Executive summary

DAVE, MonkeyOCRv2, and DocPO address different stages of a document vision-language model (VLM) lifecycle rather than competing as three interchangeable training recipes.

Method	Primary stage	Main contribution	Best use case
DAVE	Vision-encoder pretraining and multi-decoder alignment	Domain-specific MAE, supervised structural pretraining, generalist feature fusion, and decoder-agnostic weight merging	Large quantities of unlabeled document pages; reusable encoder for parsing, VQA, grounding, and web/UI tasks
MonkeyOCRv2	Visual-text encoder pretraining and downstream VLM construction	Joint text generation and pixel/stroke reconstruction on large multilingual page and element data	High-fidelity OCR, formulas, tables, multilingual documents, and reducing dependence on language priors
DocPO	RL post-training after SFT	Element-specific rewards and Step-Aware Annealing for GRPO	Improving an already accurate document parser, especially on hard text, table, and formula cases

The methods fit naturally into the following lifecycle:

flowchart TD
    A["Unlabeled document pages"] --> B["DAVE-style MAE"]
    B --> C["Visual-text alignment and structured SFT"]
    C --> D["DocPO reinforcement learning"]
    E["Labeled and pseudo-labeled crops"] --> C

MonkeyOCRv2 provides an alternative or additional bridge between the first two steps: it jointly teaches visual fidelity and text generation before the encoder is connected to the final LLM. DocPO belongs at the end; it is neither an encoder pretraining method nor a substitute for alignment and SFT.

1. Scope and terminology

A conventional document VLM contains:

document image -> vision encoder -> projector/connector -> language model -> structured output

This survey distinguishes four training phases:

1. Visual self-supervised pretraining: learn document structure from images without text labels.
2. Visual-text pretraining/alignment: make visual tokens usable by an autoregressive decoder.
3. Supervised fine-tuning (SFT): learn task instructions and output formats such as Markdown, HTML, LaTeX, JSON, bounding boxes, and reading order.
4. Reinforcement learning with verifiable rewards (RLVR): optimize sequence-level accuracy beyond teacher-forced SFT.

DAVE emphasizes phases 1 and 2. MonkeyOCRv2 combines visual reconstruction and autoregressive textual supervision before using the encoder in downstream systems. DocPO specializes phase 4.

2. DAVE

Paper: DAVE: A VLM Vision Encoder for Document Understanding and Web Agents

2.1 Motivation

General-purpose encoders such as SigLIP or DINO are largely trained on natural images. Their high-level semantic features are useful, but they are not optimized for the fine spatial structure of documents, charts, web pages, and user interfaces. DAVE therefore builds a specialist encoder while retaining access to a generalist encoder’s semantic representation.

The paper also identifies a practical compatibility problem: an encoder trained with one language decoder can become tightly coupled to that decoder and transfer poorly to another. DAVE explicitly trains with several decoder families and merges the resulting encoders.

2.2 Stage 1: self-supervised document MAE

DAVE first trains a ViT-L/16 encoder from scratch using Masked Autoencoding. With masked patch set $\mathcal M$, it reconstructs the raw pixel values of the hidden patches:

$$\mathcal L_{\text{MAE-pixel}} = \frac{1}{|\mathcal M|} \sum_{i\in\mathcal M} \left\|f_\theta(\tilde x)_i-x_i\right\|_2^2.$$

This differs from the original MAE objective, which normalizes pixels inside every patch before calculating the loss. DAVE reports that per-patch normalization becomes unstable on document and web images because many neighboring patches have low visual variance. Direct raw-pixel reconstruction stabilizes training.

Key Stage-1 settings reported in the paper are:

Item	Setting
Encoder	ViT-L/16 at 384 resolution
Initialization	From scratch
Training images	20M
Documents	10M DocFM images
Web screenshots	10M Common Screen images
Mask ratio	75%
Batch size	4,096
Training	25 epochs / approximately 120K steps
Hardware	32 H200 GPUs

After this stage, the lightweight MAE decoder is discarded and the encoder is retained.

2.3 Stage 2: supervised autoregressive pretraining

The MAE encoder is inserted into a VLM:

DAVE encoder -> MLP projector -> pretrained text decoder

The model is trained autoregressively on OCR, layout extraction, document grounding, chart/table extraction, LaTeX generation, HTML table generation, and web-UI grounding. The listed datasets contain roughly 2.08M images in total, including FM4D, PlotQA, FinTabNet, ChartQA, DaTikZ, PubTables, UGround, and a self-curated grounding/structure dataset.

