Its not really that surprising when models are trained on the exams

I work in AI evaluation, lots of problems and leakage is an issue as is ecological validity, but they definitely do not explain the progress we see.

I think Epoch has the best analysis I’ve seen on evaluation trends; they use IRT to basically model a variety of benchmark difficulties, and then model a capability parameter for each model. This is as robust a sort of “meta-study” of evaluations as I’ve seen and the trend in capabilities show no sign of slowing down.

So I think people’s feelings clash with reality, and that’s because releases are more frequent and the jumps between releases are smaller, but the growth in capabilities _over time_ has not changed for the better or worse over a very very long period of time.

Benchmarks saturate around 80-90%?

This is not "Acing" a test, this is hitting a wall.

Even on very small tests a fraction of questions might have wrong answers in the key.

If models can't get more than 90% of the benchmark right I think it's a strong indication that they were not trained on the answers and that benchmark itself is messy enough that <10% desired answers might be wrong or misleading.

Yea this may explain part of it or all of it, it’s likely a case by case kind of thing.

Also to respond to the parent comment: benchmarks have a variety of difficulty levels. Humanity’s Last Exam, though now hitting the beginning of a saturation phase with Fable, was long unsaturated while other benchmarks saturated awhile ago. So that’s what I meant by Epoch capability index: using IRT models this effect so that you gather robust signals from variety of benchmark difficulties and can track progress over time as model capabilities have evolved (and so benchmarks have had to evolve to keep up).

But yes like I was saying: all benchmarks are problematic, some are useful. Benchmark quality problems abound, so 90% being the true ceiling is not surprising. There may be other factors at play here too, I haven’t studied this problem that deeply to have a good thorough answer to this. But keep in mind there are probably 50,000 benchmarks in the literature and that is not a joke number. A crapload of noise in that signal but it’s not all noise.

Saturation is mostly just selection effects in play. Throw out the "90% easiest" of tasks, and what remains is a jagged ladder of high difficulty outliers.

Hard to climb, and hard to measure the climb - because you have less effective data points and the datapoints themselves are less linear, while you're still being subject to the measurement noise.

Not having the mislabeled tasks would reduce the saturation, but it wouldn't drive it to zero. Even without the "infinite difficulty tasks", bell curve would do its thing.

You almost can't give a correct answer to a task with expected wrong answer unless you know it expects wrong answer. ... maybe on a yes/no question by chance. But as you become more knowledgeable, the probability that you give wrong but expected answer reduces.

If a benchmark saturates to 100% it's very likely that answers leaked into the training data.

In college I had a funny exam. It was on C++. One question I had to answer incorrectly because there was a mistake in the question. So I gave two answers for it, one that answered the question as it was and the other that answered the question as I inferred it was intended to be. It was appreciated. I got a honorary mark above the top possible (I gave correct answers to all other questions). I wouldn't be surprised if across so many, so huge benchmarks, there were tasks with wrong questions or answers in the key, that some LLM answered in and expected manner in the same fashion.

Can't call it AI like that without discrediting yourself. You mean LLMs?

Jumping in here, frankly I hate the trend of calling every type of automation intelligence.

Most "AI" is really an optimization algorithm in software tools, same as its always been. This really isnt anything new, aside from adding a chatbot / MCP interface to the same tools.

When a Big Killing Robot comes to murder you be sure to always call it BKR and don't discredit yourself by calling it AI.

That's a bit hyperbolic when we're all just posting on HN

Talk about moving the goalposts. Pray tell, exactly what must an LLM do before you're willing to consider it AI? Be specific, otherwise you're just woo-mongering.

Everyone here is talking about LLMs, why bother calling them something else

I feel like LLMs refers to a model that for a given message responds with its own message.

And I feel like AI agents (powered by some LLM) are qualitatively something else.