DocPilot: Interactive Agent Orchestration and Training for Long Visual Document Understanding
- Published
- Source
- arXiv
- Paper number
- 1352
- Field
- AI Agents
- arXiv ID
- 2610.21327
Key points
- The researchers built DocPilot, a document agent framework that adaptively orchestrates six actions: page navigation, evidence extraction, local visual inspection, code-based reasoning, answer verification, and answering.
- With frontier VLM backbones, DocPilot improves MMLongBench-Doc accuracy by 10.1, 14.7, 10.8, and 13.4 points depending on the backbone.
- With curriculum-based training that moves from single-ability learning to orchestration learning, the 4B DocPilot-4B gains 4.2%, 7.5%, and 11.0% on the three benchmarks and competes with much larger models.
- With Gemini-2.5-Pro, DocPilot surpasses the prior document agent DocLens by 72.8 over 67.6, and on DocBench it exceeds DocDancer with 84.1 over 79.9.
- On the cross-document benchmark MADQA with GPT-5.4, DocPilot reaches 86.5 accuracy and records the best cross-document score of 89.5.
- Removing the answer verification action drops accuracy from 73.5 to 64.9, confirming that verification is central to the orchestration.
Paper links
External research summaries. These are not HDATF publications or measured product results.