The authors train encoder instances with three different LLMs:

• Qwen2.5-0.5B-Instruct
• Phi-4-mini-Instruct
• Granite-3.1-3B-Instruct

Supervised pretraining uses a maximum sequence length of 20K, learning rate $3\times10^{-5}$, one epoch, and AnyRes-style 384-pixel tiling.

2.4 Generalist feature ensemble

DAVE concatenates features from two encoders:

$$\phi_{\text{DAVE}}(x) = \operatorname{Concat} \left( \phi_{\text{gen}}(x), \phi_{\text{spec}}(x) \right),$$

where the generalist branch is a frozen SigLIP2 encoder and the specialist branch is the document/web MAE encoder. This lets the specialist focus on low-level spatial structure without sacrificing the general visual semantics already represented by SigLIP2.

The trade-off is that the final visual front end depends on an external generalist encoder and produces more visual features than a single-encoder design.

2.5 Multi-decoder weight merging

If $\phi_1,\ldots,\phi_n$ are encoder instances aligned with different text decoders, DAVE constructs merged weights:

$$\theta_{\text{merge}}^{(j)} = \sum_{i=1}^{n} \alpha_i^{(j)}\theta_i^{(j)}.$$

The original encoder parameters remain frozen while the merge coefficients are learned through feature distillation. This produces an encoder intended to be less dependent on any single LLM family.

This part is valuable when the vision encoder must later be tested with several LLMs. It is less valuable when the production model has one fixed decoder and training multiple full encoder–decoder combinations is too expensive.

2.6 Evidence, strengths, and limitations

The paper reports that DAVE improves the average performance over SigLIP2 by 10.5% across its document/web evaluation and improves Mind2Web agent performance by about 5% over the strongest encoder baseline. Its most convincing contribution is the demonstration that domain-specific unlabeled pretraining produces visual representations that transfer beyond pure OCR.

Strengths:

• Exploits very large unlabeled document collections.
• Learns layout and spatial priors without OCR labels.
• Covers document, chart, UI grounding, and web-agent tasks.
• Explicitly addresses compatibility with different language decoders.
• Retains general semantics through feature fusion.

Limitations:

• Fixed-resolution encoder with image tiling rather than native dynamic resolution.
• Large ViT-L training requires substantial compute.
• Dependence on an external SigLIP2 branch complicates deployment.
• Multi-decoder training and merging increase experimental cost.
• The MAE representation still requires supervised visual-language training before it is useful in a VLM.

3. MonkeyOCRv2

Paper: MonkeyOCRv2: A Visual-Text Foundation Model for Document AI

3.1 Motivation

MonkeyOCRv2 argues that document encoders must preserve character-level evidence. In natural-image recognition, small variations in texture or local appearance are often irrelevant. In documents, a decimal point, stroke, superscript, or table border can change the meaning completely.

Autoregressive text-only pretraining is also insufficient: a decoder can exploit language context to guess visually unclear characters. MonkeyOCRv2 combines text generation with image reconstruction so that visual tokens retain both semantic content and the underlying glyph evidence.

3.2 Data engine

MonkeyDoc v2 contains 113M samples across 17 languages, including Vietnamese:

Subset	Scale
Full pages	8M
Cropped document elements	105M
Real samples	61M (54%)
Synthetic samples	52M (46%)
Formula samples	Approximately 0.8M

The strong emphasis on element crops gives dense supervision for text, formulas, and tables. Real documents are labeled with a multi-expert agreement pipeline: several specialist recognizers predict each crop, and the prediction with the highest average agreement is retained.

Synthetic data are generated by rendering multilingual corpora, rare Unicode characters, random character sequences, formulas, and multilingual table content into varied fonts and layouts. Two notable filtering rules are used:

• Mask all detected regions; reject the page if a document VLM can still read residual text, indicating missed layout regions.
• Concatenate content in annotated reading order and use an LLM to reject inconsistent or incorrectly ordered pages.

