Synthetic Computers at Scale for Long-Horizon Productivity Simulation

Forum topic · 小凯 · 2026-05-02

Summary

Researchers introduce Synthetic Computers at Scale, a scalable methodology for generating realistic, user-specific computer environments featuring authentic folder hierarchies and content-rich artifacts. On top of each synthetic computer, they run long-horizon simulations in which one agent creates productivity goals tailored to the computer's owner—typically requiring multiple specialized deliverables and roughly a month of human effort—while another agent plays that user, continuously working toward the goal by browsing the file system for context, coordinating with simulated collaborators, and producing professional artifacts. In preliminary experiments, the team created 1,000 synthetic computers and ran long-horizon simulations, each requiring over 8 hours of agent runtime and averaging more than 2,000 turns. The simulations yield rich experiential learning signals, validated by significant agent performance improvements on both in-domain and out-of-domain productivity evaluations. Since personas are abundant at the billions scale, the methodology could in principle extend to millions or billions of synthetic user worlds given sufficient compute. Full paper: arXiv:2604.28181.

Overview

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

English Translation of the Post Summary

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, they run long-horizon simulations:

Goal-setting agent: creates productivity goals tailored to the computer's user, requiring multiple specialized deliverables and roughly a month of human effort.
User agent: plays that user, continuously working on the computer—browsing the file system for grounding information, coordinating with simulated collaborators, and producing professional artifacts—until the goal is completed.

Key Results

Created 1,000 synthetic computers and ran long-horizon simulations in preliminary experiments.
Each simulation run required over 8 hours of agent runtime, spanning on average more than 2,000 turns.
The simulations produce rich experiential learning signals, validated by significant agent performance improvements on both in-domain and out-of-domain productivity evaluations.
Since personas are abundant at the scale of billions, the methodology can in principle scale to millions or even billions of synthetic user worlds with sufficient compute.