Debating with More Persuasive LLMs Leads to More Truthful Answers
paper Your tags
Your notes
ICML 2024 Best Paper from UCL DARK (Akbir Khan first author; Anthropic collaboration) — the canonical scalable-oversight result that debate between more persuasive LLM debaters leads a weaker judge to more truthful answers, rather than less. Code at ucl-dark/llm_debate (~131 stars).