3.3 Dual-objective encoder pretraining

The pretraining system contains a vision encoder $E_v$, vision decoder $D_v$, and autoregressive text decoder $D_t$:

$$\mathbf z=E_v(I), \qquad \hat I=D_v(\mathbf z).$$

The default reconstruction loss is full-image pixel MSE:

$$\mathcal L_{\text{pix}} = \frac{1}{3HW}\left\|\hat I-I\right\|_2^2.$$

The text decoder consumes the same visual tokens and predicts OCR or structured text using cross-entropy. The joint objective is:

$$\mathcal L_{\text{pretrain}} = \mathcal L_{\text{text}} + \lambda\mathcal L_{\text{rec}}.$$

The paper sets $\lambda=1$. Unlike DAVE’s masked-patch reconstruction, MonkeyOCRv2 describes full-image reconstruction from the encoded tokens.

The authors also evaluate a stroke-aware variant:

$$\mathcal L_{\text{rec}} = \mathcal L_{\text{pix}} + \alpha\mathcal L_{\text{struct}},$$

where $\mathcal L_{\text{struct}}$ compares Sobel-derived edge maps and differentiable distance-to-edge maps. The reported coefficients are $\alpha=0.5$, $\beta=0.25$, and $\lambda=1$. Most experiments use MSE-only reconstruction; the stronger edge-aware variant is evaluated in the controlled document-understanding experiment.

3.4 Encoder family and training

Variant	Backbone	Parameters	Intended role
MonkeyOCRv2-S	ViT-Small	28M	Recognition and compact VLMs
MonkeyOCRv2-B	ViT-Base	113M	Parsing and document understanding
MonkeyOCRv2-AS	ViTAEv2-Small	21M	Detection, segmentation, and multi-scale localization

The encoders are trained from scratch on 64 A800 GPUs with a peak learning rate of $10^{-3}$, global batch size 256, dynamic resolution, and approximately one-million-pixel input budget for the ViT variants.

After pretraining, both the vision decoder and the temporary text decoder are discarded. Only the vision encoder is transferred.

3.5 Downstream VLM construction

For document parsing, the retained encoder is connected through an MLP projector to Qwen3-0.6B:

MonkeyOCRv2-S/B -> MLP projector -> Qwen3-0.6B

The encoder remains frozen throughout:

1. Train only the projector for alignment at learning rate $2\times10^{-4}$.
2. Train the projector and Qwen3-0.6B jointly at learning rate $2\times10^{-5}$.

The parsing system first generates layout categories, bounding boxes, and reading order from the page. It then crops each predicted element and applies the same VLM again with an element-specific prompt. The outputs are assembled into the final document representation.

For document-understanding experiments, the paper instead pairs frozen encoders with Qwen3-1.7B through an MLP projector.

3.6 Evidence, strengths, and limitations

The most important ablation is the controlled comparison where the downstream LLM, data, training, and decoding remain fixed. The average document-understanding score increases from 50.7 with text-generation-only pretraining to 51.7 with MSE reconstruction and 55.9 with edge/distance-aware reconstruction for the Small encoder. The Base edge-aware model reaches 57.2.

In the scrambled-text experiment at resolution 448, reconstruction improves accuracy from 55.4% to 72.1% and reduces the coherent-versus-scrambled accuracy gap from 29.3% to 15.3%. This supports, but does not fully prove, the claim that reconstruction reduces dependence on language priors.

MonkeyOCRv2-B-Parsing reports 83.3 on multilingual MDPBench versus 80.5 for dots.mocr. On OmniDocBench 1.6 it reaches 91.57 with the encoder frozen and without progressive post-training, but remains below the strongest specialized parsers. The authors appropriately warn that full-system comparisons are not controlled because systems differ in data, layout models, post-processing, and post-training.

Strengths:

• Strong character-level visual grounding.
• Explicit multilingual and rare-character supervision.
• Encoder remains useful for recognition, detection, segmentation, parsing, and VQA.
• Dynamic-resolution training and compact encoder sizes.
• Controlled evidence that reconstruction improves document VLM performance.

Limitations:

• Requires large quantities of labeled, pseudo-labeled, or synthetic data.
• 113M-sample pretraining is expensive to reproduce.
• English and Chinese remain dominant despite 17-language coverage.
• Pretraining decoder designs and loss schedules are not fully isolated.
• Crop-then-recognize inference improves fidelity but adds latency.
• Page-level structure receives far fewer samples than element-level recognition.

4. DocPO

Paper: DocPO: Advancing Document Policy Optimization via Tailored Step-Aware Rewards

4.1 Motivation

DocPO starts from an already SFT-trained document VLM. It addresses two problems in applying GRPO to document parsing:

1. Plain string edit distance is unsuitable for heterogeneous document elements.
2. When the model is already accurate, multiple rollout answers receive very similar high scores, yielding weak within-group advantages.

