内容来源：https://www.quantamagazine.org/he-wants-ai-benchmarking-to-be-a-science-not-a-dark-art-20261016/

内容总结：

AI基准测试陷“暗术”困局，学者呼吁建立科学评估体系

核心观点： 马克斯·普朗克智能系统研究所主任莫里茨·哈特在其2026年新书《机器学习基准测试的新兴科学》中指出，AI基准测试生态已“严重失灵”，大语言模型时代的评估体系正面临根本性危机。

背景回顾： 基准测试堪称AI领域的“标尺”——既是衡量工具，也主导着整个领域的发展方向。1987年，美国国防高级研究计划局首次语音转文字基准测试显示，当时最先进的AI竟不敌所谓“笨拙”的统计方法，直接引发长达25年的AI寒冬。2012年，ImageNet基准测试又扭转乾坤，八层神经网络在图像分类任务中碾压其他算法，开启“深度学习革命”。如今，针对GPT-6 Astra等大语言模型的新基准测试层出不穷，但学界日益质疑其有效性。

问题症结： 哈特指出，机器学习领域长期忽视“留出法”警告，将测试集公开放到互联网上，导致测试数据轻易进入模型开发循环。更关键的是，不同数据集上同一模型的准确率可能天差地别——某数据集上90%、另一数据集上73%，绝对数值毫无意义。然而，OpenAI、Anthropic等公司仍将基准分数作为推理能力的证明，哈特认为这属于“严重误用”。

意外可靠性： 有趣的是，尽管绝对数值不可靠，但模型间的相对排名在不同数据集上却高度稳定。哈特认为，正是这种“排名稳定性”让机器学习侥幸逃过科学复制危机——90%的模型在其它数据集上同样会胜过80%的模型。但他强调，这纯属“运气”，学界至今缺乏完整理论解释。

大模型时代的新危机： 排名稳定性有一个重要前提：所有模型须在同一训练集上训练。但大语言模型的训练数据不透明，模型开发者可策略性地筛选数据以“应试”——针对已发布的基准测试进行“后期训练”，即所谓“基准刷分”。这导致新模型在旧基准上表现惊艳，在新基准上却黯然失色。哈特称，这正是当前基准测试危机的根源：更好的模型未必胜出，更差的模型可能只是更会“备考”。

解决思路——AI界的“高原训练”： 哈特援引1968年墨西哥城奥运会典故：美国1500米夺冠热门意外败给肯尼亚高原选手，引发是否禁止高原训练的争议，最终共识是“让所有人都练”。同理，基准测试的问题不在优化本身，而在于优化程度不均。他建议，取消传统训练集，改为在测试前对所有模型进行统一的“基准相关数据”微调，以拉平起跑线。

深层悖论： 哈特坦言，即便此法奏效，我们也只是回到原点——获得进步指标，却仍无有效测量。基准测试的胜利某种程度上是科学的遗憾：我们放弃了理解机制这一科学核心。但基准测试也有其悖论式力量——CASP蛋白质折叠基准测试在缺乏蛋白质折叠科学理论的情况下推动了科学进步。哈特表示，我们不知道大模型如何工作，基准测试也不会告诉我们答案，但从长远看，它仍可能催生新的科学知识。

中文翻译：

他想让人工智能基准测试成为一门科学，而非一门玄学

引言

基准测试是人工智能的标尺，无论从哪个意义上说都是如此。作为衡量工具，它们的运作逻辑极其简单：如果你想知道一个人工智能模型能做什么，就设计一套标准化测试来衡量它。但人工智能基准测试也可能主导整个领域的行为。

1987年，美国国防高级研究计划局首个自动语音转文字翻译基准测试表明，当时最先进的人工智能竟然比不过所谓的“笨拙”统计方法。这一糟糕表现助推了人工智能寒冬的到来。此后25年间，人工智能研究的热情逐渐消退，资金也随之枯竭。

然后，2012年，另一个名为ImageNet的基准测试让钟摆又荡了回来。一个八层神经网络在图像分类任务上碾压了其他算法，掀起了“深度学习革命”。

如今，针对OpenAI的GPT-6 Astra等大语言模型的新基准测试，以钟表般的规律不断被创建和攻克，声称能衡量从特定领域专业知识到“通用智能”的一切。但领域内许多人——无论是怀疑者还是支持者——并不将此趋势视为人工智能潜力飞升的证明，反而认为这是基准测试本身已存在致命缺陷的证据。

计算机科学家、马克斯·普朗克智能系统研究所所长莫里茨·哈特对此判断表示认同。ImageNet结果出炉后不久，他就开始思考机器学习系统可能“不做任何智能的事”就能在排行榜上钻空子。“这事没有两面性，”他说，“基准测试生态系统已经坏了。”

