InterveneBench: leaderboard
Metric: Final score (times 100): the 45-point rubric total normalized to 0-1 (model type, core independent variable and group definition 10 points each, control and dependent variables 5 each, reasoning 2, explanation 3), on InterveneBench's 74-study 2025 test set (each instance an empirical social-science study of a real policy intervention; the model proposes the causal study design from the policy context, without causal graphs); a Gemini-3-Pro grader compares the design with the expert-verified ground truth; direct inference; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.1 | 57.8 | #131 |
| 2 | Claude 3.7 Sonnet | 54.4 | #241 |
| 3 | Claude Sonnet 4 | 53.9 | #194 |
| 4 | GLM-4.6 | 53.1 | #246 |
| 5 | Gemini 2.5 Pro | 49.7 | #145 |
| 6 | Qwen 3 235B A22B | 48.2 | #304 |
| 7 | Gemini 3 Flash | 46.6 | #93 |
| 8 | Grok 4 | 46.4 | #169 |
| 9 | GPT-4.1 | 45.6 | #240 |
| 10 | GPT-OSS-120B | 44.3 | #330 |
| 11 | DeepSeek V3.2 | 42.5 | #198 |
| 12 | Kimi K2 | 41.5 | #236 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=intervenebench · How It Works · Data refreshed daily, snapshot 2026-10-11.