Tencent Hunyuan and SSV Digital Culture Lab, in collaboration with the Institute of Computing Technology under the Chinese Academy of Sciences, released Chronicles-OCR on May 18, the first evaluation benchmark covering seven ancient font styles from oracle bone script to cursive script. The benchmark contains 2,800 expert-annotated images.
Testing of 28 mainstream multimodal large language models showed poor performance on ancient characters. GPT-5 and Gemini 2.5 Pro achieved near-zero scores on cross-era character detection, while the best-performing model reached only 16.5. Even with bounding boxes provided to skip localization, the highest accuracy was 27.1%, with Gemini 3.1 Pro achieving just 14.0% on oracle bone script.
Related News
Samsung and Intel team up to pressure, TSMC’s 18 plants launch the largest expansion plan in history! Plant maintenance materials stocks will benefit
Charms.ai completes a $1.5 million funding round to launch an AI character economy; Pennsylvania sues Character.ai for practicing medicine
Edge AI breakthrough: TetraMem releases results from its MLX200 platform built on chips based on TSMC’s 22nm chips, following development progress