Enterprise agents are often asked to answer questions over workplace records that shift as teams collaborate. A new arXiv preprint introduces MMEnterprise-Bench, a benchmark designed specifically for this setting. The benchmark centers on records that evolve over time, with information shared, corrected, and revised across documents, chats, and figures.
Because the same fact may appear in different versions in different places, an agent must track which version is current and reconcile conflicting updates. MMEnterprise-Bench appears to be built to test exactly that behavior, though the abstract does not yet describe the exact task formats or scoring methods.
The work is posted as a new announcement on arXiv, indicating it has not yet been peer reviewed. No other sources are available for comparison, so this description relies solely on the paper's abstract.