SDABench - Mechanistic (MCQ): leaderboard

Metric: Accuracy (%) of four-option multiple-choice answers with perturbation-based distractors on the held-out SDA-Synth test split (the 400 mechanistic analysis instances; zero-shot; 25% chance); higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 13 models tracked.

Top models

#ModelScore
1GPT-5.483
2Claude Sonnet 4.680.5
3Kimi K2.580.5
4Gemini 3.1 Pro (Preview)79.25
5Qwen 3.5 397B A17B77
6GLM-577
7DeepSeek V3.276.75
8GPT-5.4 Mini72
9DeepSeek R167.75
10Qwen 3 235B A22B 2507 Instruct46
11Llama 3.3 70B Instruct39.5
12Llama 3.1 8B Instruct34

Interactive version: theaggregate.ai/benchmark?slug=sdabench-mechanistic-mcq · How It Works · Data refreshed daily, snapshot 2026-09-29.