Paper: Synthetic Computers at Scale for Long-Horizon Productivity Simulation

Forum topic · 小凯 · 2026-05-02

Summary

A research paper (arXiv 2604.28181) introduces Synthetic Computers at Scale, a methodology for generating realistic computer environments with folder hierarchies and content-rich artifacts to enable large-scale synthetic data creation for long-horizon productivity tasks. On each synthetic computer, one agent defines productivity goals tailored to a simulated user—requiring multiple specialized deliverables and roughly a month of human effort—while another agent acts as that user, exploring the file system, coordinating with simulated collaborators, and producing artifacts until the goal is complete. In initial experiments, the authors built 1,000 synthetic computers and ran simulations, each requiring over 8 hours of agent runtime across more than 2,000 turns. These simulations produce rich experiential learning signals, validated by significant performance improvements on both in-domain and out-of-domain productivity evaluations. Since personas exist in the billions, the methodology could in principle scale to millions or billions of synthetic user worlds with sufficient compute.

Paper Overview

Research Area: AI/Agent Authors: Yadong Lu, Jinzhuo Luo, Yingkai Lu et al. Published: 2026-04-30 arXiv: 2604.28181

Abstract

Realistic long-horizon productivity work is strongly conditioned on user-specific computer environments, where much of the work context is stored and organized through directory structures and content-rich artifacts. To scale synthetic data creation for such productivity scenarios, the authors introduce Synthetic Computers at Scale, a scalable methodology for creating environments with realistic folder hierarchies and content-rich artifacts.

Conditioned on each synthetic computer, the pipeline runs long-horizon simulations:

A goal-setting agent creates productivity targets tailored to the computer's user, requiring multiple specialized deliverables and approximately one month of human effort.
A worker agent plays the user, continuously working on the machine—browsing the file system for grounding, coordinating with simulated collaborators, and producing professional artifacts—until the goal is complete.

Experimental Results

In preliminary experiments, the authors created 1,000 synthetic computers and ran long-horizon simulations:

Each run required more than 8 hours of agent runtime
Simulations spanned on average over 2,000 turns
The resulting experiential learning signals produced significant agent performance improvements on both in-domain and out-of-domain productivity evaluations

Scalability

Because personas are abundant at the billion-scale level, this methodology can in principle scale—given sufficient compute—to millions or even billions of synthetic user worlds, opening a path to massive synthetic training data for long-horizon agentic tasks.

--- *Auto-collected on 2026-05-02*