在他2026年出版的《机器学习基准测试的新兴科学》一书中，哈特提出了一个挑衅性的论点：这些标尺已经坏了几十年了。他写道，基准测试是“造就机器学习的那个错误”。值得注意的是，他认为，基准测试非但没有引发可复现性危机，反而推动了 legitimate 的进步——“很大程度上纯属运气”，在他看来。理解其中的原因，我们就能解释为什么在LLM时代基准测试已经失去了这种偶然的可信度，以及它们如何才能重新获得。

Quanta与哈特聊了聊基准测试哪里出了问题、哪里做对了，以及下一步该往哪里走。访谈经过压缩和编辑以确保清晰。

先把基础概念理清楚：基准测试到底是什么？

配方很简单。要评估一个机器学习模型，你把数据分成训练集和测试集。你让模型构建者在训练集上为所欲为：什么都行，没有规则。但测试集只用来比较和排名模型——让排行榜说出“这是当前最先进的，这是第一名，这是第二名”。

所以今天针对大语言模型的基准测试，比如GPT-6 Astra或Claude Fable，可能测试的是解决困难数学问题或在软件中发现安全漏洞的能力。

对。事实上，第一个基准测试创建于1959年。那是一组印在打孔卡上的数字和字母，用来测试不同扫描仪识别书写字符的能力。有人引入了训练-测试分割，并开始了一个排行榜。

并不是有什么总体规划。但这就是我们现在所说的基准测试，大约60年来它一直是人工智能进步的引擎。

基准测试应该告诉我们什么？

传统上，基准测试应该像温度计。人们训练模型做某件事，然后你设计某种测量程序来看看模型有多好。

但在实践中，基准测试更像体育比赛或市场。基准测试定义游戏规则，人们积极围绕它们竞争，我们让最好的模型获胜。

它们不能两者兼是吗？问题出在哪？

任何机器学习教科书都会告诉你“留出法”——你应该锁住测试集，只用一次来给模型排名。如果你忽视这一点，你就如履薄冰。你可能会得到非常误导性的结果，因为你评估的模型在某种意义上对测试集有所了解。

有趣的是，机器学习作为一门学科直接无视了这个警告。它把所有这些测试集放到互联网上供所有人使用。把这些测试集放进模型的开发循环中并不断迭代是很容易的。

这就是基准测试“坏掉”的地方吗？

如果你看一个模型在某个基准测试上的准确率数字，然后你只稍微改一下测试集，你就会得到完全不同的准确率数字。这就是基准测试失灵的意义所在。如果你在一个数据集上有90%的准确率，在另一个上有73%，那到底是哪个？我完全不知道这个模型有多好。

OpenAI、Anthropic和其他科技公司仍然宣传基准测试数字作为其模型推理能力的衡量标准。这纯属胡说吗？

这些分数确实被滥用了。如果你说“在这个数学基准测试上，模型得了90%，这意味着这个模型现在等同于一位专家级数学家”，你就是在从那个数字中读出一些给你错误结论的东西。我们应该避免那些隐含要求绝对数字有意义的比较或结论。

如果绝对数字没有意义，你怎么能不陷入可复现性危机？

这就是事情变得真正有趣的地方。如果你不看绝对数字，而是看排名——模型之间的相对比较——这些排名在不同数据集之间往往非常稳定。这在机器学习历史上一直相当一致。如果一个模型在某个测试集上得了90%，另一个得了80%，那么90%的模型就是你在现实世界中想用的模型。它在其他数据集上仍然会表现得更好，即使是那些与测试集相当不同的数据集。

对我来说，这就是机器学习逃脱科学危机的主要原因。

这很令人惊讶。所以90%这个分数没有多大有效性，但它“赢得”了竞争这个事实有？

是的。所以基准测试告诉你什么时候你击败了之前的最好成绩。而这些改进指标相当可靠。

但它能这么有效是运气好——我们没有理由期望会是这样。我们还没有一个好的综合理论来解释为什么会这样。这主要是一个经验现象。

但你也认为，现在在大语言模型时代，基准测试不再能提供这种“偶然的”可靠性。为什么不能？

模型排名的稳定性有一个重要前提：模型需要在相同的训练集上训练。在过去，这总是成立的。但我们不知道LLM是在什么上训练的。事实上，模型构建者可以战略性地组装数据组合，以最好地提升基准测试表现。这有点像应试学习，我们在高中都这么干。

这就是为什么我们目前处于这场基准测试危机中——因为更好的模型必然获胜这一点不再成立。一个更差的模型可能只是花更多精力为特定基准测试做准备。

这其实不算犯罪：每次新模型发布时都会发生。如果你看最新版Gemini或GPT的说明，你会发现它在新模型发布之前发布的基准测试上表现往往好得惊人，而在新模型发布后不久发布的基准测试上表现则低得多。这始终是一个信号，表明该模型在较早的基准测试上进行了“后训练”——有些人称之为“基准刷分”。而且这被普遍接受。

