BaFCo - Layout Analysis (CoT): leaderboard
BaFCo (2026) is a Bangla form-comprehension benchmark built from 200 multi-page Bangladeshi government forms. Layout analysis over 26 fine-grained entity types with a chain-of-thought prompt, scored by mean average precision at IoU 0.3 on a 0 to 100 scale; higher is better, each model at low and high reasoning effort.
Metric: Mean average precision at IoU 0.3 (0-100): class-wise average precision of predicted layout boxes, greedily matched to ground truth at IoU 0.3, averaged over the 26 granular entity classes, x100; 200 Bangladeshi government forms (316 pages) with 16,382 annotated entities; chain-of-thought prompt; reasoning_effort low or high as labelled; schema-invalid outputs excluded. Source: arxiv.org. Saturation forecast: Around 2029. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro (High) | 11.34 | #137 (Gemini 3 Pro) |
| 2 | Gemini 3 Pro (Low) | 8.53 | #137 (Gemini 3 Pro) |
| 3 | GPT-5.2 (Low) | 6.43 | #176 (GPT-5.2) |
| 4 | GPT-5.2 (High) | 6.15 | #176 (GPT-5.2) |
| 5 | Kimi K2.5 (High) | 4.76 | #256 (Kimi K2.5) |
| 6 | Qwen 3.6 Plus (High) | 1.66 | #222 (Qwen 3.6 Plus) |
| 7 | Claude Opus 4.6 (High) | 0.65 | #102 (Claude Opus 4.6) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bafco-layout-analysis-cot · How It Works · Data refreshed daily, snapshot 2026-11-13.