TIDE-Bench - ExpoDesign: leaderboard
The ExpoDesign track of TIDE-Bench (2026), a diagnostic benchmark for tool-integrated reasoning: from a research brief that hides the source paper, the model uses search and Python tools to build a structured experiment card, which a Kimi K2 judge scores from 1 to 5 for experimental design, tool success and consistency of results and conclusions. Mean of the three scores (1-5); higher is better.
Metric: Overall judge score (1-5): mean of the experimental-design, tool-success and result-conclusion consistency scores (each 1-5) a kimi-k2-0905-preview judge gives the structured experiment card the system builds, with search and Python tools, from a paper-agnostic research brief (TIDE-Bench ExpoDesign track, STEM and computer-science papers); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 4.79 | #135 |
| 2 | GPT-5 | 4.68 | #178 |
| 3 | Qwen 3 32B | 3.03 | #972 |
| 4 | Qwen 3 8B | 2.98 | #1873 |
| 5 | Qwen 3 4B | 2.83 | #2348 |
| 6 | Tool-Star | 2.53 | |
| 7 | GPT-4o | 2.52 | #751 |
| 8 | SearchR1 | 1.77 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=tide-bench-expodesign · How It Works · Data refreshed daily, snapshot 2026-11-21.