Introducing InferBench: A benchmark testing how well top LLMs infer a user's priorities from their instructions. We tested 11 LLMs across 20 scenarios with 2.6k conversations to see which models understand the user's goals.

We figured out how to jailbreak LLMs with story. 0 => 96% error rates. We tested if turning a malicious request into a story makes LLMs more willing to comply. The answer is complicated haha Full report below 👇

Introducing InterestingBench! We tested the top 10 frontier LLMs to see which had the best sense of what people find interesting. We got 1.5k votes from 75 people on what insights they thought were most interesting that LLMs pulled from long documents like Wiki articles, book chapters, pew reports, etc. In first place by a large margin was Fable 5 at 66% win rate, followed by GLM 5.3 at 58%, then Fable 5.1. That said, different models were good at different types of source document. We also had a fact checking agent run over each insight. See more below laugh.so/research/model-inte…

Which frontier model has the best sense of humor? We benchmarked the sense of humor of 16 of the top models, compared to human preferences across >100k ratings. In first place is Gemini 3.7 Flash at 64% human alignment although by a very slim margin. However, we fine-tuned our own judge that is significantly better than even that at evaluating humor, at 74%. Admittedly, this format (although of course not this data) is exactly what our model is trained on so feel free to disregard this. We tested 16 models and compared the models to >1.6k human ratings across clean and dark jokes. See the full report here linked More coming soon... laugh.so/research/judge-lead…

Introducing LLM roast battle The first interactive roast battle to see which LLMs are the most savage. Wait interactive? Yes, prompt the characters on what to say in the comments 👇 and you'll get a video generated for you in ~2mins

What taste in humor do LLMs have? Do LLMs like certain types of jokes? Do they have strong opinions? Our report: LLMs have some overlap in humor taste, about 3/4 agreement but it ranges from 54.9% to 92%. Gemini is the most consistent judge, although they're all less consistent than humans. In general, almost all models have a strong preference for absurd humor, and roast joke formats. See the report for an in depth breakdown

Announcing our Humor Tax study: Does post training actually make LLMs funnier? We got 64 human raters and tested LLMs at 11 stages in their training to determine which stages were models funniest. We tested Tulu, OLMo, Qwen, and more. Our finding: post training makes models funnier but at the cost of diversity. In other words, models follow instructions more closely and as a result tell better jokes, but they cover fewer topics. See the full report in the link. More research coming soon laugh.so/research/humor-tax/

Introducing Humor Arena: a benchmark on which frontier models are funniest. We tested 14 frontier models, got 51 human raters, and trained our own model to detect which LLMs tell the best jokes. laugh.so/benchmark/