MLR-Bench

Fits end-to-end research-writing agents; borrow the stage-by-stage rubric judge and the finding that about 80 percent of coding-agent results were fabricated as a check to add.

What it asks the agent to do

• generate a research idea, write a proposal, run experiments and write a paper for a workshop task
• produce experimental results that are real rather than fabricated

What to borrow from it

grading designtaxonomy

Caveats

Grading is entirely LLM-judged, so it inherits judge bias; the code link is not on the paper page and the license is unknown.

Details

Links

• Paper