摘要

面对现有AI基准测试的不足，研究人员正致力于设计更严格的评估方法。乔纳森·罗伯茨推出的ZeroBench，利用AI目前难以应对的视觉谜题检验多模态模型，结果所有模型均未能得分。Scale AI的EnigmaEval则以其高难度问题著称，领先模型也几乎无法解答。新兴测试如“人类的终极考验”及升级版ARC-AGI，旨在规避以往问题，如数据污染或题目过于简单，以评估AI超越记忆的能力。但AI技术的迅猛发展令这些测试迅速过时，ARC-AGI在短时间内便被超越。有观点指出，基准测试或许难以真实反映智能水平，因此出现了基于人类偏好的评价体系，如Chatbot Arena。随着ARC-AGI 3的启动，开发难以规避的评估工具成为新的竞争焦点。

Amid concerns that existing AI benchmarks are flawed or saturated, researchers are developing more rigorous tests. Jonathan Roberts’ ZeroBench challenges multimodal models with visual puzzles solvable by humans but currently impossible for AI, scoring all current models zero. Similarly, Scale AI’s EnigmaEval features brutally complex questions where top models score near zero. New benchmarks like "Humanity’s Last Exam" and updated ARC-AGI aim to avoid past pitfalls—such as training data contamination or simplistic questions—while measuring capabilities beyond memorization. However, rapid AI progress quickly outpaces these tests (e.g., ARC-AGI was surpassed within months). Critics argue benchmarks may never capture true intelligence, prompting alternatives like Chatbot Arena’s human preference rankings. The race for evasion-proof evaluations continues as ARC-AGI 3 development begins.