Financial document parsing requires accuracy, structural consistency, and verifiability that current benchmarks often fail to reflect. We present FinixDoc, an end-to-end agentic parsing system for real-world financial documents, with FinixDoc-VL, a 4B-scale vision-language model built on Qwen3-VL-4B, as its core parser. To characterize the gap between benchmark and deployment performance, we introduce a Document Parsing Capability Matrix organized along two practical axes: visual quality and document scale. Guided by this matrix, FinixDoc-VL is trained with a domain-adapted recipe combining homoglyph-aware contrastive learning and multi-stage reinforcement learning with composite domain-specific rewards. To better leverage our accumulated advantage in low-quality financial-document data and support large-scale, high-quality data production, we further build a human-in-the-loop Data Factory pipeline with confidence-aware expert review. For evaluation, we construct FinixDocBench, a financial-domain evaluation suite covering digital-native, camera-captured, ultra-large-page, and internal-workflow scenarios, with a compliance-reviewed subset released alongside this technical report. On its main subsets, FinixDoc-VL achieves the highest overall score (81.43) among evaluated baselines, outperforming the next-best open-source model by 5.13 points, with the largest gains on internal financial workflows (FinixInner: 84.08 vs. 78.73).
Figures & tables
Figure 2: Document Parsing Capability Matrix. The horizontal axis represents decreasing document quality, ranging from digital or scanned documents to camera-captured documents affected by perspective distortion, creases, and blur. The vertical axis represents increasing document scale, ranging from single-page or single-image inputs to ultra-long documents and ultra-large images. Based on this matrix, document parsing scenarios can be categorized into the Benchmark-Converged Zone, Low-Quality Zone, Underexplored Large-Scale Zone, and Ambiguous-Unrecoverable Zone.
Figure 3: Overall architecture of FinixDoc. The output format (JSON or Markdown) is specified by the user prompt.
Figure 4: Parsing output examples on example.pdf : Markdown (top) vs. JSON (bottom).
Figure 5: Overall training framework of the proposed financial visual representation adaptation method via contrastive learning. Given a document image and its ground-truth text, hard negative texts are generated through financial homoglyph substitution. The model is then optimized with dual-path contrastive objectives over both visual and multimodal generative representations, jointly with the supervised fine-tuning objective.
Figure 6: Overview of reinforcement learning fine-tuning with domain-specific rewards. The RL stage jointly optimizes JSON and Markdown generation through multi-stage sub-capability calibration followed by unified GRPO alignment. Task-dependent reward masks combine text fidelity, detection quality, and reading-order consistency according to the output format.
Source
Dataset / Document Type
Size
Out-of-domain
DocLayNet
45k
Out-of-domain
Infinity-Doc-55K
55k
In-domain
Identity Documents
20k
In-domain
Medical Examination Reports
10k
In-domain
Financial Research Reports
10k
In-domain
Insurance Policies
10k
Table 1: Composition of the training data used in the reinforcement learning stage.
Task
α
β
1−α−β
δtext
δdet
δorder
JSON
0.5
0.4
0.1
1
1
1
Markdown
0.5
0.4
0.1
1
0
1
Table 2: Reward hyperparameters and mask settings for the JSON and Markdown tasks.
Figure 7: Overview of the Data Factory pipeline.
Document Type
Source Type
Evaluation Track
Volume (Pages)
Financial Reports & Insurance Clauses
Digitally Native
FinixDigital
500
Medical Receipts
Mobile Captured
FinixPhoto
300
Large-Scale Financial Pages
Mixed Quality
FinixHuge
200
Identity Documents
Mobile Captured
FinixInner
800
Medical Examination Reports
Mobile Captured
FinixInner
800
Claim Medical Records
Mobile Captured
FinixInner
800
Table 3: Data composition of FinixDocBench.
Model
Document-level Accuracy ( ↑ )
Donut
0.922
MinerU
0.924
Qwen3-VL-2B
0.917
Qwen3-VL-4B
0.929
Qwen3-VL-4B (ViT-only adapted)
0.935
Table 4: Ablation results on the internal driving license recognition task. Performance is measured by Document-level Accuracy.
Model
SODBwo\mbox−formula↑
SODB↑
Etext↓
CDMformula↑
TEDStable↑
TEDS\mbox−Stable↑
Eorder↓
Qwen3-VL-4B
86.90
86.78
0.055
86.55
79.29
84.26
0.084
FinixDoc-VL
92.65
87.40
0.042
76.89
89.50
92.76
0.055
Specialized Models
DeepSeek-OCR-2
90.35
89.17
0.049
86.85
85.60
90.06
0.060
Dots.OCR
90.99
88.41
0.048
83.22
86.78
90.62
0.053
MinerU 2.5
91.97
90.93
0.045
88.86
88.44
92.42
0.044
Table 5: Performance on OmniDocBench v1.5.
Model
SFDB↑
Etext↓
TEDStable↑
TEDS\mbox−Stable↑
Eorder↓
Qwen3-VL-4B
80.18
0.145
76.04
81.23
0.210
FinixDoc-VL
93.19
0.039
92.07
93.67
0.086
Specialized Models
DeepSeek-OCR-2
82.80
0.139
90.00
92.32
0.277
FireRed-OCR
83.10
0.119
87.10
89.14
0.259
PaddleOCR-VL-1.5
85.41
0.116
86.12
88.26
0.183
Table 6: Performance on FinixDigital.
