Grounding, not scolding: retrieval with citations, permission to answer that it does not know, tool calls for facts and arithmetic, and a verification pass. The model optimises for the likeliest next token, not for truth, so instructions alone will not.
Card 1 of 292. Answer: Grounding, not scolding: retrieval with citations, permission to answer that it does not know, tool calls for facts and arithmetic, and a verification pass. The model optimises for the likeliest next token, not for truth, so instructions alone will not.
About · Quizzes · History of AI · To suggest new questions or report an error send me a dm.
Hallucinations drop when the model is grounded and checked: retrieval with citations, tools for facts and arithmetic, permission to say "I don't know", and a verification pass. A hallucination is fluent output that the source or the world does not support. Next-token prediction rewards plausible text, not true text, so instructions alone barely help.
Kalai et al., "Why Language Models Hallucinate" (2025), argue that training and benchmarks reward guessing. Score 1 for right and 0 for wrong or "I don't know": a model 30% sure expects 0.3 from guessing, 0 from abstaining, so it guesses. With a 1-point penalty for errors, the expected score is p − (1 − p), so guessing pays only above 50% confidence; a penalty of t/(1 − t) sets the threshold at t.
Interviewers ask why each fix works: citations make claims checkable against text, and a calculator replaces memory with computation. Lower temperature cuts randomness, not confident errors. RAG still hallucinates when retrieval misses, so give a no-answer path. SelfCheckGPT (Manakul et al., 2023) flags facts that vary across sampled answers.