DocPO keeps the Qwen2.5-VL architecture unchanged and operates entirely at the reward and optimization level.

4.2 Element-specific base rewards

All base rewards are normalized to $[0,1]$.

For text, DocPO uses character-level normalized edit similarity:

$$R_{\text{text}} = 1- \frac{\operatorname{EditDist}(y,y^*)} {\max(|y|,|y^*|)}.$$

For tables, it uses weighted TEDS/APTED over the HTML tree:

$$R_{\text{table}} = 1- \frac{\operatorname{TreeEditDist}_{w}(y,y^*)} {\max(|y|,|y^*|)}.$$

Structural node insertion/deletion and span errors receive cost 2, while content mismatches receive cost 1. This penalizes a broken rowspan or colspan more strongly than a character typo.

For formulas, it combines syntax validity, semantic equivalence, and edit similarity:

$$R_{\text{formula}} = v_{\text{syn}} \left( 0.8r_{\text{sem}}+0.2r_{\text{struct}} \right).$$

Malformed LaTeX receives zero through $v_{\text{syn}}$. Semantic equivalence is judged zero-shot by Qwen2.5-7B-Instruct, while the structural component is normalized edit similarity.

4.3 Step-Aware Annealing

DocPO’s central contribution is a power-law transformation of the base reward $M$:

$$R_{\text{shaped}}=M^\gamma.$$

The exponent grows during training:

$$\gamma(s) = \gamma_{\text{init}} + \Delta_\gamma \left[ 1-\exp\left(-\frac{s}{\tau_{\text{adaptive}}}\right) \right],$$

with $\gamma_{\text{init}}=1$ and $\Delta_\gamma=8$. Early training therefore sees almost unchanged rewards. Later training increasingly separates high-quality candidates while compressing weak candidates toward zero.

For example:

Base score	$\gamma=1$	$\gamma=8$
0.90	0.900	0.430
0.95	0.950	0.663
Gap	0.050	0.233

Because the transformation is monotonic, it preserves ranking. Its purpose is to make small differences among near-correct rollouts more discriminative after GRPO group normalization.

A Dynamic Dispersion Controller computes the rolling coefficient of variation separately for text, table, and formula rewards. It changes the annealing time constant so that the sharpening rate adapts to recent task-specific reward dispersion.

4.4 Training recipe

The paper uses:

Item	Setting
Base model	Qwen2.5-VL-3B; scaling test on 7B
Initial SFT data	490K full-page samples
RL data	612K element patches
RL tables	206K
RL formulas	196K
RL text blocks	210K
Rollouts per example	8
RL learning rate	$10^{-6}$
Global RL batch size	128
Input/output limits	4K / 8K
KL constraint	None

Before RL, the authors remove examples for which all eight rollouts are perfect and examples for which all eight receive zero. Both groups have no useful reward variance for GRPO. Training focuses on examples where the current policy sometimes succeeds and sometimes fails.

4.5 Evidence, strengths, and limitations

Element	SFT baseline	Intermediate RL	Final DocPO
Text NED ↓	0.0358	0.0238 with edit reward	0.0125
Formula CDM ↑	92.61	93.93 with Rubric+edit	94.70
Table TEDS ↑	89.30	91.70 with APTED	93.01

Dynamic SAA outperforms fixed exponents in the table ablation. It also reaches a table TEDS threshold of 91 in 587 steps versus 1,070 without SAA, approximately 1.8 times faster according to the paper.

Strengths:

• No modification of the vision encoder or VLM architecture.
• Interpretable rewards aligned with each element’s native structure.
• Focuses compute on informative hard cases.
• Improves a strong model in the high-accuracy regime.
• Does not require a separately trained reward model for text and tables.

Limitations:

• Primarily element-level RL; it does not directly optimize full-page coverage or reading order.
• Formula rewards require an external LLM semantic judge.
• Training rewards are closely related to evaluation metrics, so downstream reasoning gains are not established.
• Removing the KL constraint may increase specialization and forgetting risk.
• RL is unlikely to help when all sampled answers are uniformly incorrect.
• The self-constructed DocElemHard benchmark requires independent external validation.

