EvoCowork: A Benchmark for Real-World Office Tasks
We introduce EvoCowork, a 240-task benchmark for evaluating harness-level recursive self-improvement across four office modalities, with model weights held fixed.
ProofPaper ↗
Key points
- Office workflows are a demanding testbed for self-improving agents because their deliverables must be correct, natively editable, and mutually consistent.
- A central question is whether gains on feedback tasks reflect reusable procedures rather than task-specific fixes.
- Its protocol separates feedback, revision selection, and final testing: methods learn from single-modality tasks, then are evaluated on unseen single-modality tasks and cross-office workflows under a shared initial harness and explicit budgets.
- 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.
Sources (1)
- [1]EvoCowork: A Benchmark for Real-World Office TasksarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 11:43 AM
We introduce EvoCowork, a 240-task benchmark for evaluating harness-level recursive self-improvement across four office modalities, with model weights held fixed.
Office workflows are a demanding testbed for self-improving agents because their deliverables must be correct, natively editable, and mutually consistent.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 10, 2026EviScene: Beyond "Is It AI?" for Agentic Image Factuality Assessment
- Oct 10, 2026LightVAE: Towards Compact and Efficient Video Autoencoders
- Oct 9, 2026Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction
- Oct 9, 2026Latent World Action Model for GUI Agents
- Oct 8, 2026openai/codex rust-v0.162.0: 0.162.0
- Oct 8, 2026Understanding Jev, the new model everyone is talking about