- Who it helps
- Researchers evaluating recovery in computer-use agents.
- Availability
- October 20 paper is public. The linked repository still says setup code will be released soon in the inspected snapshot.
Test the next actions
The authors inject errors, then replay goal-related actions. Recovery passes when the recovered state produces the same downstream outcomes as the target state. A similar-looking screen alone cannot establish this.
ReTrace tests whether an agent actually recovered from a wrong action ↗
Release status needs care
The abstract says code and data are available, but the linked repository’s setup section says the code is forthcoming. That mismatch prevents treating this as a ready-to-run benchmark.
ReTrace tests whether an agent actually recovered from a wrong action ↗ReTrace repository and release status ↗
What you can do
- Read the recovery definition and check the repository’s current setup status. Open reference ↗
Where the evidence stops
Paper and repository read November 3. No benchmark executed or recovery gain reproduced.
Sources read. Sources read November 3, 2026 UTC. No third-party software, model inference or full workflow executed.
Sources & further reading
Paper · paper
ReTrace tests whether an agent actually recovered from a wrong action ↗
Identifies the subject, documented capability or specific verification limit.
arxiv.org · page checked 2026-11-03
GitHub · implementation
ReTrace repository and release status ↗
Identifies the subject, documented capability or specific verification limit.
Byeonghun Bang and collaborators · page checked 2026-11-03
Coverage note: Independent demonstrations and complete workflow verification remain unavailable.
The record over time (2)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Source recheck
ReTrace evaluates whether a computer-use agent can repair an action error and continue the intended task.
- · Added to Current
ReTrace evaluates whether a computer-use agent can repair an action error and continue the intended task.