Research map / Large language models / AI and society
GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks
Tejal Patwardhan, Rachel Dias, Elizabeth Proehl, Grace Kim, Michele Wang, Olivia Watkins, Simon Posada Fishman, Marwan Aljubeh, Phoebe Thacker, Laurance Fauconnet, Natalie S. Kim, Samuel Miserendino and 7 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase reasoning effort (low to high) for o3 and GPT-5 (ablation) | o3, GPT-5 · GDPval gold subset (220 tasks) | win rate vs human expert | Improved |
| Add tuned prompt for rigorous deliverable checking (ablation) | GPT-5 · GDPval gold subset (220 tasks) | win rate vs human expert: +5 percentage points | Improved |
| Add tuned prompt (ablation) | GPT-5 · GDPval gold subset (220 tasks) PowerPoint | egregious formatting error rate: 86% -> 64% | Improved |
| Add tuned prompt (ablation) | GPT-5 · GDPval gold subset (220 tasks) agents | multimodal self-inspection rate: 15% -> 97% | Improved |
| Add GET requests and best-of-4 sampling with GPT-5 judge (ablation) | GPT-5 · GDPval gold subset (220 tasks) | win rate vs human expert | Improved |
| Shorten prompts to 42% length (under-contextualized) (ablation) | GPT-5 · Under-contextualized GDPval | win rate vs human expert | Hurt |
| Expert try-once-then-fix workflow with GPT-4o | GPT-4o · GDPval gold subset (220 tasks) | speed improvement: 0.87x | Hurt |
| Expert try-once-then-fix workflow with GPT-5 | GPT-5 · GDPval gold subset (220 tasks) | speed improvement: 1.12x | Improved |
| Repeat sampling up to n times before expert fix with GPT-4o | GPT-4o · GDPval gold subset (220 tasks) | speed improvement: 0.87x -> 0.46x | Hurt |
| Repeat sampling n times with GPT-5 vs try once | GPT-5 · GDPval gold subset (220 tasks) | cost improvement: 1.18x -> 1.63x | Improved |
| Newer OpenAI model generation (GPT-4o to GPT-5) | OpenAI models · GDPval gold subset (220 tasks) | win rate vs human expert: 12.5% -> 39.0% | Improved |
Closest papers by meaning
- AGI-Elo: How Far Are We From Mastering A Task? (NeurIPS 2025 · similarity 0.86)
- BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows (NeurIPS 2026 · similarity 0.86)
- AI Evaluation Should Require Standardized Item-Level Data Releases (NeurIPS 2026 · similarity 0.85)
- RE-Bench: Evaluating Frontier AI R&D Capabilities of Language Model Agents against Human Experts (ICML 2025 · similarity 0.85)
- BetterBench: Assessing AI Benchmarks, Uncovering Issues, and Establishing Best Practices (NeurIPS 2024 · similarity 0.85)
- Unearthing Skill-level Insights for Understanding Trade-offs of Foundation Models (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
