P2TPO: Structure-Aligned Page-to-Token Credit Assignment for Full-Page Document Parsing

Yong Zu; Ninghua Yang; Zezhong Tan; Xinhong Ma; Ziqiang Dong; Yang Yang

cs.LGOct 15, 2026

P2TPO: Structure-Aligned Page-to-Token Credit Assignment for Full-Page Document Parsing

Authors: Yong Zu, Ninghua Yang, Zezhong Tan, Xinhong Ma, Ziqiang Dong, Yang Yang

Organizations: Qwen Business Unit, Alibaba · Xidian University

Abstract

End-to-end full-page document parsing must preserve the content and structure of heterogeneous elements, yet otherwise accurate outputs can contain local errors and omissions. A single page-level advantage does not explicitly identify which elements differ in quality across sampled outputs or where their errors occur. We introduce Page-to-Token Policy Optimization (P2TPO), a structure-aligned credit-assignment framework that operates on full-page policy trajectories. Reference-element alignment first associates block-level quality contrasts with generated spans. Unified Anchor Pricing (UAP) incorporates matched and missing element outcomes and provides feedback relative to a fixed quality ceiling when sampled outputs receive identical but imperfect scores for a reference element. Within matched blocks, Span Credit Redistribution (SCR) redistributes existing block credit according to aligned correct and erroneous spans, using boundary proxies for within-block omissions while preserving the total block-derived advantage across its tokens. Under a fixed-weight additive page utility, we show that element-indexed contrasts retain local quality differences that page-level aggregation can cancel out. In the main experimental comparison, P2TPO outperforms all evaluated policy optimization baselines in seed-averaged scores on each of the six document parsing benchmarks. Additional experiments demonstrate gains across model scales and initializations.

Figures & tables

Figure 1: Schematic motivation for structure-aware feedback. (a) Similar page rewards hide local differences. (b) Coarse feedback obscures block identity, shared errors, and error/omission locations. (c) Structure-aware feedback retains element identity and assigns local credit. Rewards and serialized content are illustrative; training outputs use HTML tables.

Figure 2: P2TPO overview. (a) P2BPO aligns reference elements, normalizes matched outcomes, and broadcasts block credit. (b) UAP incorporates missing outcomes and imperfect-consensus feedback; SCR redistributes matched-block credit under a conserved budget. (c) The complete pipeline combines local token credit with page feedback. Missing-block coefficients reach boundary tokens directly, bypassing SCR.

