How is Echoverse different from static computer-use benchmarks like WebArena?

Echoverse's environments EVOLVE alongside the agent — task graders, world dynamics, and difficulty curve up together, so a policy cannot memorise a fixed set. Microsoft reports zero-shot transfer to held-out real-web suites (WebVoyager 66.5→71.5, Online-Mind2Web 40.5→43.4), which is the signal that the co-evolutionary loop produces general skill, not overfit trajectories.

What can I actually download from Microsoft's Echoverse release?

Echoverse ships the training framework (github.com/microsoft/Echoverse, MIT-licensed) and a public Hugging Face dataset (microsoft/Echoverse) covering 722 tasks across six environment families — datepicker, nested-filter, email, customer support and two capability-focused worlds. The 22-page tech report at aka.ms/echoverse-techreport documents the training loop, EchoBank grader, and results.

How does an Echoverse-trained 9B compare to GPT-5.4 on computer-use tasks?

After co-evolutionary training, Microsoft's 9B model averages 67.1% across Echoverse's 12 worlds — within 14 points of GPT-5.4 at 80.7% — and passes GPT-5.4 outright on the EchoBank held-out set and the nested-filter environment. On WebVoyager and Online-Mind2Web the same weights transfer zero-shot, showing the gains are not benchmark-specific.

Who is Echoverse aimed at?

Teams building computer-use agents — open-source coders, enterprise RPA vendors, and academic RL groups who cannot afford GPT-5-tier per-token training. Echoverse hands them a reproducible synthetic curriculum plus a grader that verifies real database state after each step, so a small model can be pushed toward large-model behaviour without proprietary web logs.

Microsoft Research · 2026-07-30 · major

Microsoft Echoverse — deep, evolving environments train a 9B computer-use agent within 14 points of GPT-5.4

Echoverse is Microsoft Research's co-evolutionary training loop for computer-use agents. Synthetic environments, task graders, and the model improve together across 12 worlds and lift a 9B model from 36.5% to 67.1% average, within 14 points of GPT-5.4.

Microsoft Research Echoverse blog banner — evolving environments for computer-use agents

Echoverse's environments, graders and agent all improve together, taking a 9B computer-use policy from 36.5% to 67.1%.

Key specs

9 b agent average	36.5% → 67.1%
Training worlds	12
Gap to gpt 5.4	within 14 pts

Quick facts

Maker	Microsoft Research (AI Frontiers Lab)
Artifact	Training framework + 12 co-evolving environments
Code	github.com/microsoft/Echoverse (MIT)
Dataset	huggingface.co/datasets/microsoft/Echoverse (722 tasks, 6 envs)
Tech report	aka.ms/echoverse-techreport (PDF)
Baseline lift	9B model 36.5% → 67.1% average
Zero-shot transfer	WebVoyager 66.5 → 71.5, Online-Mind2Web 40.5 → 43.4

Benchmarks

Echoverse average (12 worlds)

9B base		36.5%
9B + Echoverse		67.1%
GPT-5.4		80.7%

source ↗

What is it?

Echoverse is Microsoft Research's co-evolutionary training framework for computer-use agents. Twelve synthetic worlds — email, datepicker, nested filters, customer support and two capability-focused environments among them — grow harder in step with the model, and the training code plus a Hugging Face dataset of 722 tasks ship under MIT.

How does it work?

Instead of freezing a benchmark, the environment, its task grader and the model are updated in a shared loop. Each step's outcome is checked against the underlying database, so reinforcement signal is grounded rather than judged by another LLM. The training worlds level up when the agent starts to solve them, keeping the curriculum ahead of the policy.

Why does it matter?

A 9B model trained inside Echoverse averages 67.1% across the 12 worlds, up from 36.5%, and lands within 14 points of GPT-5.4 while transferring zero-shot to WebVoyager (66.5→71.5) and Online-Mind2Web (40.5→43.4). Small open agents finally have a way to close the gap on proprietary flagships without buying browser logs.

Who is it for?

computer-use agent teams, RL researchers, RPA vendors

Frequently asked questions

How is Echoverse different from static computer-use benchmarks like WebArena?: Echoverse's environments EVOLVE alongside the agent — task graders, world dynamics, and difficulty curve up together, so a policy cannot memorise a fixed set. Microsoft reports zero-shot transfer to held-out real-web suites (WebVoyager 66.5→71.5, Online-Mind2Web 40.5→43.4), which is the signal that the co-evolutionary loop produces general skill, not overfit trajectories.
What can I actually download from Microsoft's Echoverse release?: Echoverse ships the training framework (github.com/microsoft/Echoverse, MIT-licensed) and a public Hugging Face dataset (microsoft/Echoverse) covering 722 tasks across six environment families — datepicker, nested-filter, email, customer support and two capability-focused worlds. The 22-page tech report at aka.ms/echoverse-techreport documents the training loop, EchoBank grader, and results.
How does an Echoverse-trained 9B compare to GPT-5.4 on computer-use tasks?: After co-evolutionary training, Microsoft's 9B model averages 67.1% across Echoverse's 12 worlds — within 14 points of GPT-5.4 at 80.7% — and passes GPT-5.4 outright on the EchoBank held-out set and the nested-filter environment. On WebVoyager and Online-Mind2Web the same weights transfer zero-shot, showing the gains are not benchmark-specific.
Who is Echoverse aimed at?: Teams building computer-use agents — open-source coders, enterprise RPA vendors, and academic RL groups who cannot afford GPT-5-tier per-token training. Echoverse hands them a reproducible synthetic curriculum plus a grader that verifies real database state after each step, so a small model can be pushed toward large-model behaviour without proprietary web logs.

Try it

git clone github.com/microsoft/Echoverse