Research map / Large language models / Language, safety and interpretability
LLMCO4MR: LLMs-aided Neural Combinatorial Optimization for Ancient Manuscript Restoration from Fragments with Case Studies on Dunhuang
Yuqing Zhang, Hangqi Li, Shengyu Zhang, Runzhong Wang, Baoyi He, Huaiyong Dou, Junchi Yan, Yongquan Zhang, Fei Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add OT layer (cardinality constraint) to GNN fragment selection instead of greedy top-K (ablation) | CO solver · Synthetic 2x2 selection (pool 15) | selection accuracy: 0.3125 -> 0.5825 | Improved |
| add OT layer to GNN selection (ablation) | CO solver · Synthetic 3x3 selection (pool 30) | selection accuracy: 0.5389 -> 0.6044 | Improved |
| add second-order matching strategy to Top-9 selection (ablation) | CO solver · Synthetic Top-9 of 30 | selection accuracy: 0.6044 -> 0.6633 | Improved |
| add second-order matching strategy (ablation) | CO solver · Synthetic Top-9 of 20 | selection accuracy: 0.7256 -> 0.7311 | No clear effect |
| add MLLM (GPT-4V) pairwise matching after CO solver selection (ablation) | LLMCO4MR · Dunhuang case study (2x2, pool 10) | success rate: 0.5750 -> 0.6750 | Improved |
| add MLLM matching on top of CO solver for larger 3x3 pool (ablation) | LLMCO4MR · Dunhuang case study (3x3, pool 25) | success rate: 0.4667 -> 0.5556 | Improved |
| add CO solver outlier filtering before GPT-4V matching (ablation) | GPT-4V · Dunhuang case study (2x2, pool 15) | success rate: 0.1750 -> 0.6250 | Improved |
| increase negative pairs in siamese training (ratio 1:2 vs 1:1) (ablation) | Siamese network · Top-4 performance, 120K samples | Top-4 performance: 0.7525 -> 0.7800 | Improved |
| increase siamese training data from 40K to 120K (ratio 1:2) (ablation) | Siamese network · Top-4 performance | Top-4 performance: 0.7100 -> 0.7800 | Improved |
| use full LLMCO4MR instead of prior restoration methods | LLMCO4MR · Dunhuang case study (3x3, pool 25) | success rate: 0.4111 -> 0.5556 | Improved |
Closest papers by meaning
- 3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping (AAAI 2025 · similarity 0.84)
- EpiAgent: An Agent-Centric System for Ancient Inscription Restoration (CVPR 2026 · similarity 0.84)
- Towards Automated Chinese Ancient Character Restoration: A Diffusion-Based Method with a New Dataset (AAAI 2024 · similarity 0.84)
- OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? (ICLR 2025 · similarity 0.82)
- Re-assembling the past: The RePAIR dataset and benchmark for real world 2D and 3D puzzle solving (NeurIPS 2024 · similarity 0.82)
- VaseVQA-3D: Benchmarking 3D VLMs on Ancient Greek Pottery (ICLR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
