DocFormBench (DocFormFlow) - Hallucinated Formatting: leaderboard

Metric: Hallucinated formatting rate (%) on DocFormBench: share of applied formatting changes that fall outside the intended scope (1 - precision), DocFormFlow workflow (localize targets, then call formatting tools; temperature 1.0, up to 3 verification and self-repair rounds; multimodal models also receive page images); reasoning disabled; lower is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 15 models tracked.

Top models

#ModelScore
1GLM-4.6 (Non-reasoning)1.24
2DeepSeek V4 Pro (Non-reasoning)1.57
3Gemini 3 Flash (Minimal)1.58
4DeepSeek V4 Flash (Non-reasoning)1.8
5Qwen 3 VL 32B Instruct1.94
6DeepSeek V3.2 (Non-reasoning)1.99
7Qwen 3 VL 235B A22B Instruct2.03
8GLM-4.6V (Non-reasoning)2.14
9Qwen 3 VL 8B Instruct2.45
10Qwen 3 235B A22B 2507 Instruct2.69
11Qwen 3.5 397B A17B (Non-reasoning)2.69
12Qwen 3.5 122B A10B (Non-reasoning)3.06
13GPT-5 (Minimal)3.44
14Qwen 3.5 35B A3B (Non-reasoning)3.45
15Qwen 3 VL 30B A3B Instruct3.8

Interactive version: theaggregate.ai/benchmark?slug=docformbench-docformflow-hallucinated-formatting · How It Works · Data refreshed daily, snapshot 2026-09-29.