Your model is a liar. It's also a fact-checker. Use the right one. A 2023 Meta… · @ReignSwords · TwKit
@ReignSwords
Your model is a liar. It's also a fact-checker. Use the right one. A 2023 Meta paper ("Chain-of-Verification") proved something I've felt in my bones: making the model check its own draft cuts hallucinations hard — but only if it checks BLIND. Let the model see its own draft while verifying, and it just copy-pastes its own lie back with full confidence. Which is exactly what happens when I reread my own writing: same eyes, same blind spot. Here's the 4-step recipe, copyable as-is: 1. DRAFT — answer the question normally. No tricks. Let it hallucinate; you need the raw material. 2. PLAN — feed the draft back and ask: "List verification questions that would expose false claims in this draft. Use general questions, not yes/no. Ask 'When was X founded?' — not 'Is it true X was founded in 2004?'" 3. VERIFY BLIND — new chat, no draft in context. Answer each verification question from scratch, independently. The paper's bluntest finding: independent verification beats in-context checking every time. The draft poisons its own interrogation. 4. REVISE — only now bring the draft and the verified answers together: "Revise the draft using only the verified answers. Cut anything contradicted. Keep what's confirmed." The numbers: on closed-book QA, F1 went from 0.39 to 0.48, and it beat zero-shot, few-shot, and chain-of-thought on long-form factuality. The honest caveat: it reduces hallucination, doesn't eliminate it. I run the human version of this on every post I publish — ~40 a day, two languages, every stat listed and re-checked in a fresh search before it ships. Meta just gave my paranoia a citation. Stop telling the model to "be careful." Make it interrogate itself — with the draft locked out of the room. Four model calls instead of one. Worth it, or is "first try good enough" the whole point of a chatbot? 🔵 Fellow Premium members, let's connect! I post about AI & tech — follow for more 👥
曝光 30 · 评论 6 · 点赞 0 · 书签 0
查看原帖