ReasonTabQA - Chinese: leaderboard
ReasonTabQA accuracy on its Chinese-language industrial tables: the model writes pandas code that reads the table files (multi-sheet directories, nested headers, very large tables) and prints an answer, which an LLM judge compares with the gold answer; percent, higher is better. TeleAI (China Telecom), 2026.
Metric: Answer Accuracy (%; LLM judge against gold answers, program-aided). Source: arxiv.org. Saturation forecast: Estimated already saturated. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Doubao-1.5-Thinking-Pro | 68.06 | #236 |
| 2 | Gemini 3 Pro (Preview) | 66.96 | #111 |
| 3 | Claude Opus 4.5 (Thinking) | 65.74 | #152 (Claude Opus 4.5) |
| 4 | Claude Sonnet 4 (Thinking) | 64.35 | #397 (Claude Sonnet 4) |
| 5 | Kimi K2 | 63.43 | #435 |
| 6 | O1 Mini | 61.57 | #761 |
| 7 | Qwen 3 32B (Thinking) | 60.19 | #960 (Qwen 3 32B) |
| 8 | GPT-4o | 58.8 | #741 |
| 9 | GPT-5.2 (Non-reasoning) | 58.14 | #172 (GPT-5.2) |
| 10 | DeepSeek R1 | 56.3 | #467 |
| 11 | Qwen 3 32B (Non-reasoning) | 56.07 | #960 (Qwen 3 32B) |
| 12 | Qwen 3 30B A3B (Thinking) | 54.77 | #1124 (Qwen 3 30B A3B) |
| 13 | Qwen 3 8B (Thinking) | 54.63 | #1838 (Qwen 3 8B) |
| 14 | QwQ-32B | 53.11 | #909 |
| 15 | DeepSeek R1 Distill Qwen 32B | 52.78 | #1780 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=reasontabqa-chinese · How It Works · Data refreshed daily, snapshot 2026-11-12.