What if LLMs were good banter? Well we tested which LLMs have the best sense of humor. More specifically, we tested 16 top LLMs align most with human sense of humor. We ran >100k judgements and got >1.6k human ratings to see where do we agree across clean and dark jokes. The main findings: - Gemini 3.7 Flash has the best taste at 64% although it's really close at the top - However, we've trained our own judge that reaches 74% - Humans agree up to ~80%, the other part is subjectivity Other notes: - All frontier models have worse taste on dark jokes except Grok (unsurprisingly haha) - Model intelligence correlates decently (+0.42) with judgement of humor. As many say, humor is a sign of intelligence! Check our website to see what humor research is coming next! Full report here: https://laugh.so/research/judge-leaderboard. |