cs-ai/agents IB-arxiv-2607.28074-shallow-env-live-site-drop measured

Training a computer-use agent on shallow Echoverse environments reduces live-site accuracy below the base model, from 80.0% to 75.0%.

Active Evidence state is recorded explicitly and may be revised by a later event.

Observed quantities

live_site_accuracy_before_pct 80.0 · live_site_accuracy_after_pct 75.0

01

Primary source

  1. arxiv · 2607.28074 · v1 · abstract
Title
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
and 1 moreAkshay Nambi
Version
v1
Check
Support quote found word for word in the arXiv abstract v1, 2026-10-08
Supporting quote from the source
shallow environments push live-site accuracy below the base model ($80.0 \to 75.0$)

02

Boundary of this block

No additional public limitation was recorded.

03

Connections

Outgoing

No outgoing public edges.

Incoming

No incoming public edges.