Mahal Systems

Overview › Benchmark catalog

ResearchRubrics

Fits any report-writing agent that needs rubric grading. Borrow the expert-written weighted rubrics (20-43 criteria per prompt, with negative-weight criteria) and the three-level satisfied / partial / not-satisfied verdict scheme with multiple judge models.

What it asks the agent to do

  • write a long evidence-backed answer to an open-ended research prompt
  • infer implicit context the prompt does not spell out
  • reason over retrieved material rather than paraphrase it

What to borrow from it

grading designcase templatetaxonomy

Caveats

Only 101 prompts; the paper says prompts, rubrics and code are released but the arXiv page gives no repository, only a Scale project page; license not stated.

Details

Workloadsresearch, web browsing, documents
Interactionautonomous long horizon
Environmentbrowser
End state checkedNo
Toolsweb search
Gradingllm judge, human
Size101
Datareal public
Personal datanone
Licenseunknown
Harnessnone
Statusactive

Links