Interactive Benchmarks

Baoqing Yue, Zihan Zhu, Yutong Han +6

Existing reasoning evaluation paradigms suffer from different limitations: fixed benchmarks are increasingly saturated and vulnerable to contamination, while preference-based evalu…