Microsoft unveils Echoverse to train computer-use AI

Microsoft Research has introduced Echoverse, a suite of twelve training environments—ten deep domain worlds and two capability worlds—designed to train computer-use agents. The ten domains are EchoMail, EchoCalendar, EchoChat, EchoML, EchoForge, EchoBank, EchoCare, EchoForum, EchoTunes, and EchoStay. Training a Qwen3.5-9B model on these worlds nearly doubled its base score from 36.5% to 67.1%, placing the 9B model within fourteen points of GPT-5.4, which scored 80.7%. Microsoft has released four of these worlds publicly.

Fidelity is crucial: on WebVoyager's Allrecipes, shallow training dropped performance from 80.0% to 75.0%, while deep training raised it to 85.0%. On Hugging Face, shallow training stayed at 48.0%, but deep training reached 65.0%, cutting step-budget exhaustion from 15 to 9 out of 37 tasks. Evaluating the model on live-web benchmarks showed improvements on WebVoyager from 66.5% to 71.5% and Online-Mind2Web from 40.5% to 43.4%; without BrowserBase, these scores shifted from 50.9% to 55.6% and 29.5% to 37.2% respectively. Adding EchoForge raised GitHub scores from 58.5% to 63.4%, WebVoyager from 50.9% to 52.9%, and Online-Mind2Web from 29.5% to 31.1%.

The capability worlds target date pickers (6 core widgets across 10 contexts, 10 held out) and nested filters (20 widget families, 9 held out). Datepicker training raised in-distribution scores from 60.0% to 82.6% and held-out layouts from 34.0% to 54.0%, while filter training raised held-out filters from 62.8% to 84.1%, also lifting Online-Mind2Web from 29.5% to 34.3%. Scaling trajectories from 6,400 to 20,000 on a fixed set of worlds kept WebVoyager flat at 54.8% to 55.6% and slipped Online-Mind2Web from 40.1% to 37.2%, showing that environment diversity, not volume, drives generalization.

Tasks are often five to twenty actions deep, and Echoverse uses database-grounded verification, accepting semantic matches like $288 for $287.62. EchoForum features 2.55 million comments, and EchoStay has 87 routes and 23 tables. System co-evolution resolved critical bugs: fixing EchoStay's guest-count control raised completed bookings from 48% to 78% (recovering 15 of 24 blocked runs). EchoForum fixes took 37 tasks from 0 solved to 36, and aligning EchoChat's verifier lifted gradable tasks from 34% to 99%.