He Wants AI Benchmarking To Be a Science, Not a Dark Art

The computer scientist Moritz Hardt calls benchmarks “the mistake that made machine learning.” He’s studied where they went wrong, and why they’ve managed to drive progress anyway.

计算机科学家莫里茨·哈特 (Moritz Hardt) 将基准测试称为“机器学习的错误”。他研究了他们哪里出了问题，以及为什么他们仍然能够推动进步。

Benchmarks are the rulers of AI, in both senses of the word.

从两个意义上来说，基准都是人工智能的统治者。

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.

作为衡量标准，它们按照极其简单的逻辑工作：如果你想知道人工智能模型能做什么，就设计一个标准化测试来衡量它。

But AI benchmarks can also dominate the behavior of the entire field.

但人工智能基准也可以主导整个领域的行为。

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.

1987 年，国防高级研究计划局的第一个自动语音到文本翻译基准表明，最先进的人工智能无法胜过所谓的“愚蠢”统计方法。

The poor performance helped to usher in an AI winter.

糟糕的表现帮助迎来了人工智能的冬天。

For 25 years, interest in AI research waned and funding dried up.

25 年来，人们对人工智能研究的兴趣逐渐减弱，资金也枯竭。

Then, in 2012, another benchmark called ImageNet swung the pendulum back.

然后，在 2012 年，另一个名为 ImageNet 的基准测试又扭转了局面。

An eight-layer neural network trounced other algorithms at classifying images and kicked off a “deep learning revolution.”

八层神经网络在图像分类方面击败了其他算法，并引发了一场“深度学习革命”。

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 .

如今，针对 OpenAI 的 GPT-6 Astra 等大型语言模型的新 AI 基准测试有规律地创建和解决，旨在衡量从特定领域的专业知识到“通用智能”的一切。但该领域的许多人（无论是怀疑论者还是支持者）并没有将这一趋势视为人工智能飞速发展潜力的证明，而是将其视为基准测试本身存在致命缺陷的证据。

Moritz Hardt , a computer scientist and a director at the Max Planck Institute for Intelligent Systems, agrees with this diagnosis.

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 (Moritz Hardt) 同意这一诊断。

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.

在 ImageNet 结果出来后不久，他开始认为机器学习系统可以“不做任何智能的事情”来玩排行榜。 “没有两种方法可以解决这个问题，”他说。

“The benchmarking ecosystem is broken.”

“基准测试生态系统已被破坏。”

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.

哈特在其 2026 年出版的《机器学习基准的新兴科学》一书中提出了一个颇具争议性的论点，即这些衡量标准已经被打破了数十年：他写道，基准是“导致机器学习的错误”。值得注意的是，他认为，他们并没有造成可复制性危机，反而推动了合法的进步——在他看来，“主要是靠运气”。

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.

通过理解原因，我们可以解释基准如何在法学硕士时代失去这种意外的可信度，以及它们如何重新获得它。

Quanta spoke to Hardt about what went wrong, and right, with benchmarks, and where to take them next.

广达与哈特讨论了基准的错误和正确之处，以及下一步的发展方向。

The interview has been condensed and edited for clarity.

为了清晰起见，采访内容已经过精简和编辑。

Hardt is fascinated by how benchmarks paradoxically helped start both a 25-year AI winter in the late 1980s and a deep learning revolution in the 2010s.

Hardt 着迷于基准测试如何自相矛盾地帮助开启了 20 世纪 80 年代末长达 25 年的人工智能寒冬和 2010 年代的深度学习革命。

The recipe is simple.

食谱很简单。

To evaluate a machine learning model, you take your data and split it into a training set and a test set.

要评估机器学习模型，您需要获取数据并将其分为训练集和测试集。

You let the model builders do whatever they want on the training set: Anything goes, no rules.

您可以让模型构建者在训练集上做任何他们想做的事：任何事情都可以，没有规则。

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.”

但你使用测试集只是为了比较和排名模型 - 让排行榜说，“这是目前最先进的，这是第一，这是第二。”

There’s no two ways about it.

没有两种方法可以解决这个问题。

The benchmarking ecosystem is broken.

基准测试生态系统已被破坏。

Right.

正确的。

The first benchmark, in fact, was created in 1959.

事实上，第一个基准创建于 1959 年。

It was a set of digits and letters printed on punch cards, meant to test how well different scanners could recognize written characters .

这是一组印在打孔卡上的数字和字母，旨在测试不同扫描仪识别书面字符的能力。

Somebody introduced the train-test split and started a leaderboard.