Benchmark	Split	Init	GRPO	GSPO	GDPO	RDPO	P2BPO	+UAP	+SCR	P2TPO
OmniDocBench	v1.6	92.17	$92.25{\scriptstyle\,\pm\,0.34}$	$92.38{\scriptstyle\,\pm\,0.18}$	$92.31{\scriptstyle\,\pm\,0.24}$	$92.32{\scriptstyle\,\pm\,0.35}$	$92.48{\scriptstyle\,\pm\,0.15}$	$\underline{92.84}{\scriptstyle\,\pm\,0.30}$	$92.56{\scriptstyle\,\pm\,0.21}$	$\mathbf{93.12}{\scriptstyle\,\pm\,0.23}$
PureDocBench	Clean	81.28	$83.19{\scriptstyle\,\pm\,0.70}$	$83.44{\scriptstyle\,\pm\,0.31}$	$82.68{\scriptstyle\,\pm\,0.65}$	$83.10{\scriptstyle\,\pm\,0.54}$	$\underline{84.16}{\scriptstyle\,\pm\,0.62}$	$83.98{\scriptstyle\,\pm\,0.37}$	$83.78{\scriptstyle\,\pm\,0.40}$	$\mathbf{84.68}{\scriptstyle\,\pm\,0.74}$
	Digital	76.66	$79.89{\scriptstyle\,\pm\,0.48}$	$79.50{\scriptstyle\,\pm\,0.33}$	$79.24{\scriptstyle\,\pm\,0.50}$	$79.26{\scriptstyle\,\pm\,0.50}$	$\underline{80.31}{\scriptstyle\,\pm\,0.48}$	$80.26{\scriptstyle\,\pm\,0.39}$	$79.79{\scriptstyle\,\pm\,0.31}$	$\mathbf{81.03}{\scriptstyle\,\pm\,0.60}$
	Real	63.29	$68.89{\scriptstyle\,\pm\,0.64}$	$67.64{\scriptstyle\,\pm\,0.38}$	$68.23{\scriptstyle\,\pm\,0.69}$	$68.27{\scriptstyle\,\pm\,0.28}$	$69.36{\scriptstyle\,\pm\,0.46}$	$\underline{69.87}{\scriptstyle\,\pm\,0.49}$	$69.08{\scriptstyle\,\pm\,0.35}$	$\mathbf{70.33}{\scriptstyle\,\pm\,0.70}$
FinixDocBench	—	65.30	$69.36{\scriptstyle\,\pm\,0.49}$	$69.24{\scriptstyle\,\pm\,0.44}$	$69.21{\scriptstyle\,\pm\,1.42}$	$68.15{\scriptstyle\,\pm\,0.74}$	$70.03{\scriptstyle\,\pm\,0.77}$	$\underline{70.29}{\scriptstyle\,\pm\,0.62}$	$69.37{\scriptstyle\,\pm\,0.44}$	$\mathbf{70.92}{\scriptstyle\,\pm\,0.92}$
Real5- OmniDocBench	Scanning	91.06	$92.07{\scriptstyle\,\pm\,0.20}$	$92.00{\scriptstyle\,\pm\,0.17}$	$92.00{\scriptstyle\,\pm\,0.21}$	$91.90{\scriptstyle\,\pm\,0.28}$	$\underline{92.44}{\scriptstyle\,\pm\,0.29}$	$\mathbf{92.52}{\scriptstyle\,\pm\,0.16}$	$92.25{\scriptstyle\,\pm\,0.31}$	$92.40{\scriptstyle\,\pm\,0.12}$

Table 1: Cross-benchmark Overall scores (

\uparrow

, 0–100) from Qwen3.5-2B initialized by SFT on 700K pages. Init reports pre-RL scores; RL entries report mean

\pm

standard deviation over five seeds after 100 updates.

Benchmark	Split	Full P2TPO	(a) UAP interventions		w/o UAP	(b) SCR allocation controls		w/o SCR
Benchmark	Split	Full P2TPO	w/o consensus feedback	w/o omission routing	w/o UAP	Circular shift	Random support	w/o SCR
OmniDocBench	v1.6	$\mathbf{93.12}{\scriptstyle\,\pm\,0.23}$	$92.92{\scriptstyle\,\pm\,0.45}$	$92.86{\scriptstyle\,\pm\,0.23}$	$92.56{\scriptstyle\,\pm\,0.21}$	$92.73{\scriptstyle\,\pm\,0.17}$	$92.82{\scriptstyle\,\pm\,0.25}$	$92.84{\scriptstyle\,\pm\,0.30}$
PureDocBench	Clean	$\mathbf{84.68}{\scriptstyle\,\pm\,0.74}$	$84.38{\scriptstyle\,\pm\,0.94}$	$83.74{\scriptstyle\,\pm\,0.35}$	$83.78{\scriptstyle\,\pm\,0.40}$	$83.85{\scriptstyle\,\pm\,0.79}$	$83.84{\scriptstyle\,\pm\,0.16}$	$83.98{\scriptstyle\,\pm\,0.37}$
	Digital	$\mathbf{81.03}{\scriptstyle\,\pm\,0.60}$	$80.79{\scriptstyle\,\pm\,0.88}$	$80.15{\scriptstyle\,\pm\,0.36}$	$79.79{\scriptstyle\,\pm\,0.31}$	$80.08{\scriptstyle\,\pm\,0.73}$	$80.43{\scriptstyle\,\pm\,0.45}$	$80.26{\scriptstyle\,\pm\,0.39}$
	Real	$\mathbf{70.33}{\scriptstyle\,\pm\,0.70}$	$70.05{\scriptstyle\,\pm\,0.93}$	$69.67{\scriptstyle\,\pm\,0.64}$	$69.08{\scriptstyle\,\pm\,0.35}$	$69.69{\scriptstyle\,\pm\,0.78}$	$69.33{\scriptstyle\,\pm\,0.47}$	$69.87{\scriptstyle\,\pm\,0.49}$
FinixDocBench	—	$70.92{\scriptstyle\,\pm\,0.92}$	$70.42{\scriptstyle\,\pm\,1.36}$	$70.05{\scriptstyle\,\pm\,0.80}$	$69.37{\scriptstyle\,\pm\,0.44}$	$\mathbf{70.98}{\scriptstyle\,\pm\,1.22}$	$70.03{\scriptstyle\,\pm\,1.09}$	$70.29{\scriptstyle\,\pm\,0.62}$

