Researchers have built a way to shrink AI agent benchmarks by up to 40x without losing much accuracy.
Evaluating AI agents is far pricier than testing a standard language model, since agents run multi-step trajectories instead of answering single prompts. The paper's authors analyzed large-scale agent trajectories and identified six process signals, patterns in how an agent works through a task, that reliably predict final performance. Their method, called DualViewEval, combines those process signals with outcome data to pick a small, representative set of tasks and then predict scores on the full benchmark. Tested across five agent benchmarks against five existing compression baselines, it beat every one of them, hitting 24x to 40x compression on APEX-Agents and BFCL using just 20 tasks.
That's a real cost saving as more labs ship agents that browse, code, and call tools autonomously, each requiring longer and pricier evaluation runs. DualViewEval also cut prediction error by 14.5% to 28.2% against the strongest rival methods, and its selected task sets surfaced capability gaps between agents that score-only compression missed.
It's a useful fix for a real bottleneck, but a 40x cheaper test suite is still a proxy. Teams shipping agents will still want to run the full benchmark before betting real money on the result.