Echoverse: Deep, Evolving Environments for Training Computer-Use Agents at Scale
Echoverse reports a 9B computer-use agent rising from 36.5% to 67.1% after training in stateful synthetic apps that evolve with the model.
The paper argues that agent-training environments need behavioral depth, targeted failure coverage, and repair loops, not just scale. Echoverse compiles specs into apps with database-grounded grading, then uses each rollout both to improve the environment and to train the model. In the reported tests, shallow environments hurt live-site accuracy, while deeper ones improved it; repairing one environment lifted its trained model from 16.2% to 38.5%. The authors also release four environments as a benchmark with apps, seed data, and graders. HF Daily Papers' note
The paper argues that agent-training environments need behavioral depth, targeted failure coverage, and repair loops, not just scale. Echoverse compiles specs into apps with database-grounded grading, then uses each rollout both to improve the environment and to train the model. In the reported tests, shallow environments hurt live-site accuracy, while deeper ones improved it; repairing one environment lifted its trained model from 16.2% to 38.5%. The authors also release four environments as a benchmark with apps, seed data, and graders. HF Daily Papers' note
score 5