论文38

GPT-6.2 Astra 棋类变体测试表现引质疑

原标题:When asked to design chess variants to test out-of-distribution generalization, GPT-6.2 Astra simply swapped the piece labels. Fable 5.2 ...

Epoch AI@EpochAIResearch发布时间 (北京时间)

当被要求设计国际象棋变体以测试分布外泛化能力时,GPT-6.2 Astra 只是调换了棋子标签。Fable 5.2 和 Opus 5.5 则使用了已有的国际象棋变体,却声称是自己发明的。 测试分布外泛化能力的更强方法,涉及对规则和获胜条件做出真正新颖、更实质性的改变。

中文译文尚未完成,先为你展示原文。

When asked to design chess variants to test out-of-distribution generalization, GPT-6.2 Astra simply swapped the piece labels. Fable 5.2 and Opus 5.5 used pre-existing chess variants that they claimed they invented.

A stronger approach for testing out-of-distribution generalization involves more substantial changes to rules and win conditions that are genuinely novel.

推理能力论文研究评测基准#论文/研究#评测/基准#推理