The paper was submitted to arXiv on October 25, 2026.
CLE contains 120 human-authored tasks arranged as five-level ladders, for 600 levels in total. It covers 89 professional roles across 12 occupational domains. An agent must pass a level to unlock the next one, and later assignments build on earlier work. More than 50 domain experts contributed to the benchmark. Submissions are assessed by an evaluation agent against human-authored rubrics.
The authors tested 17 harness-model configurations using frontier LLMs. Providing diagnostic feedback after failure and allowing one revision produced a median completion gain of 5.3 percentage points. The best system completed 47.7% of levels and completed all five levels in 23.3% of tasks.
The paper does not establish that these results represent performance in every workplace, nor that feedback produces durable learning outside CLE. It is an arXiv preprint, and its scores depend on this benchmark's tasks, rubrics, and evaluation-agent setup.