Microsoft Open-Sources Echoverse: High-Fidelity Worlds Boost 9B Agents to Rival GPT-5.4

Microsoft Research · rss · 2026-07-31

Microsoft Research introduced Echoverse, a set of deep, evolving environments designed to train computer-use agents. The project includes 12 training worlds (10 deep domain worlds and 2 capability worlds targeting specific UI controls like date pickers).

Key Concepts & Findings:

Fidelity Matters: Shallow worlds cause model regression, while deep builds significantly improve performance. Trained on all 12 worlds, a 9B model nearly doubled its score (36.5% to 67.1%), rivaling GPT-5.4.
Co-evolution: The system treats model training and the improvement of environments, tasks, and verifiers as a single loop. Each graded run generates both training data and environmental repairs.
RL over Imitation: Using grounded verifiers as rewards, reinforcement learning pushes agents past imitation, teaching them to reach goals in fewer steps.

Microsoft has open-sourced 4 of the worlds with their code, data, and grounded graders to support further research.

Original post →

Microsoft Open-Sources Echoverse: High-Fidelity Worlds Boost 9B Agents to Rival GPT-5.4

More from coding & agent