有人引入了训练-测试的划分并创建了排行榜。

It wasn’t like there was some master plan.

好像没有什么总体规划。

But that’s what we now call benchmarking, and it’s been the engine of progress in AI for about 60 years or so.

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。

Traditionally, a benchmark is supposed to be like a thermometer.

传统上，基准应该像温度计一样。

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.

人们训练模型来做某事，然后你想出某种测量程序来看看模型有多好。

But in practice, benchmarking is more like a sports competition or a market.

但在实践中，标杆管理更像是一场体育比赛或市场。

Benchmarks define the rules of the game, and people actively compete over them, and we let the best model win.

基准定义了游戏规则，人们积极地竞争它们，我们让最好的模型获胜。

Hardt cautions that the scores AI models receive on benchmarks don’t mean much.

哈特警告说，人工智能模型在基准测试中获得的分数并没有多大意义。

But comparisons, or rankings, among models do.

但模型之间的比较或排名确实如此。

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.

任何机器学习教科书都会告诉你“保留方法”——你应该锁定测试集，并且只使用它一次来对模型进行排名。

If you ignore this, you’re on thin ice.

如果你忽视这一点，你就会如履薄冰。

You might get very misleading results because you’re evaluating models that in some sense have knowledge of the test set.

您可能会得到非常误导的结果，因为您正在评估在某种意义上了解测试集的模型。

The fascinating thing is that machine learning as a discipline just ignored the warning.

令人着迷的是，机器学习作为一门学科却忽略了这个警告。

It took all these test sets and put them on the internet for everyone to use.

它采取了所有这些测试集并将它们放在互联网上供每个人使用。

It’s easy to put these test sets in a model’s development loop and to keep iterating on them.

将这些测试集放入模型的开发循环中并不断迭代它们很容易。

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.

如果您在一个基准上查看模型的准确度数字，然后仅稍微更改测试集，您就会得到完全不同的准确度数字。

That’s the sense in which the benchmark is not working.

从这个意义上说，基准测试不起作用。

If you have 90% accuracy on one dataset and 73% on another, which is it?

如果一个数据集的准确率是 90%，而另一个数据集的准确率是 73%，那么哪个数据集准确率高？

I have no idea how good this model is.

我不知道这个模型有多好。

These scores really get misused.

这些分数确实被滥用了。

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.

如果你说，“在这个数学基准上，模型得到了 90%，这意味着该模型现在相当于专家数学家”，那么你从这个数字中得到的信息是错误的。

We should avoid comparisons or conclusions that implicitly require the absolute number to make sense.

我们应该避免隐含地需要绝对数字才有意义的比较或结论。

“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,” Hardt said.

“基准测试的胜利令人悲伤，因为我们似乎正在放弃科学的真正意义：理解机制，”哈特说。

“But benchmarking can, in the long run, lead to new scientific knowledge.”

“但从长远来看，基准测试可以带来新的科学知识。”

This is where things get really interesting.

这就是事情变得非常有趣的地方。

If, instead of absolute numbers, you look at rankings — relative comparisons between models — those rankings tend to be very stable across different datasets.

如果您查看排名（模型之间的相对比较）而不是绝对数字，那么这些排名在不同数据集上往往非常稳定。

And this has been pretty consistent across the history of machine learning.

这在机器学习的历史上是非常一致的。

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.

如果一个模型在特定测试集上获得 90%，而另一个模型获得 80%，那么 90% 的模型就是您想要在现实世界中使用的模型。

It’ll still do better on other datasets, even ones that are quite different than the test set.

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完全不同的数据集。

That, to me, is the main reason why machine learning escaped a scientific crisis.

对我来说，这就是机器学习逃脱科学危机的主要原因。

We’re currently in this benchmarking crisis — because it’s no longer true that the better model necessarily wins.

我们目前正处于这场基准危机中——因为更好的模型一定会获胜已经不再是事实。

Yeah.

是的。

So benchmarks tell you when you beat the previous best.

因此，基准测试会告诉您何时击败之前的最佳成绩。

And these improvement indicators are pretty reliable.

而且这些改进指标是相当可靠的。

But it’s lucky that it worked as well as it did — we had no reason to expect this to be the case.

但幸运的是，它的效果很好——我们没有理由期望会出现这种情况。

And we don’t yet have a good comprehensive theory for why this happens.

我们还没有一个很好的综合理论来解释为什么会发生这种情况。

It’s mostly an empirical phenomenon.

