UniDoc: Towards Unified Thinking in Documents
Organizations: Tsinghua University · University of Waterloo · The Hong Kong University of Science and Technology · Imperial College London
Abstract
Answering questions over long, visually rich documents demands both cross-page structural reasoning and fine-grained visual understanding. Existing approaches sacrifice one for the other: retrieval-augmented methods fragment documents and lose inter-page structure; monolithic vision-language models preserve structure but scale poorly beyond a few pages. This impasse stems from a false premise: that the macroscopic task of evidence retrieval and the microscopic task of detailed reasoning must either be split into separate systems or be processed identically. By decoupling their perceptual resolutions rather than their architectures, a single model can achieve the structural holism of monolithic processing alongside the efficiency of targeted retrieval. UniDoc realizes this through two resolution-adaptive stages: a Fast Reading stage processes all pages at low resolution to structurally identify evidence pages, and a Focused Thinking stage re-processes only those pages at high resolution for multimodal inference. To train this pipeline, we construct 30K multi-hop reasoning instances with stage-level supervision. On MMLongBench-Doc, UniDoc yields relative accuracy gains of 55.7% on Qwen2.5-VL-7B and 16.9% on Qwen3-VL-8B, while significantly reducing token cost relative to uniform high-resolution processing. Error analysis identifies page retrieval and fine-grained visual reasoning as the remaining bottlenecks, delineating concrete targets for future work.
Figures & tables
| Model | Size | Paradigm | DUDE | MMLong | SlideVQA | MP- | ||
| bench-doc | DocVQA † | |||||||
| ANLS | Acc. | F1 | Acc. | F1 | Acc. | |||
| Average Images Per Question | - | - | 5.7 | 47.5 | 20.0 | 8.3 | ||
| GPT-4o mini ( Hurst et al., 2024 ) | - | End-to-End | 46.5 | 28.6 | 29.4 | 48.1 | 60.7 | - |
| GPT-4o ( Hurst et al., 2024 ) | - | End-to-End | 52.7 | 41.6 | 42.3 | 53.1 | 65.8 | - |
| GPT-4.1 ( OpenAI, 2025a ) | - | End-to-End | 50.2 | 45.6 | 49.7 | 61.5 | 74.7 | - |
| Method | Acc. | Tokens/Doc | Wall-Clock (s/doc) |
| Full high-res (all pages) | 21.2 | 49.8k | 11.0 |
| ColPali-RAG ( Faysse et al., 2025 ) | 31.9 | 47.4k 6.1k | 5.3 |
| UniDoc-7B | 33.0 | 26.3k | 9.4 |
| Our Components | Benchmarks | ||||||
| Model Configuration | Two-Stage Framework | Data Engine (Synthesized) | RL Training (Direct RL) | DUDE (ANLS) | MMLong. (Acc.) | SlideVQA (Acc.) | MP-DocVQA (Acc.) |
| Base Model | 41.8 | 21.2 | 38.9 | 57.3 | |||
| Ours (Full Model) | 46.2 | 33.0 | 58.3 | 75.0 | |||
| w/o Two-Stage Framework | 40.5 | 22.1 | 47.5 | 63.9 | |||
| w/o Synthesized Data | 44.9 | 26.7 | 49.9 | 67.6 | |||
| w/o RL Training | 45.4 | 28.3 | 40.4 | 65.4 | |||
| Model | DUDE | MMLong. | SlideVQA | MP-DocVQA |
| (ANLS) | (Acc.) | (F1) | (Acc.) | |
| Avg. Images per Question | 5.7 | 47.5 | 20.0 | 8.3 |
| UniDoc-7B (Direct RL) | 46.2 | 33.0 | 67.9 | 75.0 |
| Unified SFT+RL | 47.2 | 28.6 | 68.7 | 78.7 |
| Delta | +1.0 | -4.4 | +0.8 | +3.7 |