Table 2: UAP ablations and SCR allocation controls across six benchmarks and 13 splits. Entries report mean Overall score (

\uparrow

, 0–100)

\pm

standard deviation over five runs; settings follow Table 1 .

	Parsing metrics					Omission rates (%)
Method	Text ED $\downarrow$	Formula CDM $\uparrow$	Formula ED $\downarrow$	Table TEDS $\uparrow$	Order ED $\downarrow$	Element $\downarrow$	Local $\downarrow$
P2BPO	$0.0360{\scriptstyle\,\pm\,0.0018}$	$0.9340{\scriptstyle\,\pm\,0.0028}$	$0.1279{\scriptstyle\,\pm\,0.0020}$	$0.8765{\scriptstyle\,\pm\,0.0046}$	$0.0475{\scriptstyle\,\pm\,0.0025}$	$1.09$	$6.72$
P2BPO + UAP	$\mathbf{0.0333}{\scriptstyle\,\pm\,0.0017}$	$\underline{0.9398}{\scriptstyle\,\pm\,0.0042}$	$\underline{0.1261}{\scriptstyle\,\pm\,0.0041}$	$0.8786{\scriptstyle\,\pm\,0.0059}$	$\underline{0.0459}{\scriptstyle\,\pm\,0.0020}$	$\underline{1.01}$	$\underline{5.59}$
P2BPO + SCR	$0.0364{\scriptstyle\,\pm\,0.0008}$	$0.9325{\scriptstyle\,\pm\,0.0046}$	$0.1275{\scriptstyle\,\pm\,0.0031}$	$\underline{0.8807}{\scriptstyle\,\pm\,0.0080}$	$0.0500{\scriptstyle\,\pm\,0.0017}$	$1.08$	$5.99$
P2TPO	$\underline{0.0343}{\scriptstyle\,\pm\,0.0015}$	$\mathbf{0.9446}{\scriptstyle\,\pm\,0.0027}$	$\mathbf{0.1212}{\scriptstyle\,\pm\,0.0023}$	$\mathbf{0.8833}{\scriptstyle\,\pm\,0.0061}$	$\mathbf{0.0423}{\scriptstyle\,\pm\,0.0024}$	$\mathbf{0.95}$	$\mathbf{5.53}$

Table 3: Component diagnostics on OmniDocBench v1.6 after 100 RL updates from the shared 700K-page SFT initialization. Parsing metrics report five-seed means

\pm

standard deviations; best and second-best values are bold and underlined, respectively.

