Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale

Authors: Yash Pandya, Sahil Gupta, Sarthak Harne, Archana Yadav, Kavyansh Chourasia, Hussein Mozannar, Vibhav Vineet, Sara Abdali, Corby Rosset, Yash Lara, Ahmed Awadallah, Ece Kamar, Akshay Nambi (Microsoft Research)

arXiv ID: 2607.28074

Problem: Computer-use agents learn from what their actions change, so training them needs applications they can act on, break and reset. The most important real applications are login-gated and stateful, so synthetic environments stand in; but recent pipelines generate environments in bulk, moving the bottleneck from how many exist to what is inside each one.

Key Methodology:

Echoverse compiles specifications into stateful applications whose tasks are graded against the application's own database (grounded verifiers)
A co-evolution loop reads every graded rollout twice: as repairs to the environment, its tasks and its verifier, and as training signal for the model
Ablations isolate three properties: behavioral depth, targeting the interaction the agent actually fails, and improvement alongside the model
Twelve environments train a 9B model; four environments release as a benchmark with applications, seed data and grounded graders

Key Results:

The 9B model improves from 36.5% to 67.1% across fourteen evaluation splits, within fourteen points of the much larger frontier model that taught it
On the same domains, shallow environments push live-site accuracy below the base model (80.0 to 75.0) while deep ones raise it (80.0 to 85.0 and 48.0 to 65.0)
Drilling one interface control across many renderings transfers to held-out widget families and the open web
Repairing a single environment lifts the model trained on it from 16.2% to 38.5%
In RL training, a reward combining the grounded verifier with a dense per-step judge raises held-out score from 58.8% to 68.0%

What it means for developers: The quantity of training environments matters less than their depth and how they evolve: a small set of stateful, verifier-graded, self-repairing environments can carry a small model near frontier performance, while shallow environments can actively regress it. For anyone building GUI or computer-use agents, this argues for investing in few, deep, database-grounded training worlds with repair loops over generating more environments, and for checking live transfer rather than just environment scores.

Paper: arXiv:2607.28074