ReceiptBench - Reasoning: leaderboard

Semantic reasoning fields (expense type, place, departure and arrival, ISO currency, merchant city) of ReceiptBench (2026), structured extraction of 19 fields from photos and scans of real receipts and invoices (purchase, dining, travel, hotel, taxi and more), 2,000 held-out test images, JSON output. Field-level F1 times 100 (0-100), exact, numeric or list matching with a small LLM judge for near misses; higher is better.

Metric: Semantic reasoning fields (expense type, place, departure, arrival, ISO currency, merchant city), field-level F1 (times 100, so 0-100) on the 2,000-image held-out test set of ReceiptBench (real travel and expense receipts and invoices, 19 annotated fields), one-shot schema prompt with JSON output; fields are scored by exact, numeric or structured-list matching with a Qwen3-4B semantic judge when string similarity falls below the match threshold; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL Plus87.87#514
2Gemini 3 Pro87.14#146
3GPT-587.06#185
4Qwen 3 VL 32B86.9#624
5PaddleOCR-VL80.67
6InternVL3.5-241B80.24#632
7Qwen 3 VL 8B79.36#927
8InternVL3-78B78.74#765
9olmOCR-7B78.54
10Qwen 3 VL 4B75.88#1163
11InternVL3-8B73.71#1795
12DianJin-OCR-R159.92
13DeepSeek-OCR-small55.17
14InternVL3-2B46.37#3318

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=receiptbench-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-22.