2026-10-04 · America/Los_Angeles · 社区动态 · #2

An LLM scoreboard for whether an AI falls for the same trick questions people do

https://preview.redd.it/t0osnuw50fth1.png?width=1401&format=png&auto=webp&s=a0924bdfdccf1c81f096af95c9200b33410074d2 People are bad at a few famous trick questions. A bat and a ball cost $1.10. The bat costs $1 more than the ball. Almost everyone says the ball is 10 cents. It is 5 cents. Cognit gives those kinds of questions to AI models and puts the results on a board: https://cognit.rajtilak.tech/ Each model takes the quiz twice. The first time it just answers. The second time it is told to answer the way a person would. 100% means it gave the careful answer. 0% means it gave the answer…

热度 44.6 / 100；排名与评分保留该期记录。