There Is No Neutral Harness: Modern LLM Leaderboards Are Manufactured by Config-Fragile Items

没有中立的评估框架：现代大模型排行榜是由“配置脆弱”的题目制造出来的

Abstract: Multiple-choice benchmarks fix the questions and the correct answers, but not the harness: the order of the options, the wording of the prompt, and whether a language model’s answer is read from generated text or from per-option likelihoods. Work on this harness sensitivity reports it as aggregate score variance, leaving unexamined which items the variance falls on and whether they are the items that separate one model from the next.

摘要： 多项选择基准测试虽然固定了问题和正确答案，但并未固定评估框架（harness）：选项的顺序、提示词（prompt）的措辞，以及语言模型的答案是通过生成文本读取还是通过各选项的似然概率读取，这些因素均未统一。以往关于评估框架敏感性的研究通常将其表现为总分方差，却未深入探究这些方差具体落在哪些题目上，以及这些题目是否正是区分不同模型的关键。

We treat the evaluation harness of large language models (LLMs) as an independent variable and resolve its effect to single items. We introduce the fragility grid: 12 open-weight instruction-tuned LLMs from 4 families answer the same 3,679 items from 4 benchmarks (ARC, HellaSwag, MMLU, TruthfulQA) under 26 equally defensible harness configurations, recording one correctness bit for every model, item, and configuration. The comparison is matched, since the items, the weights, and the greedy decoding stay fixed while only the harness varies.

我们将大语言模型（LLM）的评估框架视为一个自变量，并将其影响细化到单个题目。我们引入了“脆弱性网格”（fragility grid）：来自 4 个系列的 12 个开源指令微调大模型，在 26 种同样合理的评估框架配置下，回答来自 4 个基准测试（ARC、HellaSwag、MMLU、TruthfulQA）的相同 3,679 道题目，并记录每个模型、题目和配置下的正确性比特。由于题目、模型权重和贪婪解码保持不变，仅评估框架发生变化，因此这种比较是匹配的。

Under the grid a model’s score is a band rather than a point: gemma4-31b scores between 31 and 89 percent depending only on the harness. Three results follow. On the items that two adjacent models both answer stably the pair is tied, and config-fragile items carry 95.7 percent of a pair’s gap on average. Four of the 12 models reach rank one under some configuration, so the harness selects the winner. Item discrimination, the property that benchmark-compression methods maximize, correlates with fragility at 0.28 (95 percent CI 0.25 to 0.30), so compression keeps the fragile items rather than removing them.

在网格测试下，模型的得分是一个区间而非一个点：例如 gemma4-31b 的得分在 31% 到 89% 之间波动，完全取决于评估框架。研究得出三个结论：在两个相邻模型都能稳定回答的题目上，两者表现持平；而“配置脆弱”的题目平均贡献了模型间 95.7% 的分差。在 12 个模型中，有 4 个模型在某些配置下能达到排名第一，这意味着评估框架决定了胜者。基准测试压缩方法所最大化的“题目区分度”与脆弱性呈 0.28 的相关性（95% 置信区间为 0.25 至 0.30），这意味着压缩过程保留了脆弱题目，而非将其剔除。

The scoring choice, not the option order that protocols usually fix, is the load-bearing axis. We release the per-item records and the analysis script, from which every number regenerates on a CPU in seconds, and we position the fragility grid as a check a leaderboard can run before it reports an order.

评分方式的选择（而非协议通常固定的选项顺序）才是影响结果的核心轴。我们发布了逐题记录和分析脚本，任何人在 CPU 上几秒钟内即可复现所有数据。我们将“脆弱性网格”定位为排行榜在发布排名之前可以运行的一项检查工具。