5. Direct comparison

Dimension	DAVE	MonkeyOCRv2	DocPO
Model component improved	Vision encoder	Vision encoder	Full VLM policy
Starting point	Randomly initialized ViT	Randomly initialized compact encoder	SFT document VLM
Primary supervision	Unlabeled images, then structured labels	Paired visual-text data plus reconstruction	Reference outputs and verifiable rewards
Main objective	Masked raw-pixel reconstruction, then autoregressive multitask learning	Text cross-entropy plus full-image/stroke reconstruction	GRPO with task-specific shaped rewards
Document scale	20M pages/screens in SSL; ~2.08M supervised images	113M samples, mostly element crops	490K SFT pages and 612K RL crops
Resolution	Fixed base resolution plus tiling	Dynamic resolution	Inherited from Qwen2.5-VL
LLM relationship	Trains with multiple LLMs and merges encoders	Temporary decoder discarded; downstream Qwen3-0.6B/1.7B	Directly post-trains Qwen2.5-VL
Key advantage	Efficient use of unlabeled data and structural learning	Fine visual fidelity and reduced language shortcuts	Fine-grained improvement after SFT saturation
Main cost	Large MAE and multi-decoder training	Massive labeled/pseudo-labeled corpus	Multiple rollouts and reward computation
Best evaluation level	Encoder transfer and VLM tasks	Encoder transfer, parsing, and understanding	Element parsing after SFT

5.1 DAVE versus MonkeyOCRv2

These methods share the goal of building a document-native encoder, but their supervision differs:

• DAVE first learns from unlabeled full pages using masked reconstruction, then adds supervised parsing and grounding.
• MonkeyOCRv2 jointly learns textual output and full visual reconstruction on paired page/element data.

DAVE is preferable when unlabeled page scale is the main asset. MonkeyOCRv2 is preferable when reliable OCR, table, formula, and synthetic labels can be generated at scale and character fidelity is the dominant requirement.

They can be combined, but the hybrid is an inference from the two papers rather than a recipe validated by either paper:

DAVE-style MAE initialization
-> MonkeyOCRv2-style text generation + reconstruction continuation
-> final LLM alignment and structured SFT

5.2 Encoder pretraining versus DocPO

DocPO is complementary to both encoder methods. Better vision features reduce perceptual errors; DocPO improves how the complete autoregressive model chooses and structures its outputs. DocPO cannot recover visual details that the encoder discarded, and encoder pretraining alone cannot optimize sequence-level TEDS, LaTeX validity, or output formatting.

6. Recommended recipe for the current project

Assumed current assets:

• Approximately 20M document pages.
• A document MAE encoder already trained on about 2.6M images.
• Approximately 100K annotated images for visual-language alignment.
• One DGX node with 8 H200 GPUs.
• A target document VLM/decoder such as Chandra 2.

The most practical strategy is not to restart all three papers literally.

Stage A — preserve the existing MAE checkpoint

Treat the current encoder as the DAVE Stage-1 result. Before additional training, evaluate it with frozen linear or lightweight heads on:

• OCR character recognition
• Layout element classification/localization
• Reading-order prediction
• Table-region recognition
• Document VQA with a controlled projector and LLM

Compare against the original general-purpose encoder under identical downstream training.

Stage B — add MonkeyOCRv2-style visual-text continuation

Create a mixture of full pages and fine-grained crops from the 100K verified annotations and pseudo-labeled portion of the remaining corpus. Continue training the MAE encoder with:

$$\mathcal L = \mathcal L_{\text{text}} + \lambda_{\text{pix}}\mathcal L_{\text{pix}} + \lambda_{\text{stroke}}\mathcal L_{\text{struct}}.$$

Recommended targets include OCR transcription, table HTML, formula LaTeX, element class plus bounding box, reading order, and page Markdown. This stage should begin with a low encoder learning rate because the encoder is already pretrained; the exact schedule requires ablation because MonkeyOCRv2 trains from scratch.

Stage C — align with the final LLM

Use the production decoder rather than reproducing DAVE’s three-decoder merge initially:

1. Freeze the encoder and LLM; train the projector.
2. Keep the encoder frozen; train the projector and LLM on structured document targets.
3. Optionally unfreeze only the last encoder blocks at a substantially lower learning rate if controlled evaluation shows a consistent gain.

Multi-decoder merging should be a separate research branch only if encoder portability across several LLM families is an explicit goal.

