Question Answering on CUAD (ANLS, ROUGE-L, METEOR)
0.2498ANLSStructure-based chunking
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Structure-based chunkingtop-k={1, 4}, Models=Llama-3.2-3B, Mistral-8B, Qwen-2.5-7B (averaged)2026.04 | 0.2498 | 15.56 | 15.91 | |
| Length chunkingtop-k={1, 4}, Models=Llama-3.2-3B, Mistral-8B, Qwen-2.5-7B (averaged)2026.04 | 0.2585 | 16.77 | 16.62 | |
| Semantic chunkingtop-k={1, 4}, Models=Llama-3.2-3B, Mistral-8B, Qwen-2.5-7B (averaged)2026.04 | 0.2593 | 14.91 | 14.68 | |
| Perplexity chunkingtop-k={1, 4}, Models=Llama-3.2-3B, Mistral-8B, Qwen-2.5-7B (averaged)2026.04 | 0.2641 | 16.46 | 15.24 | |
| LumberChunkertop-k={1, 4}, Models=Llama-3.2-3B, Mistral-8B, Qwen-2.5-7B (averaged)2026.04 | 0.2657 | 16.3 | 16.5 | |
| MultiDocFusiontop-k={1, 4}, Models=Llama-3.2-3B, Mistral-8B, Qwen-2.5-7B (averaged)2026.04 | 0.2738 | 17.62 | 16.5 | |
| Structure-based chunkingtop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 24.98 | 15.56 | 15.91 | |
| Length chunkingtop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 25.85 | 16.77 | 16.62 | |
| Semantic chunkingtop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 25.93 | 14.91 | 14.68 | |
| Perplexity chunkingtop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 26.41 | 16.46 | 15.24 | |
| LumberChunkertop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 26.57 | 16.3 | 16.5 | |
| MultiDocFusiontop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 27.38 | 17.62 | 16.5 | |
| M3DOCDEPtop-k={1, 2, 3, 4}, Readers=LLaVA-OneVision-1.5, InternVL-3.5, and Qwen2.5-VL, RAG protocol=shared corpus-level RAG protocol2026.04 | 29.25 | 19.78 | 18.23 |