EvoCowork is a 240‑task benchmark designed to evaluate recursive self‑improvement at the harness level across four office modalities while keeping model weights fixed. The benchmark spans presentations, documents, spreadsheets, and cross‑office workflows, scoring artifacts on content, native structure, formulas, cross‑file consistency, and visual quality where applicable.
Its protocol separates feedback, revision selection, and final testing: methods first learn on single‑modality tasks, then are evaluated on unseen single‑modality tasks and cross‑office workflows using a shared initial harness and explicit budgets. In static evaluations of two models and four harnesses, the best configuration achieves an overall score of 42.25 and 30.62 on cross‑office tasks. Starting from the same MSA harness with GLM‑5.3, ACE and AHE improve in‑domain scores by 4.90 and 2.64 points, respectively, but change cross‑office scores by only +0.03 and -1.72. This shows that gains on file types that provide feedback do not necessarily transfer to coordinated workflows. EvoCowork makes this distinction explicit while recording capability, transfer, and the cost of harness evolution.
Code and task packages: https://github.com/DeepThinkingZhouLiu/cowork-evolution-benchmark
Review