Stage D — structured SFT

Train a balanced instruction mixture covering:

• Page-to-Markdown parsing
• Layout elements, bounding boxes, and reading order
• Text blocks and paragraphs
• Tables with spanning cells in HTML/JSON
• Formulas in normalized LaTeX
• Charts/figures and caption association
• Schema-based information extraction
• General multimodal instructions to limit catastrophic forgetting

Stage E — DocPO on informative hard cases

After SFT produces valid outputs reliably:

1. Generate multiple rollouts per element.
2. Remove all-perfect and all-zero groups.
3. Apply NED to text, TEDS to tables, and syntax-gated semantic/edit reward to formulas.
4. Add project-specific rewards for layout class, bounding-box matching, reading order, and full-page element coverage.
5. Apply SAA separately to normalized rewards.
6. Retain a small KL penalty or general-data replay unless experiments demonstrate that specialization causes no unacceptable forgetting.

DocPO should begin with element patches. A later page-level RL stage can add:

$$R_{\text{page}} = w_cR_{\text{coverage}} +w_oR_{\text{order}} +w_lR_{\text{layout}} +w_tR_{\text{text}} +w_bR_{\text{table}} +w_fR_{\text{formula}}.$$

This page-level extension is a proposed adaptation, not part of the original DocPO method.

7. Minimum experiment matrix

To identify which component creates the gain, avoid comparing only the final hybrid against a baseline.

Experiment	Encoder initialization	Continued pretraining	VLM training	RL
E0	Original general encoder	None	Same SFT	None
E1	Existing document MAE	None	Same SFT	None
E2	Existing document MAE	Text generation only	Same SFT	None
E3	Existing document MAE	Text + pixel reconstruction	Same SFT	None
E4	Existing document MAE	Text + pixel + stroke reconstruction	Same SFT	None
E5	Best E2–E4	Same structured SFT	Basic element GRPO	Basic rewards
E6	Best E2–E4	Same structured SFT	DocPO	Element rewards + SAA
E7	Best E2–E4	Same structured SFT	Extended DocPO	Element + page rewards

Use identical downstream data, resolution, projector, LLM, decoding, and evaluation for E0–E4. This is essential for attributing gains to the encoder rather than to changes in data or decoder capacity.

8. Evaluation framework

Capability	Suggested metrics
OCR fidelity	CER, NED, exact match, rare-character accuracy
Language-prior dependence	Coherent-versus-scrambled gap, CHAOS-style perturbed-word recall
Layout	Class F1, mAP/IoU, element coverage
Reading order	Pairwise-order accuracy, edit distance, graph-edge F1
Tables	TEDS, TEDS-Structure, cell content F1, span accuracy
Formulas	CDM, exact match, compile success, render similarity
Full-page parsing	OmniDocBench-style aggregate plus page validity
Document understanding	DocVQA, InfoVQA, ChartQA, domain-specific extraction accuracy
Robustness	Low resolution, blur, skew, watermark, photographed pages, mixed scripts
Efficiency	Visual tokens, latency, peak memory, throughput, output length
General capability retention	General VQA/instruction benchmark before and after specialization

For model selection, prioritize held-out in-domain documents and difficult out-of-domain subsets rather than optimizing only the same metrics used as RL rewards.

9. Final assessment

• DAVE provides the best foundation when unlabeled page-scale data are abundant. Its raw-pixel MAE formulation is directly relevant to document images and has already influenced the current encoder training direction.
• MonkeyOCRv2 contributes the most valuable next experiment: add autoregressive textual supervision while retaining pixel/stroke reconstruction so the encoder learns meaning without discarding visual evidence.
• DocPO is the appropriate final refinement method once the VLM is already accurate and produces diverse near-correct rollouts. Its reward design should be extended with layout and reading-order signals for a complete document parser.

For the current project, the recommended path is therefore:

existing DAVE-style MAE checkpoint
-> MonkeyOCRv2-inspired visual-text continuation
-> production-LLM alignment
-> structured document SFT
-> DocPO element RL
-> optional page-level RL

This combines the strongest idea from each paper while avoiding the cost of reproducing all three systems from scratch.

References

1. Huang, B. et al. DAVE: A VLM Vision Encoder for Document Understanding and Web Agents. arXiv:2512.17221, 2025.
2. Liu, Y. et al. MonkeyOCRv2: A Visual-Text Foundation Model for Document AI. arXiv:2607.11562, 2026.
3. Wang, Y. et al. DocPO: Advancing Document Policy Optimization via Tailored Step-Aware Rewards. arXiv:2608.00536, 2026.