CITBench: leaderboard
CITBench (2026) evaluates LLMs on tabular data processing tasks of matching, cleaning, augmentation and transformation. Scored in the offline single-turn setting as average performance: the mean Level-Accuracy over repeated runs, with partial credit, in percent; higher is better. This board averages the four offline difficulty tiers: base, complex rules, multiple input tables and multi-step pipelines.
Metric: Average performance (%; mean Level-Accuracy with partial credit, each of repeated runs scored 0 to 1, on the offline single-turn tabular data processing tasks of CITBench, a balanced 50% slice of its 1,296 instances in 18 matching, cleaning, augmentation and transformation task types; temperature 1.0, 8,192 output tokens; unweighted mean of the Base, Complex-Rule, Multi-Input and Multi-Step tiers). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 70.33 | #244 |
| 2 | Gemini 3 Flash | 68.5 | #159 |
| 3 | Gemini 3 Pro | 65.32 | #135 |
| 4 | GLM-4.7 | 62.14 | #365 |
| 5 | GPT-5.2 | 61.74 | #172 |
| 6 | Claude Haiku 4.5 | 54.64 | #568 |
| 7 | DeepSeek V3.2 | 53.51 | #403 |
| 8 | GPT-4o | 52.35 | #741 |
| 9 | Qwen 3 Coder Flash | 49.56 | #1624 |
| 10 | Qwen 3 235B A22B | 49.27 | #622 |
| 11 | Qwen 3 32B | 47.16 | #960 |
| 12 | Qwen 3 14B | 40.95 | #1235 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=citbench · How It Works · Data refreshed daily, snapshot 2026-11-12.