Document Question Answering on DocVQA (val)
93.72ANLSGPT-4.1 vision (direct)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| GPT-4.1 vision (direct)Notes=Upper bound; reads images directly2026.04 | 93.72 | 95 | 0 | |
| RaV-IDP fullQA answering model=GPT-4.1-mini, Notes=Best open-source pipeline2026.04 | 42.24 | 44.7 | 0.7 | |
| RaV-IDP no_ravQA answering model=GPT-4.1-mini, Notes=Without RaV components2026.04 | 42.06 | 44.3 | 0.7 | |
| UnstructuredQA answering model=GPT-4.1-mini, Notes=+8% vs. RaV full2026.04 | 39.1 | 41 | 0 | |
| DoclingQA answering model=GPT-4.1-mini2026.04 | 38.44 | 40.7 | 2.7 | |
| MarkerQA answering model=GPT-4.1-mini2026.04 | 36.19 | 38.7 | 0 | |
| LlamaParseQA answering model=GPT-4.1-mini, Notes=Cloud timeouts hurt score2026.04 | 26.74 | 28.3 | 30 | |
| RaV-IDP gate_onlyQA answering model=GPT-4.1-mini, Notes=Exclusion-only; collapses coverage2026.04 | 14.08 | 14.3 | 29.7 |