Microsoft Research releases Echoverse, a framework for training computer-use agents in deep, evolving environments

Microsoft Research introduced Echoverse, a framework designed to train computer-use agents in deep, evolving environments rather than static benchmarks. The approach emphasizes fidelity and state coherence, with twelve training worlds—ten deep domain worlds and two capability worlds—each targeting specific interaction patterns such as date pickers and nested filters.

In experiments, a 9-billion-parameter model trained on all twelve worlds improved its base score from 36.5% to 67.1%, narrowing the gap to GPT-5.4 by fourteen points. The authors report that high simulation fidelity is critical; models trained on shallow versions of the same environments regressed on those shallow tasks while improving on the deep ones.

Echoverse implements a co-evolution loop where each graded run both evaluates the agent and identifies weaknesses in the environment, tasks, or verifier. Reinforcement learning with grounded verifiers is used as a reward signal, pushing agents to reach goals in fewer steps and improving held-out performance.

The team released four of the twelve worlds as open-source artifacts, including code, data, and grounded graders, to support further research into high-fidelity computer-use environments. The resources are available on GitHub and Hugging Face under the Microsoft Echoverse repository.