SDABench (MCQ): leaderboard
Metric: Accuracy (%) of four-option multiple-choice answers with perturbation-based distractors on the held-out SDA-Synth test split (2,400 instances, macro-averaged over the six task types; zero-shot; 25% chance); higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 13 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 69.46 |
| 2 | GPT-5.4 | 66.58 |
| 3 | Claude Sonnet 4.6 | 66.38 |
| 4 | Qwen 3.5 397B A17B | 62.46 |
| 5 | DeepSeek V3.2 | 62.21 |
| 6 | GLM-5 | 60.79 |
| 7 | GPT-5.4 Mini | 58.46 |
| 8 | Kimi K2.5 | 57.04 |
| 9 | DeepSeek R1 | 56.38 |
| 10 | Qwen 3 235B A22B 2507 Instruct | 52.42 |
| 11 | Llama 3.3 70B Instruct | 30.46 |
| 12 | Llama 3.1 8B Instruct | 25 |
Interactive version: theaggregate.ai/benchmark?slug=sdabench-mcq · How It Works · Data refreshed daily, snapshot 2026-09-29.