Reproduce “The Evolution of AI Office: From Passive Assistants to Goal-Driven Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yihao Liu, Yongsiqi Tu, Haoxuan Ma, Wen Wang, Yang Hao, Chang Liu, Li Wang, Huadong Li and 7 more
A framework-driven survey unifies AI systems for documents, spreadsheets, and presentations under the concept of “AI Office,” treating the committed, editable artifact—not the conversational response or isolated action—as the primary unit of evaluation. Analysis of office-assistant requests shows a mismatch between user needs and current benchmarks: execution and deliverable-generation tasks dominate demand, while evaluation often emphasizes question answering and comprehension.
Key insights
- Documents, spreadsheets, and slides share an underlying artifact-management loop, but require different validators: executable checks for spreadsheets, visual and structural checks for slides, and provenance and long-range consistency checks for documents.
- Reliable office agents require a model-driven harness that manages intent, memory, reusable skills, permissions, routing, verification, recovery, checkpoints, and cross-file dependencies rather than relying only on application-local features.
- Artifact states, rendered outputs, execution logs, diffs, citations, and cross-file comparisons can provide training rewards, but composite objectives and regression testing are needed to prevent improvements in one quality dimension from damaging another.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.