Echoverse: Deep, evolving environments for computer-use agents
Microsoft says high-fidelity synthetic apps, not shallow clones, made its 9B computer-use agent markedly better.
Echoverse built twelve training worlds with real state, database-grounded grading, and tasks aimed at weak agent skills such as date pickers and nested filters. Trained across them, the 9B model rose from 36.5% to 67.1%, within fourteen points of GPT-5.4 on the reported average. Shallow versions hurt or stalled performance, while deeper worlds and targeted control practice transferred to held-out layouts and some live-web benchmarks. Microsoft is releasing four worlds: EchoStay, EchoForge, datepicker, and nested-filter. Microsoft Research AI's note
Echoverse built twelve training worlds with real state, database-grounded grading, and tasks aimed at weak agent skills such as date pickers and nested filters. Trained across them, the 9B model rose from 36.5% to 67.1%, within fourteen points of GPT-5.4 on the reported average. Shallow versions hurt or stalled performance, while deeper worlds and targeted control practice transferred to held-out layouts and some live-web benchmarks. Microsoft is releasing four worlds: EchoStay, EchoForge, datepicker, and nested-filter. Microsoft Research AI's note
score 5