aifollow.news
Back Rohan Paul
Rohan Paul· @rohanpaul_ai · X· · Original publication time SelectedAI score60

Amazon paper finds some rule-following benchmarks mark correct answers wrongMachine translation

AI introduction

The researchers audited nine public benchmarks by checking source rules, rerunning reference programs, and testing scorers with controlled answers. LegalBench’s tax scorer removed decimal points; after the fix, accepted answers for each of three models rose from eight or nine to between 87 and 91 out of 96.

Original post · AI translation

@rohanpaul_ai
Translation · refer to the original

亚马逊一篇新论文发现,用于评测 AI 模型处理税务、法律和商业规则能力的基准测试,常常把正确答案判为错误,因为测试自身的答案键或评分代码存在错误。

RuleArena 和 LegalBench 等基准测试检验模型能否将成文规则应用于税务、法律和商业案例。如果参考答案或评分器有误,分数反映的就是基准测试中的错误,而不是模型的表现。

研究人员审查了 9 项公开基准测试:将参考答案追溯至规则原文,重新运行参考程序,并用预设答案测试评分器。随后,他们在每次修正后,对同一批已保存的模型回答重新评分。

LegalBench 的税务评分器删除了小数点,因此将 259,487.13 美元读成了 25,948,713。修正后,3 个模型各自在 96 个答案中获接受的数量,从 8 个或 9 个升至 87 到 91 个。

在相信规则遵循基准测试给出的低分之前,应阅读评分器代码,并手动核查几个被判错的答案。

Source: Rohan Paul · x.com

Research
Found an error? Send a correction