SDABench - Mechanistic: leaderboard
SDABench (2026) mechanistic tasks, which explain the process behind an observed outcome from data and domain theory, on held-out synthetic scientific datasets from five domains, answered in open-ended form and checked against programmatically derived answers. The score is accuracy in percent; higher is better.
Metric: Accuracy (%) of open-ended answers on the held-out SDA-Synth test split (the 400 mechanistic analysis instances; zero-shot, JSON reasoning and answer fields, the answer scored by relative-error numeric, exact or structured matching with a GPT-4o fallback judge); higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 70 | #256 |
| 2 | Qwen 3.5 397B A17B | 48 | #246 |
| 3 | GPT-5.4 | 47 | #123 |
| 4 | Claude Sonnet 4.6 | 45 | #161 |
| 5 | DeepSeek V3.2 | 41 | #406 |
| 6 | Gemini 3.1 Pro (Preview) | 35.75 | #94 |
| 7 | GLM-5 | 30.25 | #260 |
| 8 | DeepSeek R1 | 23.25 | #473 |
| 9 | Qwen 3 235B A22B 2507 Instruct | 22.5 | #619 |
| 10 | GPT-5.4 Mini | 18.75 | #389 |
| 11 | Llama 3.3 70B Instruct | 2.75 | #1357 |
| 12 | Llama 3.1 8B Instruct | 0 | #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-mechanistic · How It Works · Data refreshed daily, snapshot 2026-11-13.