Chronicles-OCR - Archaic Character Spotting: leaderboard

Chronicles-OCR (2026) tests vision-language models on historical Chinese writing across the Seven Chinese Scripts, 2,800 expert-annotated images from Oracle Bone to Cursive. This board scores end-to-end spotting on the Oracle Bone, Bronze and Seal scripts: the model must box every archaic character and give its modern equivalent, scored as text-spotting H-mean (0 to 100) averaged over the three scripts; higher is better.

Metric: Text-spotting H-mean (0-100) of predicted character boxes with their modern character mapping (a hit needs IoU above 0.75 and an exact character match, undeciphered symbols excluded) on the archaic scripts (Oracle Bone, Bronze and Seal), the paper Average column over the per-script results; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 26 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Seed 2.0 ProNon-reasoning16.5#222
2Qwen 3.5 397B A17BNon-reasoning9.7#264
3Seed 1.8Non-reasoning9.2#252
4Qwen 3 VL 235B A22B Instruct7.8#532
5Qwen 3.5 35B A3BNon-reasoning5.6#492
6Kimi K2.5Non-reasoning5#266
7Qwen 3 VL 8B Instruct3.4#1029
8Gemini 3.1 Pro (Preview)2.6#99
9Gemma 4 31B (IT)2.3#350
10Qwen 3 VL 235B A22B (Thinking)2.1#459
11Qwen 3 VL 2B Instruct2.1#2367
12Ovis2.6-30B-A3BThinking1.9
13GLM-4.5V1.4#754
14Qwen 3 VL 8B (Thinking)1#910
15Gemini 2.5 Pro0.8#276

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=chronicles-ocr-archaic-character-spotting · How It Works · Data refreshed daily, snapshot 2026-11-21.