CausalMind汇流知行
← All articles

GDPval brings evaluation closer to real work

Move from quiz answers to occupational tasks, deliverables, and expert review.

Move from quiz answers to occupational tasks, deliverables, and expert review.

PUBLIC RESEARCH / METHOD BASIS

Published results and documented interfaces, not CausalMind experiment results.

Evaluate deliverables

OpenAI introduced GDPval in 2025 with real economic work tasks across 44 occupations. Tasks ask for usable work products rather than quiz answers. Source completeness, format, business constraints, and product quality become part of evaluation.

Expert comparison matters

GDPval uses domain experts to compare deliverables. Its grading guidance says rubric-based model judgments can offer rough estimates, while nuanced professional quality still benefits from human review. Preserve reviewers' reasoning, not just a score.

Apply the design lesson

Start from a real role, a clear time point, available material, and an inspectable deliverable. Use rollouts to calibrate difficulty and rubrics. CausalMind's related task design is distinct from official GDPval data.

What the numbers represent

GDPval has 1,320 tasks in its full set, with 220 in the open gold subset, spanning nine sectors and 44 occupations. Tasks ask for concrete documents, spreadsheets, presentations, and other work products. Same-occupation experts compare model and human work blind to authorship and judge it better, tied, or worse. The public automated grader offers a quick estimate; the project does not use it to replace expert review.

Beyond a one-shot deliverable

The GDPval team notes that its first version is largely one-shot and does not capture multi-round feedback. A real report may need revision after a client changes constraints or new evidence appears. We would preserve a chain of draft, expert feedback, revision, and acceptance, scoring source use, factual accuracy, usable format, and uptake of feedback separately. Comparing the same draft with and without expert feedback would test whether the agent actually improves.

Primary sources

Editorial analysis by CausalMind; benchmark names refer to the cited original projects.