Some notes i made while reading the paper. The setup: you can't edit GPT-4o's weights or read its logits. So most hallucination fixes in the literature don't apply to the models people actually ship with. What you do control is the prompt. Before each question goes to the model, another LLM tags it with 17 yes/no flags. Vague "it"? Needs domain knowledge? Asks for "the best" something? Then a bandit picks one of 5 rewrites: paraphrase, simplify, disambiguate, expand, or define the jargon. Bandit = the slot machine problem. Try the options, watch which pays out, lean on the winners, keep testing the rest a little. The flags tell it which machine to try for this particular question. Numbers on GPT-4o: 68.1% → 76.6% avg accuracy across 16 QA sets 80.7% → 88.8% on TruthfulQA Beat the untouched question 87.5% of the time ~$0.00035 per question, per the paper The part I keep thinking about: "always paraphrase" barely beat sending the question as-is. "Always expand" did worse than doing nothing. Worth remembering next time you add a one-line "rewrite the user query" step to a RAG pipeline. Pull the 17 flags out and the win rate drops from 87.5% to 81.7%. The flags carry real signal about which rewrite will help. Same rewrite, opposite results depending on the question. Simplify is fine on "can you tell me..." phrasing. On "what's the largest X" it trims away the comparison and the answer breaks. The eval detail I liked: on the original benchmark questions, the bandit mostly picked "no rewrite." GPT-4o has probably memorized them. The gains only showed up once they reworded the questions. Worth checking if your own evals lean on public benchmarks. Their scoring is worth a look if you build evals: 0.6 × LLM judge + 0.3 × fuzzy match + 0.1 × BLEU-1. A judge alone only gives 0 or 1. No partial credit for near misses, so the bandit had a harder time learning. What I'm taking into my own prompts: jargon → define it "it / they / last year" → spell it out dates, numbers, limits → add more "best / largest" → leave it alone Caveats from the paper itself: it lowers the hallucination rate, and an LLM is still grading the answers, biases included. Paper: arxiv.org/abs/2602.20332

100 years apart.