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
Figure 1: Existing paradigms face a fundamental scalability–fidelity trade-off. UniDoc addresses this by jointly optimizing evidence retrieval and multimodal reasoning within a single model through two resolution-adaptive stages: Fast Reading and Focused Thinking.
Figure 2: Overview of our methodology. The Data Synthesis Engine produces complex queries and stage-level annotations from raw documents. UniDoc processes them through Fast Reading and Focused Thinking stages, optimized end-to-end via Reinforcement Learning with stage-aware rewards.
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
-
Table 1: Quantitative experimental results. Size refers to the number of parameters. Paradigm refers to the specific method to derive the answer. Reported metrics include ANLS, Accuracy, and F1 score. Best performances among open-weight models are in bold , second best are underlined . MP-DocVQA is evaluated as a held-out benchmark and marked with † : its training set is excluded from our training data, while all other columns are trained on. Bolding and underlining are therefore applied only to the trained columns. Results with MP-DocVQA included in training are reported in Appendix E.5 . Dash (-) indicates the metric was not originally evaluated for that dataset.
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
Table 2: Comparison of accuracy and computational cost on MMLongBench-Doc. Wall-clock time includes all preprocessing and inference stages.
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
Table 3: Ablation study of our core components. We quantify the impact of each component by comparing UniDoc-7B against the Base Model and three ablated variants, each removing one of our core contributions.
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
Table 4: Performance comparison of the final policy. The Unified SFT+RL exhibits a trade-off: marginal gains on short-context tasks but degradation on long-context MMLongBench-Doc.
Figure 7Figure 8
Figure 6: Error taxonomy on MMLongBench-Doc, comparing (a) the base model and (b) our trained model. Each error is attributed to a stage (Fast Reading or Focused Thinking) and a document characteristic.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Data Engine query distribution.
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
Appendix
Table 6: Hyperparameters for supervised fine-tuning in Stage 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
Appendix
Table 7: Hyperparameters for supervised fine-tuning in Stage 2 .
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
Appendix
Table 8: Hyperparameters for supervised fine-tuning in mixed training.
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
Appendix
Table 9: Query Sources in SFT Training
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
Appendix
Table 10: Hyperparameters for reinforcement learning.
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
Appendix
Table 11: Stage 1 retrieval performance across three benchmarks, with end-to-end accuracy from Table 1 for reference.
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
Appendix
Table 12: Accuracy given gold evidence pages, with end-to-end accuracy under the same two-stage scaffold for reference. DUDE is reported in ANLS.
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
Appendix
Table 13: Shared parameters versus decoupled specialists, on Qwen3-VL-8B. Shared: both stages executed by one set of parameters (UniDoc-8B). Decoupled: two independently trained specialists composed sequentially at inference. SFT prior indicates whether stage-level SFT precedes RL. Params refers to the number of parameter sets held at inference time.
Setting
UniDoc-7B
UniDoc-8B
n
Base
+ low-res context
n
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
Appendix
Table 14: Accuracy on MMLongBench-Doc, partitioned by whether the pages returned by Fast Reading contain every ground-truth evidence page. Under the proposed variant, a missing page is present at ϕlow .
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
Appendix
Table 15: MP-DocVQA accuracy under held-out and in-domain training. Held-out excludes the MP-DocVQA training set from our training data and corresponds to the values reported in Table 1 ; trained includes 4K sampled instances from it.
Figure 8: A synthesized query example. The query provides training signals for both stages: Fast Reading must retrieve the page containing the multimodal anchor “Figure 6”, while Focused Thinking must reason over “Figure 6” and its referencing text to derive the answer.
Figure 9: A synthesized query requiring multi-page evidence. Fast Reading must retrieve both the page containing “Figure 3” and the page referencing it with deeper analysis. Focused Thinking must then synthesize information across the retrieved pages to produce the answer.
Figure 10: Fast Reading comparison. The base model fails by relying on keyword matching without holistic context. Our model cross-references pages to correctly retrieve the evidence.
Figure 11: Focused Thinking comparison. Red text indicates grounding steps and green text indicates extraction steps. The base model halts after the first failed attempt. Our model iteratively re-grounds and extracts until sufficient evidence is gathered.
School of Software Engineering, Huazhong University of Science and Technology · MiLM Plus, Xiaomi Inc. · School of Computer Science and Technology, Huazhong University of Science and Technology +1