Targeted Document Editing on Curated Document Editing Benchmark
76.48ConsistencyLEDGER
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LEDGEREdge Types=CONTAINS + REFERENCES + DEPENDS + RELATED2026.06 | 76.48 | 76.68 | 77.1 | 78.06 | 1,500 | |
| Hierarchical RAGEdge Types=CONTAINS only (DOM-tree)2026.06 | 63.47 | 68.92 | 69.85 | 52.34 | 1.7 | |
| Claude + MemoryContext Type=Interaction history + semantically retrieved chunks2026.06 | 58.34 | 62.47 | 64.21 | 43.18 | 1.6 | |
| Full Doc. (Base.)Context Type=Entire document (5K–100K tokens)2026.06 | 56.59 | 64.31 | 67.13 | 41 | 1,822 | |
| RAG (Flat Sem.)Retrieval=Top-k chunks by cosine similarity2026.06 | 49.83 | 55.62 | 58.77 | 35.21 | 1.5 | |
| Sliding WindowContext Type=Fixed 4K token window around target2026.06 | 44.21 | 50.18 | 52.34 | 31.47 | 4 |