SDABench - Descriptive (MCQ): leaderboard
SDABench (2026) descriptive tasks, which summarize measurements to characterize a phenomenon, on held-out synthetic scientific datasets from five domains, posed as four-option multiple choice with perturbation-based distractors. The score is accuracy in percent, with 25 percent at chance; higher is better.
Metric: Accuracy (%) of four-option multiple-choice answers with perturbation-based distractors on the held-out SDA-Synth test split (the 400 descriptive analysis instances; zero-shot; 25% chance); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 94.25 | #94 |
| 2 | Claude Sonnet 4.6 | 93.75 | #161 |
| 3 | GPT-5.4 | 93.5 | #123 |
| 4 | Kimi K2.5 | 92.25 | #256 |
| 5 | Qwen 3.5 397B A17B | 92 | #246 |
| 6 | GLM-5 | 91.5 | #260 |
| 7 | DeepSeek V3.2 | 90 | #406 |
| 8 | GPT-5.4 Mini | 85.75 | #389 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 85.5 | #619 |
| 10 | DeepSeek R1 | 84.25 | #473 |
| 11 | Llama 3.3 70B Instruct | 28 | #1357 |
| 12 | Llama 3.1 8B Instruct | 21.5 | #4246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sdabench-descriptive-mcq · How It Works · Data refreshed daily, snapshot 2026-11-13.