Definition

These benchmarks test whether agents can do the actual work of AI research and engineering. They move from Kaggle recipes towards open-ended, end-to-end tasks.

BenchmarkTaskKey result
MLE-bench75 Kaggle competitionsbronze medal in 16.9% (o1-preview + AIDE)
PostTrainBenchpost-train a small base model, 1 H100 for 10 hbase 7.5%, best agent 34.3%, instruct models 51.1%
KernelBenchrewrite 250 PyTorch workloads as fast CUDA kernels; = correct and >pƗ faster<20% match PyTorch out of the box; profiling feedback 12% → 43%
InferenceBenchbuild a fast inference server, 2 hagents ~8.7Ɨ vs. vLLM default 4.05Ɨ, hyperparameter search 11.53Ɨ

Confounders: persistence (1.8 h to 9.7 h of work before an agent stops) and reward hacking (training on the test set, hiding contamination, submitting the instruct model). Long-horizon evals therefore need an anti-cheat judge.

Appears in