T2S-Bench - Counterfactual Reasoning: leaderboard
Metric: Exact match (%) on the 107 Counterfactual Reasoning questions on T2S-Bench-MR (500 expert-checked multiple-choice questions over text-structure pairs from scientific papers in six domains; single- and multiple-answer items, only the text and the question given), deterministic decoding at temperature 0; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 45 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 90.65 | #145 |
| 2 | Qwen 3 32B | 89.72 | #424 |
| 3 | Claude Sonnet 4 (20250514) | 89.72 | #211 |
| 4 | Claude Sonnet 4.5 | 88.79 | #138 |
| 5 | Claude Haiku 4.5 (20251001) | 85.98 | #251 |
| 6 | GPT-5.2 | 84.11 | #105 |
| 7 | Kimi K2 0905 | 84.11 | #282 |
| 8 | GPT-5.1 | 83.18 | #131 |
| 9 | Mistral Small 3.2 | 80.37 | #587 |
| 10 | GPT-4o | 79.44 | #333 |
| 11 | Gemini 2.0 Flash Lite | 79.44 | #438 |
| 12 | Gemini 2.5 Flash | 78.5 | #237 |
| 13 | DeepSeek V3.2 | 78.5 | #198 |
| 14 | Qwen 3 235B A22B 2507 Instruct | 78.5 | #291 |
| 15 | DeepSeek V3.1 | 78.5 | #260 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=t2s-bench-counterfactual-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-11.