NeFut Logo NeFut
Admin Login

[CS.AI] WorkWorlds: An Infrastructure for Evaluating AI Agents on Workplace Tasks

Published at: 2026-09-23 22:00 Last updated: 2026-09-24 00:40
#algorithm #AI #Machine Learning

WorkWorlds introduces an evaluation infrastructure that decouples organizational state from task specification. A world is first fixed with a revision, date and employee seat, materialising the state accessible to that employee; tasks are added only afterwards. The authors implement the system in a synthetic pharmaceutical company covering six employee seats and eight tasks, and create additional organizational worlds. Across 192 matched evaluations, switching from task‑curated context to a full role‑visible workplace reduced evidence access from 90.4% to 74.5% and criterion pass from 79.4% to 68.2%, while the pass rate conditioned on evidence access remained almost unchanged, indicating that most of the gap occurs before the agent gathers sufficient evidence.

Review

Original Source: https://arxiv.org/abs/2609.23806

[h] Back to Home