ReasonTabQA: leaderboard
ReasonTabQA (TeleAI (China Telecom), 2026, with Chongqing University and Beihang University) poses questions over industrial tables in Chinese and English from 30 industry sub-domains, including multi-table and multi-sheet directories, nested headers and very large tables. Each model writes a Python program that reads the table files with pandas and prints an answer, which an LLM judge compares with the annotated gold answer; accuracy in percent, higher is better. The authors' TabCodeRL models are fine-tuned on the benchmark's own training split.
Metric: Answer Accuracy (%; LLM judge against gold answers, program-aided). Source: arxiv.org. Saturation forecast: Around February 2027. 33 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (Preview) | 67.58 | #111 |
| 2 | Claude Opus 4.5 (Thinking) | 66.2 | #152 (Claude Opus 4.5) |
| 3 | Doubao-1.5-Thinking-Pro | 64.48 | #236 |
| 4 | Claude Sonnet 4 (Thinking) | 62.53 | #397 (Claude Sonnet 4) |
| 5 | Kimi K2 | 59.57 | #435 |
| 6 | GPT-5.2 (Non-reasoning) | 59.46 | #172 (GPT-5.2) |
| 7 | Qwen 3 32B (Thinking) | 58.76 | #960 (Qwen 3 32B) |
| 8 | GPT-4o | 56.9 | #741 |
| 9 | DeepSeek R1 | 55.8 | #467 |
| 10 | O1 Mini | 55.8 | #761 |
| 11 | QwQ-32B | 54.1 | #909 |
| 12 | DeepSeek R1 Distill Qwen 32B | 53.64 | #1780 |
| 13 | Qwen 3 30B A3B (Thinking) | 53.61 | #1124 (Qwen 3 30B A3B) |
| 14 | Qwen 3 32B (Non-reasoning) | 53.13 | #960 (Qwen 3 32B) |
| 15 | DeepSeek V3 | 52.15 | #705 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=reasontabqa · How It Works · Data refreshed daily, snapshot 2026-11-12.