Echoverse: Deep Evolving Environments for Training Computer-Use Agents
Microsoft
Microsoft Research introduces Echoverse, a set of twelve synthetic worlds designed to train computer-use agents through high-fidelity simulations. A 9B model trained on these worlds nearly doubled its performance, highlighting the importance of depth over quantity in agent training.
Microsoft Research developed Echoverse, a collection of twelve synthetic worlds for training computer-use agents, comprising ten deep domain worlds and two capability worlds. These worlds replicate real application behavior with coherent state and realistic data, allowing agents to learn from genuine consequences. Training a 9B model on all twelve worlds improved its score from 36.5% to 67.1%, approaching GPT-5.4's performance. Key findings include the necessity of high simulation fidelity, the effectiveness of drilling challenging UI elements like date pickers, and the benefits of co-evolving the model, world, and verifier. Reinforcement learning using grounded verifiers as rewards outperformed imitation learning. Microsoft is releasing four worlds with code, data, and graders to support further research.
- Сокращения
- SFT = Supervised Fine-Tuning — Supervised Fine-Tuning
- RL = Reinforcement Learning — Reinforcement Learning
- SQL = Structured Query Language — Structured Query Language
Source: Microsoft Research —
original