Mahal Systems

Overview › Benchmark catalog

xbench-DeepSearch

Fits Chinese-language search assistants that need a refreshed test set. Borrow the encrypted release with canary strings and the plan to re-issue a new 100-question set periodically.

What it asks the agent to do

  • answer a Chinese open-domain question that needs several search steps
  • follow the expected reference steps to locate the fact

What to borrow from it

baselineprivacy pattern

Caveats

Chinese only; 100 questions per release and old releases are superseded; the arXiv paper covers the profession tracks, the DeepSearch details live on the dataset card and site; default judge is gemini-2.0-flash.

Details

Workloadsweb browsing, knowledge qa, research
Interactionautonomous long horizon
Environmentbrowser
End state checkedNo
Toolsweb search
Gradingllm judge
Size100
Datareal public
Personal datanone
LicenseMIT
Harnessown repo
Statusactive

Links