SCOPE Experimental Design (CoT + Search): leaderboard
SCOPE (2026) asks LLMs to design the experiments of recent ICML, NeurIPS and ICLR papers from their task description, related work and method, and a GPT-5.2 judge scores the plan against the real design with a redline rule that zeros fatal flaws. Total score out of 30 over six sub-dimensions, higher is better, with chain-of-thought prompting plus web search the model may call.
Metric: Total plan score (out of 30): sum of six sub-dimension rubric scores from a GPT-5.2 judge against the ground-truth experimental design of 300 recent ICML, NeurIPS and ICLR papers in 19 research domains; a redline rule zeros a sub-dimension with a fatal flaw such as a hallucinated dataset or baseline, an incompatible metric or a violated constraint; chain-of-thought prompting with web search the model is told to call when needed. Source: arxiv.org. Saturation forecast: Around July 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 18.38 | #244 |
| 2 | GPT-5.2 | 16.77 | #172 |
| 3 | Kimi K2 | 15.6 | #435 |
| 4 | DeepSeek V3.2 | 13.76 | #403 |
| 5 | Grok 4 | 13.12 | #306 |
| 6 | Qwen 3 Max | 12.88 | #401 |
| 7 | Gemini 3 Pro | 12.66 | #135 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=scope-experimental-design-cot-plus-search · How It Works · Data refreshed daily, snapshot 2026-11-12.