CITBench - Complex-Rule: leaderboard

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; Complex-Rule tier: one input table under tightly dependent rules). Source: arxiv.org. Saturation forecast: Around December 2026. 12 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.583.96
2Gemini 3 Flash81.18
3GLM-4.779.89
4Gemini 3 Pro74.11
5GPT-5.266.42
6GPT-4o65.99
7Qwen 3 Coder Flash65.91
8Qwen 3 235B A22B65.75
9DeepSeek V3.265.34
10Claude Haiku 4.560.56
11Qwen 3 32B58.46
12Qwen 3 14B53.14

Interactive version: theaggregate.ai/benchmark?slug=citbench-complex-rule · How It Works · Data refreshed daily, snapshot 2026-09-26.