AI EVALUATION, IN CONTEXT.BY EQUALITY MACHINE
Benchmark Radar.Benchmarks. Model results. Original sources.
/

Computer use benchmarks

Read evaluations of GUI grounding, browser actions and complete desktop or mobile workflows.

Computer use benchmarks ask a system to interact with a graphical interface. Some test one decision, such as locating the button named in an instruction. Others require a sequence of actions that changes an application's state. ScreenSpot-style grounding and OSWorld-style desktop workflows therefore answer different questions. A terminal-only coding agent belongs in Tools & Agents unless its task also requires graphical interaction.

What these benchmarks test

  • Grounding connects a textual instruction to a visible interface element or screen location.
  • Workflow evaluations test navigation, recovery from mistakes and completion across one or more applications.
  • Mobile, browser and desktop environments expose different controls, permissions and state.

How to read the results

  • Identify what the agent can observe: screenshots, an accessibility tree, a DOM, or a combination.
  • Match application versions, screen resolution, action limits and starting state before comparing results.
  • Prefer an explicit success check on the final state; distinguish it from matching a reference click sequence.

What to watch for

  • Accurate clicking alone does not demonstrate the ability to complete a long task.
  • A successful sandbox workflow does not establish reliability with live accounts, changing websites or irreversible actions.

References

  1. OSWorld and OSWorld Verified
  2. SeeClick and ScreenSpot
  3. AndroidWorld

Benchmarks in this topic 18 source records

Search and filter this topic

Related topics

Tools and agents benchmarksMultimodal benchmarksSafety and reliability benchmarks