没有中立的评测框架:现代大模型排行榜是由配置脆弱项构建的
文章背景与核心概要
尽管多项选择题基准测试在问题和标准答案上保持固定,但它们在“评测框架”(evaluation harnesses)——例如选项排序、提示词措辞,以及是通过生成文本还是计算选项似然值来获取答案——方面却缺乏标准化。这种不一致性导致了评测结果的巨大波动。
本文通过引入“脆弱性网格”(fragility grid),深入探讨了单个题目层面的框架敏感性。作者测试了12个开源指令微调大模型,涵盖了4个标准基准测试(ARC, HellaSwag, MMLU, TruthfulQA)中的3,679个题目,并应用了26种合理的配置。研究揭示,模型的得分并非一个确定的点,而是一个波动的区间。
核心结论指出,配置脆弱项(config-fragile items)贡献了相邻模型间95.7%的性能差距。这意味着,评测框架的选择实际上决定了基准测试的胜出者,而非模型本身的能力差异。此外,现有的基准压缩方法倾向于保留这些脆弱项,进一步加剧了排行榜的不稳定性。
执行摘要
尽管多项选择题基准测试固定了问题和正确答案,但它们将评估“框架”(harnesses)——例如选项排序、提示词措辞,以及答案是从生成的文本中读取还是通过选项似然值计算——留在了未标准化的状态。
While multiple-choice benchmarks fix questions and correct answers, they leave evaluation "harnesses"—such as option ordering, prompt wording, and whether answers are read from generated text or per-option likelihoods—unstandardized.
本文通过引入“脆弱性网格”(fragility grid),在单个题目层面研究了框架的敏感性。通过在26种合理的配置下,对来自4个标准基准测试(ARC, HellaSwag, MMLU, TruthfulQA)的3,679个题目进行测试,作者对12个开源指令微调大模型进行了评估,揭示了模型的得分是一个波动的区间,而非一个确定的点。至关重要的是,配置脆弱项占了相邻模型间性能差距的95.7%,这意味着评估框架实际上决定了基准测试的胜出者。
This paper investigates harness sensitivity at the individual item level by introducing the fragility grid. Testing 12 open-weight instruction-tuned LLMs across 3,679 items from 4 standard benchmarks (ARC, HellaSwag, MMLU, TruthfulQA) under 26 defensible configurations, the author reveals that a model's score is a fluctuating band rather than a point. Crucially, configuration-fragile items account for 95.7% of performance gaps between adjacent models, meaning evaluation harnesses effectively dictate benchmark winners.
关键发现
- 得分是区间而非点: 在固定模型权重和贪婪解码的情况下,模型的性能仅基于框架配置的不同就会发生剧烈变化。例如,
gemma4-31b的得分可以在 31% 到 89% 之间波动。
- Scores as Bands, Not Points: With fixed model weights and greedy decoding, a model's performance varies dramatically based solely on the harness configuration. For example,
gemma4-31bscores anywhere between 31% and 89%.
- 决定胜者: 在测试的12个模型中,有4个模型在至少一种配置下获得了排名第一,这证明了评估框架主动选择了“获胜”模型。
- Deciding the Winner: Four out of the 12 tested models achieve rank one under at least one configuration, proving that the evaluation harness actively selects the "winning" model.
- 脆弱项的作用: 配置脆弱项承担了相邻模型间平均 95.7% 的性能差距,而回答稳定的题目只会导致平局。
- The Role of Fragile Items: Config-fragile items carry an average of 95.7% of the performance gap between adjacent models, while stably answered items simply result in a tie.
- 有缺陷的压缩: 题目区分度(基准压缩方法主要最大化的指标)与脆弱性呈正相关(相关系数为 0.28)。因此,压缩方法倾向于保留脆弱项而不是丢弃它们。
- Flawed Compression: Item discrimination—the primary metric maximized by benchmark-compression methods—correlates positively with fragility (at 0.28). Consequently, compression methods tend to retain fragile items rather than discard them.
- 承载变化的轴心: 评分方法的选择(而不是协议通常标准化的选项顺序)是导致变化的主要承载轴。
- The Load-Bearing Axis: The choice of scoring method (rather than option order, which protocols usually standardize) serves as the primary load-bearing axis of variation.
工件与可复现性
作者已随论文发布了完整的逐题记录和分析脚本。所有结果均可在标准 CPU 上于数秒内重新生成。
The author has released the complete per-item records and analysis scripts alongside the paper. All results can be regenerated on a standard CPU within seconds.
- 访问选项:
- Access Options:
- View PDF
- HTML Version (Experimental)
- TeX Source
- License: CC BY 4.0