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
| Parameter | Value |
| Per device train batch size | 1 |
| Gradient accumulation steps | 4 |
| Learning rate | 1.0e-5 |
| Number of epochs | 5.0 |
| LR scheduler | cosine |
| Warmup ratio | 0.1 |
| Parameter | Value |
| Per device train batch size | 2 |
| Gradient accumulation steps | 4 |
| Learning rate | 2.0e-6 |
| Number of epochs | 2.0 |
| LR scheduler | cosine |
| Warmup ratio | 0.1 |
| Parameter | Value |
| Per device train batch size | 1 |
| Gradient accumulation steps | 8 |
| Learning rate | 1.0e-5 |
| Number of epochs | 2.0 |
| LR scheduler | cosine |
| Warmup ratio | 0.1 |
| Stage | Source | Nums |
| Stage 1 | DUDE | 5148 |
| Stage 1 | SlideVQA | 3722 |
| Stage 1 | Synthesized from Data Engine | 2580 |
| Stage 2 | SlideVQA | 13414 |
| Stage 2 | Synthesized from Data Engine | 4321 |
| Parameter | Value |
| Train batch size | 64 |
| mini batch size | 32 |
| n resp per prompt | 8 |
| max prompt length | 14500 |
| max response length | 2500 |
| actor learning rate | 1.0e-6 |
| Method | MMLongBench-Doc | SlideVQA | DUDE | ||||||
| Acc. | Rec. | E2E | Acc. | Rec. | E2E | Acc. | Rec. | E2E | |
| SV-RAG | 17.23 | 46.56 | 23.0 | 23.59 | 67.56 | 32.0 | 27.24 | 76.55 | 45.0 |
| FRAG | 3.70 | 54.81 | 31.3 | 4.97 | 71.92 | 52.4 | 23.81 | 78.10 | 43.1 |
| UniDoc-7B | 44.30 | 60.86 | 33.0 | 64.54 | 75.31 | 58.3 | 53.70 | 79.30 | 46.2 |
| UniDoc-8B | 51.64 | 72.32 | 52.5 | 70.26 | 80.59 | 65.1 | 56.65 | 83.55 | 47.8 |
| Method | MMLongBench-Doc | SlideVQA | DUDE | |||
| Gold | E2E | Gold | E2E | Gold | E2E | |
| Qwen2.5-VL-7B | 56.3 | 28.3 | 84.3 | 40.4 | 76.8 | 45.4 |
| UniDoc-7B | 72.1 | 33.0 | 86.2 | 58.3 | 80.5 | 46.2 |
| Qwen3-VL-8B | 75.6 | 43.4 | 88.1 | 52.6 | 77.6 | 46.7 |
| UniDoc-8B | 86.5 | 52.5 | 90.5 | 65.1 | 81.7 | 47.8 |
| Configuration | SFT prior | Params | MMLong. | SlideVQA | DUDE |
| (1) Shared | no | 1 | 52.5 | 65.1 | 47.8 |
| (2) Decoupled | no | 2 | 48.1 | 59.4 | 48.6 |
| (3) Decoupled | yes | 2 | 60.4 | 74.8 | 66.7 |
| Setting | UniDoc-7B | UniDoc-8B | ||||
| Base | + low-res context | Base | + low-res context | |||
| Full evidence retrieved | 664 | 51.56 | 48.85 | 789 | 70.30 | 68.21 |
| Partial evidence retrieved | 427 | 4.24 | 24.66 | 302 | 6.03 | 36.97 |
| Overall | 1091 | 33.04 | 39.38 | 1091 | 52.51 | 59.56 |
| Backbone | Base model | Held-out | Trained |
| Qwen2.5-VL-7B | 57.3 | 75.0 | 81.3 |
| Qwen3-VL-8B | 72.8 | 81.4 | 89.6 |
Doc-V*:Coarse-to-Fine Interactive Visual Reasoning for Multi-Page Document VQA
Internalized Reasoning for Long-Context Visual Document Understanding
<think> tags, gated by a <cot> control token, and the resulting reasoning capability is internalized via low-strength model merging. We study Qwen3 VL 32B and Mistral Small 3.1 24B. With Qwen3 VL, we achieve 58.3 on MMLongBenchDoc, surpassing the 7 larger Qwen3 VL 235B A22B (57.0). With Mistral, we show that synthetic reasoning outperforms distillation from the Thinking version's traces by 3.8 points on MMLBD-C, and internalized reasoning exhibits 12.4 fewer mean output tokens compared to explicit reasoning. We release our pipeline for reproducibility and further exploration.