这主要是一种经验现象。

The stability of model rankings has one big caveat: Models need to train on the same training set.

模型排名的稳定性有一个很大的警告：模型需要在相同的训练集上进行训练。

In the old days, this was always true.

在过去，这总是正确的。

But we don’t know what LLMs were trained on.

但我们不知道法学硕士接受的培训是什么。

In fact, model builders can strategically assemble data mixes that best promote benchmark performance.

事实上，模型构建者可以战略性地组合最能提升基准性能的数据组合。

It’s a little bit like studying to the test, which we all do in high school.

这有点像我们在高中时所做的准备考试。

That’s why we’re currently in this benchmarking crisis — because it’s no longer true that the better model necessarily wins.

这就是我们目前陷入这场基准危机的原因——因为更好的模型一定会获胜已经不再是事实。

A worse model might just put more effort into preparing for the specific benchmark.

更糟糕的模型可能只是投入更多的精力来准备特定的基准。

This isn’t really a crime: It happens every time a new model is released.

这并不是真正的犯罪：每次发布新型号时都会发生这种情况。

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.

如果您查看最新版本的 Gemini 或 GPT 的注释，您会发现它在新型号之前发布的基准测试中的性能通常非常出色，而在新型号之后不久发布的基准测试中的性能则要低得多。

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.

这始终表明该模型是根据早期基准进行“后训练”的——有些人称之为“benchmaxxed”。并且这是普遍接受的。

Hardt says that AI models should receive “altitude training” for benchmarks.

哈特表示，人工智能模型应该接受基准测试的“海拔训练”。

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.

我给大家讲一个故事：1968年墨西哥城奥运会上，有一位美国选手在1500米比赛中大获全胜。

He lost unexpectedly to a runner from the Kenyan highlands, by a wide margin.

他意外地以大比分输给了一位来自肯尼亚高地的选手。

The Kenyan runner had trained at high elevation and was better prepared for competing at 7,500 feet above sea level.

这位肯尼亚跑者接受过高海拔训练，为在海拔 7,500 英尺的比赛做好了更好的准备。

It caused some controversy: Should we outlaw altitude training?

这引起了一些争议：我们应该禁止高原训练吗？

The consensus in the end was, just let everyone do it.

最后的共识是，大家都去做吧。

The problem is not so much that people optimize to the benchmarks — it’s that they optimize to different degrees.

问题不在于人们针对基准进行优化，而在于他们优化的程度不同。

If everyone optimizes, or “altitude trains,” the absolute numbers will be inflated, but you can restore the validity of the rankings.

如果每个人都进行优化，或者“高原训练”，绝对数字将会被夸大，但你可以恢复排名的有效性。

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.

我们事先不知道这些模型的内容，但我们可以根据基准相关数据对它们进行一些训练，以确保我们在某种程度上公平竞争。

In this new view of benchmarks, you would have no training set.

在这种新的基准视图中，您将没有训练集。

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.

你会有一个测试集，并且在测试每个模型之前你还会有某种协议来准备它们 - 这就是高原训练。

This is a really interesting question.

这是一个非常有趣的问题。

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.

基准测试的胜利有些令人悲伤，因为我们似乎正在放弃科学的真正意义：理解机制。

But if we had some kind of grand mathematical theory to build these models with, we wouldn’t need the benchmarks.

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。

We don’t know how these models do it, and benchmarking won’t tell us.

我们不知道这些模型是如何做到这一点的，基准测试也不会告诉我们。

But benchmarking can, in the long run, lead to new scientific knowledge.

但从长远来看，基准测试可以带来新的科学知识。

The CASP protein-folding benchmark advanced science without first insisting on a scientific theory of protein folding.

CASP 蛋白质折叠基准先进了科学，但没有首先坚持蛋白质折叠的科学理论。

That’s the paradoxical power of benchmarking in the sciences.

这就是科学中基准测试的自相矛盾的力量。

Get highlights of the most important news delivered to your email inbox

获取发送到您的电子邮件收件箱的最重要新闻的摘要

Quanta Magazine moderates comments to facilitate an informed, substantive, civil conversation.

《Quanta》杂志对评论进行审核，以促进知情、实质性、文明的对话。

Abusive, profane, self-promotional, misleading, incoherent or off-topic comments will be rejected.

辱骂、亵渎、自我推销、误导、语无伦次或偏离主题的评论将被拒绝。

Moderators are staffed during regular business hours (New York time) and can only accept comments written in English.

版主在正常工作时间（纽约时间）有工作人员，并且只能接受英文评论。