Initial Conditions		Qwen3.5-2B SFT data: 50K pages					Qwen3.5-4B SFT data: 50K pages					Qwen3-VL-2B-Instruct SFT data: 700K pages
Benchmark	Split	GRPO	GSPO	GDPO	RDPO	P2TPO	GRPO	GSPO	GDPO	RDPO	P2TPO	GRPO	GSPO	GDPO	RDPO	P2TPO
OmniDocBench	v1.6	$91.00$	$90.39$	$\underline{91.05}$	$90.85$	$\mathbf{91.62}$	$\underline{92.13}$	$91.99$	$92.07$	$91.54$	$\mathbf{92.53}$	$90.45$	$90.55$	$\underline{90.57}$	$88.25$	$\mathbf{91.29}$
PureDocBench	Clean	$80.83$	$81.01$	$\underline{81.02}$	$79.75$	$\mathbf{81.98}$	$\underline{84.41}$	$84.28$	$84.07$	$82.90$	$\mathbf{84.95}$	$\underline{81.81}$	$81.80$	$81.75$	$79.67$	$\mathbf{81.95}$
	Digital	$\underline{78.26}$	$78.11$	$77.85$	$77.27$	$\mathbf{79.16}$	$\underline{81.80}$	$81.06$	$81.47$	$80.14$	$\mathbf{82.34}$	$76.08$	$76.29$	$\mathbf{76.57}$	$73.98$	$\underline{76.51}$
	Real	$\underline{69.82}$	$68.66$	$69.58$	$69.20$	$\mathbf{71.01}$	$73.55$	$72.52$	$\underline{73.61}$	$71.86$	$\mathbf{74.48}$	$65.45$	$65.44$	$\underline{65.76}$	$63.53$	$\mathbf{66.25}$
FinixDocBench	—	$69.45$	$70.11$	$\underline{70.95}$	$\underline{70.95}$	$\mathbf{73.01}$	$74.33$	$\underline{74.43}$	$73.59$	$69.38$	$\mathbf{76.74}$	$67.68$	$67.75$	$\underline{67.95}$	$65.18$	$\mathbf{68.23}$

Table 4: Cross-benchmark RL mean scores (

\uparrow

) across three SFT-initialized checkpoints. Benchmark names and splits follow Table 1 .

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Work	Target	RL algorithm	Reward construction	Optimizer signal
Infinity-Parser ( Wang et al., 2025 )	end2end	GRPO	Edit similarity $+$ paragraph count $+$ reading order	Sum $\to$ Scalar
Logics-Parsing ( Chen et al., 2025 )	end2end	GRPO	Text similarity $+$ box localization $+$ reading order	Sum $\to$ Scalar
olmOCR 2 ( Poznanski et al., 2025 )	end2end	GRPO	Unit-test pass rate $+$ termination and metadata checks	Sum $\to$ Scalar
HunyuanOCR ( Hunyuan Vision Team et al., 2025 )	end2end	GRPO	Parsing: NED-based similarity; format/length validity gate	Scalar
LightOnOCR ( Taghadouini et al., 2026 )	end2end	GRPO	OCR: unit-test pass rate with repetition, termination, math, and format checks	Scalar
TexOCR ( Wang et al., 2026a )	end2end	GRPO-style	Pass rate of binary tests derived from nine metrics	Scalar

Appendix

Table 5: RL targets and optimizer signals in document parsing. The signal column distinguishes response aggregation, functional-attribute routing, and reference-block credit assignment. Target and signal labels are defined below.

Setting	Specification
Training framework	veRL with Megatron actor training and asynchronous vLLM generation; BF16.
RL budget	100 updates; 128 page prompts per update; $G=8$ responses per prompt.
PPO minibatch	128 prompts, expanded to 1,024 responses; one PPO epoch.
Sequence limits	8,192 prompt tokens (including image tokens) and 8,192 response tokens; total model length 16,384.
Rollout sampling	Temperature 1.0, top- $p=1$ , top- $k$ disabled.
Optimizer	Megatron Adam; learning rate $10^{-6}$ , moments $(0.9,0.999)$ , weight decay $0.01$ , constant schedule, no warmup.

Appendix

Table 6: Training configuration for P2TPO and its component variants, which share optimizer and loss settings. Baseline-specific policy objectives are described in Appendix C.1.4 .

