Reproduce “AlphaOracle: Oracle bone script decipherment via human-workflow-inspired deep learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yuliang Liu, Haisu Guan, Pengjie Wang, Xinyu Wang, Jinpeng Wan, Kaile Zhang, Handong Zheng, Xingchen Liu and 6 more
AlphaOracle analyzes oracle-bone rubbings by detecting characters, generating readings from glyph and historical forms, reranking them using inscriptional context, and retrieving related texts and scholarship. For the disputed character 勞, this evidence supports a place-name reading. It selects contextually appropriate candidates with 95.2% accuracy versus 65.4% for Gemini 2.5 Pro, but does not settle rare or damaged characters.
Key insights
- Context is especially valuable after visual analysis has narrowed the candidate set: contextual candidate re-ranking reached 95.2%, far above the reported language-model baseline of 65.4%.
- Decipherment benefits from aggregating repeated occurrences of a glyph and linking them to diachronic variants, formulaic contexts, and philological references instead of treating each rubbing as an isolated image.
- Human evaluation supports a complementary role for AI: 78.7% of 86 experts found the system helpful or very helpful, while the authors emphasize that unresolved cases still require scholarly adjudication.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.