原帖引用 · 原文
该语言的正文暂不可用,当前显示已有版本。
Rohan Paul@rohanpaul_ai
full technical report of FrontierML
Synthetic Sciences, a Y Combinator startup, has launched FrontierML, a benchmark so hard that none of the 5 frontier AI models it tested hit the strong-result bar on any of its 25 ML research problems.
They hide each objective, so agents must build their own tests and decide how much to trust them.
This is a test of whether AI agents can do machine learning research on their own without ever seeing the score they’ll be graded on.
Each agent gets 12 hours and real GPUs, and has to work out for itself how to measure progress.
FrontierML runs each task as a loop with the objective kept out of reach:
> The agent gets an open-ended ML problem and 12 hours in Lab0
> It designs a local proxy for progress
> It runs experiments on H100s and calibrates how far that proxy can be trusted
> It submits a model or method
> FrontierML scores it on a held-out evaluation
Synthetic Sciences argues step 3 is where much of empirical research happens. Here, a lot of the score rides on it.
在 X 查看这条帖子 ↗
Introducing FrontierML, a benchmark for evaluating AI agents on autonomous empirical ML R&D.
Most existing benchmarks for ML "research" give the agent a validation set or verifier that closely tracks its final score. In practice, human researchers rarely have that. Much of empirical research is deciding how to measure progress on a problem and how far to trust those measurements.
In FrontierML's 25 open-ended problems, agents have to operationalize progress through their own experiments, constructing local proxies and calibrating how far to trust them, while working toward an objective they never directly observe. Their submitted model or method is then scored on a held-out evaluation.
在 X 查看这条帖子 ↗
· UTC+8