有没有办法走出你所说的这场基准测试危机？

让我给你讲个故事：1968年墨西哥城奥运会上，有一位美国跑者被极度看好赢得1500米比赛。他意外地输给了一位来自肯尼亚高地的跑者，而且差距很大。那位肯尼亚跑者一直在高海拔训练，为在海拔7500英尺处比赛做了更好的准备。这引发了一些争议：我们应该禁止高海拔训练吗？最终共识是，让所有人都练吧。

问题不在于人们针对基准测试进行优化——而在于他们优化的程度不同。如果每个人都优化，或者说都“高海拔训练”，绝对数字会被夸大，但你可以恢复排名的有效性。

人工智能基准测试的“高海拔训练”会是什么样？

我们事先不知道这些模型里装了什么，但我们可以用与基准测试相关的数据对它们进行少量训练，只是为了确保我们在一定程度上拉平了竞争环境。

在这种新的基准测试视角下，你不会有训练集。你会有一个测试集，还会有某种协议，在测试之前对每个模型进行准备——这就是高海拔训练。

但即使这可行，我们只是回到了起点：进步指标，但没有有效的测量。没有那个，我们怎么能了解模型是如何工作的？

这是一个非常有趣的问题。基准测试的胜利有些令人悲哀，因为我们似乎放弃了科学真正要做的：理解机制。但如果我们有某种宏大的数学理论来构建这些模型，我们就不需要基准测试了。

