Logical-CommonsenseQA - Mixed: leaderboard
Questions whose four options use different plausibility operators (Mixed). Logical-CommonsenseQA (Junias and Pacheco, University of Colorado Boulder, 2026) rewrites CommonsenseQA questions so that every option joins two atomic statements with a plausibility operator, and the model picks the one option whose composite plausibility fits. Models answer zero-shot (the authors' fine-tunes on the training split share this board); macro-F1 in percent over the 250 human-validated test questions, with a single-letter answer at temperature 0; higher is better.
Metric: Macro-F1 (%; operators mixed across the four options of a question; zero-shot prompt, or the authors' fine-tunes on the training split; 250 human-validated four-option test questions). Source: arxiv.org. Saturation forecast: Around October 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 61.8 | #141 |
| 2 | Gemini 2.5 Flash | 60.9 | #452 |
| 3 | Qwen 2.5 7B Instruct | 53.2 | #2583 |
| 4 | Llama 3.3 70B Instruct | 53 | #1353 |
| 5 | Llama 3.1 8B Instruct | 41.8 | #4277 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=logical-commonsenseqa-mixed · How It Works · Data refreshed daily, snapshot 2026-11-12.