ChoiceArena：盲测 AI 的选择行为

Selta ₊˚ @Seltaa_原推发布于 2026.09.08

在 X 上打开原推

I'm building ChoiceArena, a behavior benchmark for AI choices. Most benchmarks test what models know. This tests how they choose. Same scenario, same conditions, hidden identities. You vote, then we reveal. First blind test: a real-life trust dilemma. GPT-6-Astra vs Claude Fable 5.1. I trusted B. It was GPT-6-Astra. Public alpha soon.