我们不知道这些模型是怎么做到的，基准测试也不会告诉我们。

但从长远来看，基准测试可以带来新的科学知识。CASP蛋白质折叠基准测试推进了科学，而没有先坚持一套关于蛋白质折叠的科学理论。这就是基准测试在科学中悖论式的力量。

英文来源：

He Wants AI Benchmarking To Be a Science, Not a Dark Art
Introduction
Benchmarks are the rulers of AI, in both senses of the word. As measuring sticks, they work according to an extremely simple logic: If you want to know what an AI model can do, devise a standardized test to gauge it. But AI benchmarks can also dominate the behavior of the entire field.
In 1987, the Defense Advanced Research Projects Agency’s first benchmark on automatic speech-to-text translation showed that state-of-the-art AI could not outperform so-called “dumb” statistical approaches. The poor performance helped to usher in an AI winter. For 25 years, interest in AI research waned and funding dried up.
Then, in 2012, another benchmark called ImageNet swung the pendulum back. An eight-layer neural network trounced other algorithms at classifying images and kicked off a “deep learning revolution.”
Nowadays, new AI benchmarks, geared at large language models such as OpenAI’s GPT-6 Astra, get created and solved with clockwork regularity, purporting to measure everything from domain-specific expertise to “general intelligence.” But rather than hail this trend as proof of AI’s skyrocketing potential, many in the field — skeptics and boosters alike — take it as evidence that benchmarking itself has become fatally flawed.
Moritz Hardt, a computer scientist and a director at the Max Planck Institute for Intelligent Systems, agrees with this diagnosis. Shortly after the ImageNet result, he began to think that machine learning systems could game leaderboards “without doing anything intelligent.” “There’s no two ways about it,” he said. “The benchmarking ecosystem is broken.”
In his 2026 book The Emerging Science of Machine Learning Benchmarks, Hardt makes the provocative case that these measuring sticks have been broken for decades already: Benchmarks, he writes, are “the mistake that made machine learning.” Notably, he argues, instead of creating a replicability crisis, they drove legitimate progress anyway — “mostly by sheer luck,” in his view. By understanding why, we can explain how benchmarks have since lost this accidental credibility in the age of LLMs, and how they might regain it.
Quanta spoke to Hardt about what went wrong, and right, with benchmarks, and where to take them next. The interview has been condensed and edited for clarity.
Let’s get the basics straight: What even is a benchmark?
The recipe is simple. To evaluate a machine learning model, you take your data and split it into a training set and a test set. You let the model builders do whatever they want on the training set: Anything goes, no rules. But you use the test set only to compare and rank models — to have a leaderboard say, “This is currently state-of-the-art, this is number one, this is number two.”
So a benchmark for a large language model today, like GPT-6 Astra or Claude Fable, might test an ability to solve difficult math problems or find security bugs in software.
Right. The first benchmark, in fact, was created in 1959. It was a set of digits and letters printed on punch cards, meant to test how well different scanners could recognize written characters. Somebody introduced the train-test split and started a leaderboard.
It wasn’t like there was some master plan. But that’s what we now call benchmarking, and it’s been the engine of progress in AI for about 60 years or so.
What are benchmarks supposed to tell us?
Traditionally, a benchmark is supposed to be like a thermometer. People train a model to do something, and then you come up with some kind of measurement procedure to see how good the model is.
But in practice, benchmarking is more like a sports competition or a market. Benchmarks define the rules of the game, and people actively compete over them, and we let the best model win.
Can’t they be both? What’s the problem?
Any machine learning textbook will tell you about the “holdout method” — that you should lock away the test set, and only use it once to rank the models. If you ignore this, you’re on thin ice. You might get very misleading results because you’re evaluating models that in some sense have knowledge of the test set.
The fascinating thing is that machine learning as a discipline just ignored the warning. It took all these test sets and put them on the internet for everyone to use. It’s easy to put these test sets in a model’s development loop and to keep iterating on them.
Is that what’s “broken” about benchmarks?
If you look at a model’s accuracy numbers on one benchmark, and then you change the test set only slightly, you get completely different accuracy numbers. That’s the sense in which the benchmark is not working. If you have 90% accuracy on one dataset and 73% on another, which is it? I have no idea how good this model is.
OpenAI, Anthropic, and other technology companies still advertise benchmark numbers as a measurement of their models’ reasoning capabilities. Is that just nonsense?
These scores really get misused. If you say, “On this math benchmark, the model got a 90%, which means that the model is now equivalent to an expert mathematician,” you’re reading something into that number that gives you the wrong takeaway. We should avoid comparisons or conclusions that implicitly require the absolute number to make sense.
If the absolute numbers don’t make sense, how do you not end up with a replication crisis?
This is where things get really interesting. If, instead of absolute numbers, you look at rankings — relative comparisons between models — those rankings tend to be very stable across different datasets. And this has been pretty consistent across the history of machine learning. If one model gets 90% on a particular test set, and another model gets 80%, then the 90% model is the model you’d want to use in the real world. It’ll still do better on other datasets, even ones that are quite different than the test set.
That, to me, is the main reason why machine learning escaped a scientific crisis.
That’s surprising. So the 90% score doesn’t have much validity, but the fact that it “won” the competition does?
Yeah. So benchmarks tell you when you beat the previous best. And these improvement indicators are pretty reliable.
But it’s lucky that it worked as well as it did — we had no reason to expect this to be the case. And we don’t yet have a good comprehensive theory for why this happens. It’s mostly an empirical phenomenon.
But you also argue that now, in the era of large language models, benchmarks are no longer able to deliver this “accidental” reliability. Why not?
The stability of model rankings has one big caveat: Models need to train on the same training set. In the old days, this was always true. But we don’t know what LLMs were trained on. In fact, model builders can strategically assemble data mixes that best promote benchmark performance. It’s a little bit like studying to the test, which we all do in high school.
That’s why we’re currently in this benchmarking crisis — because it’s no longer true that the better model necessarily wins. A worse model might just put more effort into preparing for the specific benchmark.
This isn’t really a crime: It happens every time a new model is released. If you look at the notes on the latest version of Gemini or GPT, you’ll see that its performance on benchmarks that were released before the new model is often fantastic, while its performance on benchmarks released shortly after is way lower. That’s always an indication that the model was “post-trained” on the earlier benchmark — what some people call “benchmaxxed.” And it’s generally accepted.
Is there a way out of this benchmarking crisis, as you call it?
Let me tell you a story: At the 1968 Olympics in Mexico City, there was an American runner who was vastly favored to win the 1,500-meter race. He lost unexpectedly to a runner from the Kenyan highlands, by a wide margin. The Kenyan runner had trained at high elevation and was better prepared for competing at 7,500 feet above sea level. It caused some controversy: Should we outlaw altitude training? The consensus in the end was, just let everyone do it.
The problem is not so much that people optimize to the benchmarks — it’s that they optimize to different degrees. If everyone optimizes, or “altitude trains,” the absolute numbers will be inflated, but you can restore the validity of the rankings.
What would altitude training for AI benchmarks look like?
We don’t know ahead of time what went into these models, but we can train them for a little bit on benchmark-relevant data, just to make sure that we’re leveling the playing field somewhat.
In this new view of benchmarks, you would have no training set. You would have a test set, and you would also have some kind of protocol for preparing each of the models before you test them — that’s the altitude training.
But even if this works, we’re just back to where we started: progress indicators, but no valid measurements. Without that, how can we learn anything about how the models work?
This is a really interesting question. There’s something sad to the triumph of benchmarking, because it seems like we’re giving up what science is actually about: understanding the mechanisms. But if we had some kind of grand mathematical theory to build these models with, we wouldn’t need the benchmarks.
We don’t know how these models do it, and benchmarking won’t tell us.
But benchmarking can, in the long run, lead to new scientific knowledge. The CASP protein-folding benchmark advanced science without first insisting on a scientific theory of protein folding. That’s the paradoxical power of benchmarking in the sciences.