Microsoft's Echoverse doubles 9B agent accuracy to 67.1%, nears GPT-5.4
Deep synthetic worlds lift a 9B model within 14 points of GPT-5.4—without live systems.
Microsoft Research's Echoverse tackles a fundamental bottleneck in training computer-use agents: shallow environments that fail to capture real consequences. Rather than just adding hundreds of low-fidelity worlds, Echoverse builds twelve deep worlds—ten focused on domains like email, banking, and cloud consoles, plus two capability worlds drilling tricky UI controls like date pickers and nested filters. These worlds keep state coherent across users and screens, meaning actions genuinely change data, fail realistically, and are graded from ground truth, not screenshots.
Results show depth matters: a 9B model trained on all twelve worlds doubled its base accuracy from 36.5% to 67.1%, coming within fourteen points of GPT-5.4. The team found that shallow versions of the same sites actually caused regression, while deep versions improved performance. Reinforcement learning with grounded verifiers further boosted held-out scores and reduced task steps. Microsoft released four worlds with code, data, and verifiers on GitHub and Hugging Face, fueling research on high-fidelity agent training environments.
- Nine-billion model jumped 36.5% to 67.1% accuracy after training on Echoverse's twelve deep worlds, landing within 14 points of GPT-5.4
- Ten domain worlds plus two capability worlds target weak UI skills (date pickers, nested filters), improving generalization to unseen domains
- Open-sourced four worlds with code, data, and grounded graders—not just environments but verifiers that reward true outcomes over appearance
- RL against grounded verifiers taught the agent to achieve goals in fewer steps while boosting held-out performance
Why It Matters
Echoverse proves deep, evolving synthetic environments—not just scale—can close the gap between small open models and frontier agents.