Microsoft Research presents Echoverse, a suite of twelve training worlds for computer-use agents: ten deep domain environments and two capability-focused worlds targeting controls such as date pickers and nested filters. The environments reproduce application behavior, contain realistic seeded data, and preserve coherent state across screens and users. A 9B model trained across all twelve worlds improved from 36.5% to 67.1%, reportedly coming within 14 percentage points of GPT-5.4. The study emphasizes simulation fidelity, varied control practice, and co-evolution of models, worlds, and verifiers.
No heat snapshots are available in the last 24 hours.