Multi-page Document Understanding on DUDE
64.5ANLSDoc-V* (GRPO)
Evaluation Results
| Method | Links | |
|---|---|---|
| Doc-V* (GRPO)Backbone=Qwen2.5-VL, Param=7B, Paradigm=Agent2026.04 | 64.5 | |
| Claude-3.7-SonnetBackbone=-, Param=-, Paradigm=E2E2026.04 | 58.1 | |
| Doc-V* (SFT)Backbone=Qwen2.5-VL, Param=7B, Paradigm=Agent2026.04 | 58.1 | |
| URaGBackbone=Qwen2.5-VL, Param=7B, Paradigm=RAG2026.04 | 57.6 | |
| GPT-4oBackbone=-, Param=-, Paradigm=E2E2026.04 | 54.1 | |
| CREAMBackbone=Pix2Struct / LLaMa2, Param=7B, Paradigm=RAG2026.04 | 52.5 | |
| Qwen2.5-VL (RAG Top-5)Backbone=Qwen2.5-VL, Param=7B, Paradigm=RAG2026.04 | 52.2 | |
| Qwen2.5-VL (Baseline)Backbone=Qwen2.5-VL, Param=7B, Paradigm=E2E2026.04 | 51.9 | |
| GPT-4.1Backbone=-, Param=-, Paradigm=E2E2026.04 | 50.2 | |
| DocVLMBackbone=Qwen2-VL, Param=7B, Paradigm=E2E2026.04 | 47.4 | |
| InternVL3Backbone=InternViT / Qwen2.5, Param=8B, Paradigm=E2E2026.04 | 47.4 | |
| mPLUG-DocOwl2Backbone=ViT / LLaMa, Param=8B, Paradigm=E2E2026.04 | 46.8 | |
| GPT-4o miniBackbone=-, Param=-, Paradigm=E2E2026.04 | 46.5 | |
| CogDocBackbone=Qwen2.5-VL, Param=7B, Paradigm=Agent2026.04 | 46.2 | |
| Gemini-1.5-ProBackbone=-, Param=-, Paradigm=E2E2026.04 | 46 | |
| SV-RAGBackbone=InternVL2, Param=4B, Paradigm=RAG2026.04 | 45 | |
| VDocRAGBackbone=Phi3-Vision, Param=4B, Paradigm=RAG2026.04 | 44 | |
| VisRAGBackbone=MiniCPM-V 2.6, Param=8B, Paradigm=RAG2026.04 | 43.1 | |
| DocopilotBackbone=InternVL2, Param=8B, Paradigm=E2E2026.04 | 40.7 | |
| M3DocRAGBackbone=Qwen2-VL, Param=7B, Paradigm=RAG2026.04 | 39.5 | |
| HiVT5 (PR)Backbone=DiT / T5, Param=0.3B, Paradigm=E2E2026.04 | 23.1 |