Things Wrong
39mLLMs and Hallucinations: Understand, Measure, Control
10mHallucinations: Harvesting Uncertainty
28mWhat Are AI Hallucinations?
34mDisincentivizing Hallucination
26mReducing Hallucinations and Evaluating LLMs for Production