Study	Initialization	RL records / updates	Summary	Evaluation
Primary comparison and UAP/SCR combinations	Qwen3.5-2B; 700K SFT	12,800 / 100	$\mu\pm\sigma$ ; 5 runs	All splits; content and omission diagnostics on OmniDocBench
Internal UAP interventions	Qwen3.5-2B; 700K SFT	12,800 / 100	$\mu\pm\sigma$ ; 5 paired runs	All splits
SCR allocation controls	Qwen3.5-2B; 700K SFT	12,800 / 100	$\mu\pm\sigma$ ; 5 runs	All splits
Initialization and scale	Qwen3.5-2B and 4B; same 50K SFT manifest	12,800 / 100	$\mu\pm\sigma$ ; 5 data seeds	All splits
Additional backbone	Qwen3-VL-2B-Instruct; 700K SFT	12,800 / 100	$\mu\pm\sigma$ ; 5 data seeds	All splits
Parameter sensitivity	Qwen3.5-2B; 700K SFT	12,800 / 100	$\mu\pm\sigma$ ; 5 runs	All splits

Appendix

Table 7: Configuration of each experimental series. SFT counts specify initialization data; RL counts specify training records. “All splits” denotes the six benchmarks and 13 splits in Table 1 . Overall scores are summarized by means and standard deviations across five runs.

Benchmark	Pages	Evaluated Scope	Engine
OmniDocBench v1.6 ( Ouyang et al., 2025 ; OmniDocBench contributors, 2026 )	1,651	Official v1.6 ground truth, including 1,355 v1.5 pages.	OmniDocBench v1.6
Real5-OmniDocBench ( Zhou et al., 2026 )	1,355 per condition	Scanning, Warping, Skew, Illumination, Screen photo; shared v1.5 ground truth.	OmniDocBench v1.6
Wild-OmniDocBench ( Li et al., 2026b )	1,350	Real captures of a subset of v1.5 pages.	OmniDocBench v1.6
DocPTBench SynReal ( Du et al., 2025 ; DocPTBench authors, 2025 )	1,381	Photographed set: simulated and real captures.	OmniDocBench v1.6
DocPTBench Dewarp ( Du et al., 2025 ; DocPTBench authors, 2025 )	1,381	Unwarping set: rectified Photographed images.	OmniDocBench v1.6
PureDocBench ( Li et al., 2026c ; PureDocBench authors, 2026 )	1,475 per condition	Clean, Digital, Real; ground truth uses the v1.6 page schema.	OmniDocBench v1.6

Appendix

Table 8: Benchmark scope and evaluation engines. Page counts specify the evaluated sets. Engine names identify the shared OmniDocBench v1.6 engine or the official FinixDocBench Markdown evaluator.

Figure 5: Frozen responses / no policy update. Final token-wise advantages on four selected saved rollout groups. Methods are evaluated on common responses; each method row contains eight rollout bands. Horizontal position indexes generated tokens, and panel headers give page-reward ranges. Red/blue denote negative/positive advantages on a shared zero-centered scale, with colors saturating beyond

\pm 1

. Panels show (a,b) shorter/longer responses, (c) magnitude variation across rollouts, and (d) within-response sign changes.

Panel of Table 9 — Table 9: Cross-benchmark RL evaluation across model scales and SFT-initialized checkpoints. Metrics and splits follow Table 1 . Within each panel, Init reports pre-RL scores and RL methods share the starting checkpoint and 100-update budget.

Document Parsing

Document parsing converts PDFs and page images into structured, machine-readable text, tables, formulas, layouts, and reading order.

