LogicSkills - Countermodel Construction: leaderboard
Metric: Accuracy (fraction times 100) on 300 language-neutral countermodel items: for an invalid argument the model gives a finite structure on a fixed small domain that makes every premise true and the conclusion false, checked by Z3; greedy decoding, responses normalized by a GPT-4o extractor and checked with the Z3 solver; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 | 99 | #121 |
| 2 | Qwen 3 32B (Thinking) | 89 | #424 (Qwen 3 32B) |
| 3 | Claude 3.7 Sonnet | 42 | #241 |
| 4 | Qwen2.5-Math-72B-Instruct | 15 | #708 |
| 5 | Phi-4 | 14 | #701 |
| 6 | Gemini 2.5 Flash | 13 | #237 |
| 7 | Llama 3.1 70B Instruct | 12 | #548 |
| 8 | GPT-4o | 10 | #333 |
| 9 | Llama 3.1 8B Instruct | 0 | #1018 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=logicskills-countermodel-construction · How It Works · Data refreshed daily, snapshot 2026-10-11.