See all 63 papers

Jul 21, 2026

HPD-Parsing: Hierarchical Parallel Document Parsing

Efficient teamwork typically combines global coordination with parallel execution, a principle not yet fully reflected in unified Vision-Language Model (VLM)-based document parsers. Existing unified parsers process an entire page jointly but generate its output through a single token-by-token autoregressive trajectory, creating a sequential bottleneck that grows with document length. Such full-page sequential generation overlooks a key property of document parsing: layout must be analyzed globally, whereas block content can be parsed in parallel. Based on this observation, we introduce HPD-Parsing, which replaces full-page autoregressive generation with a Hierarchical Parallel Decoding paradigm. A main layout branch organizes the overall document structure and dynamically assigns block-level content decoding to concurrent branches, while progressive multi-token prediction (P-MTP) further reduces the decoding steps within each branch. Experiments on public benchmarks show that HPD-Parsing achieves 4,752 tokens per second, delivering

2.62\times

the throughput of the fastest existing document parsing model and

3.06\times

that of the vanilla autoregressive baseline, while maintaining competitive parsing accuracy. These results establish hierarchical parallel decoding as an effective alternative to full-page autoregressive generation, opening a new direction for efficient unified document parsing.

Shu Wei, Jingjing Wu, Lingshu Zhang +10

Efficient VLM Inference Document Parsing Parallel Decoding

Ask PDF

Aug 1, 2026

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

Reinforcement learning (RL) for document parsing often relies on reference-based rewards rooted in edit distance (e.g., tree edit distance), yet it remains hard to optimize in the high-accuracy regime because such rewards become weakly discriminative: near-correct outputs receive very similar scores, providing limited learning signal for hard cases. We propose Step-Aware Annealing (SAA), a plug-and-play reward sharpening mechanism that progressively increases reward curvature during training, amplifying subtle quality differences among high-scoring samples while preserving stability in early learning. Built on SAA, we introduce DocPO, a document policy optimization framework with element-specific, reference-based rewards anchored by edit-distance signals: normalized string edit distance (NED) for text, tree edit distance similarity (TEDS) for tables, and a hybrid Rubric+edit reward for formulas. Experiments on OmniDocBench and DocElemHard show that SAA consistently improves GRPO-style RL across document elements over non-annealed rewards, without requiring additional human supervision for reward construction.

Yunhao Wang, Binghong Wu, Zhenyu Huang +4

Tencent Hunyuan · Beijing, China · Shanghai, China +1

Reward Shaping Document Parsing Reward Design for RL

Ask PDF

Aug 6, 2026

PaDoc: Layout-Grounded Parallel Decoding for Document Parsing

End-to-end document parsers provide a unified interface, but serialize page layouts and regional contents into one autoregressive sequence. This formulation forces independent regions onto a decoding path whose length grows with the total content, whereas crop-based two-stage parsers expose region-level parallelism at the cost of repeated visual prefills and fragmented page context. To retain full-page context while removing dependencies, we propose PaDoc, a layout-grounded parser that treats the predicted layout as a branching structure over a shared page representation. Under a region-sufficiency assumption, we derive a prefix-conditioned factorization in which the layout stream and regional content branches advance concurrently, reducing the decoding depth to the longest layout-content path. We realize this factorization within a single MLLM: packed variable-length ancestor attention preserves the visibility under standard next-token training, while masked parallel decoding creates branches that the evaluated vLLM backend serves as concurrent requests with cache-resident shared-prefix reuse. On OmniDocBench Full, PaDoc attains an Overall layout F1 of 91.1 and, among end-to-end parsers, a top-tier Overall score of 94.24 together with the best Text Edit (0.038) and Formula CDM (95.59). On a 384-page subset and one A800 GPU, it is the fastest end-to-end parser at five concurrency levels, improving valid-page throughput by 67.4-118% and reducing P95 latency by 39.2-54.9% relative to a same-backbone Sequential SFT baseline. Code is available at https://github.com/Longin-Yu/Padoc

Hao Yu, Jiabo Zhan, Kang Liu +8

1Tsinghua University · 2Wechat Vision, Tencent

LLM Inference Acceleration Document Parsing Parallel Decoding Document Layout Analysis

Ask PDF

Abstract

Document Parsing

Explore similar work

HPD-Parsing: Hierarchical Parallel Document Parsing

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

PaDoc: Layout-Grounded Parallel Decoding for Document Parsing