Model
SFDB↑
Etext↓
TEDStable↑
TEDS\mbox−Stable↑
Eorder↓
Qwen3-VL-4B
54.28
0.408
50.13
63.04
0.465
FinixDoc-VL
67.03
0.276
69.08
77.69
0.404
Specialized Models
PaddleOCR-VL-1.5
41.28
0.512
34.54
46.72
0.595
MinerU 2.5
43.08
0.513
35.54
48.58
0.550
DeepSeek-OCR-2
43.20
0.459
30.69
43.52
0.552
Table 7: Performance on FinixPhoto.
Model
Success Rate ↑
SFDB↑
Etext↓
TEDStable↑
TEDS\mbox−Stable↑
Eorder↓
FinixDoc
0.92
68.23
0.357
57.09
60.10
0.167
Qwen3-VL-235B-A22B-Instruct
0.68
34.85
0.847
47.05
63.20
0.578
GLM-OCR
0.34
38.06
0.816
59.39
62.43
0.636
Table 8: Performance on FinixHuge. FinixDoc denotes the complete system, including the Document Preprocessing Tool Layer’s split-then-merge (tiling-and-merging) module.
Model
Overall
Claim Medical Records
Expense Statements
Expense Receipts
Medical Examination Reports
Identity Documents
Qwen3-VL-4B
66.42
50.24
64.85
70.55
60.21
86.26
FinixDoc-VL
84.08
85.42
80.79
85.15
78.38
90.64
Specialized Models
DeepSeek-OCR-2
52.77
42.40
54.37
66.92
53.64
46.51
FireRed-OCR
64.93
50.23
62.67
66.24
69.45
76.08
PaddleOCR-VL-1.5
61.06
45.17
65.19
62.86
56.30
75.79
Table 9: Performance on FinixInner.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Internal pipeline of the split-then-merge strategy in the Document Preprocessing Tool Layer. The strategy follows a three-stage process: Dynamic Routing, Structure-Aware Segmentation, and Reconstruction. Depending on token feasibility and structural complexity, an input page is processed either by direct VLM parsing or by a split-then-merge path for ultra-large-page handling.
Target Character
Example Homoglyphs
Target Character
Example Homoglyphs
0
O, 8, 6, 9, D
维
唯, 淮, 推, 准
O
0, Q, D
清
靖, 情, 请, 洁
I
l, 1, | , !
请
情, 清, 谱, 讲, 证
l
I, i, | , !
组
细, 织, 绪, 级, 纸
|
I, l, !, 丨
痛
疼, 病, 痹, 痪
1
I, l, 7, !
靖
情, 晴, 清
Appendix
Table 10: Representative examples from the financial homoglyph vocabulary.
Stroke Count
Example Characters
1
一, 丨, 丶, 丿, 乙
2
丁, 七, 乃, 九, 了, 二, 乜
3
万, 丈, 三, 上, 下, 久, 义, 亡, 个, 丫
4
不, 丑, 专, 中, 丰, 乌, 书, 云, 互, 五
5
且, 世, 丘, 业, 东, 丝, 乎, 乏, 乐, 主
6
丞, 丢, 两, 争, 亘, 亚, 交, 亥, 亦, 产
Appendix
Table 11: Representative examples from the stroke-count lexicon.
Figure 9: FinixDigital Sample 1. Comparison of parsing results for an insurance clause page containing a two-column disease list table and footnote references. FinixDoc correctly preserves the two-column table structure for the 15 infectious diseases, maintains proper heading hierarchy ( ## , ### ), and retains superscript footnote references ( 11 , 12 ). In contrast, MinerU 2.5 collapses the table into a flat list, fails to encode the heading hierarchy, merges neighboring text around the disease-list introduction, and introduces OCR errors. Qwen3-VL-235B-A22B-Instruct preserves the table layout more completely than MinerU 2.5, but still fails to encode the heading hierarchy, changes domain terms in the disease-list entries (e.g., 特定传染疾病定义 becomes 特定传染病定义 ), and introduces a footnote-level OCR error (e.g., 入住医院 becomes 入往医院 ).
Figure 10: FinixDigital Sample 2. Comparison of parsing results for an insurance product specification page containing a financial illustration table and section headings. FinixDoc preserves the numeric table grid, reconstructs the heading hierarchy ( ### 3 , #### 3.1 , #### 3.2 ), and maintains consistent list formatting with (1) and (2) numbering. MinerU 2.5 fails to render the circled section marker and subsection hierarchy as Markdown headings and mixes bullet styles in the notes. Qwen3-VL-235B-A22B-Instruct misinterprets the first numeric data row as a table header and introduces an HTML/dataframe-style table with an extra index-like column.
Figure 11: FinixDigital Sample 3. Comparison of parsing results for an insurance contract page with nested headings, a structured table, and footnote references. FinixDoc accurately reconstructs the heading hierarchy ( ### 3.3 , ### 3.4 , ## 4 , ### 4.1 – ### 4.4 ), preserves the table with proper headers, and retains superscript footnote references ( 12 , 13 ). MinerU 2.5 flattens the heading hierarchy, loses or inconsistently renders superscript footnotes (e.g., 12 becomes plain text), and degrades the page-ending footnote area. Qwen3-VL-235B-A22B-Instruct preserves the table but inserts spurious spacing between Chinese characters, appends extraneous AI commentary after the parsed page, and shows inconsistent heading formatting.
Figure 12: FinixPhoto Sample 1. Per-field comparison on a Shanghai inpatient receipt captured under challenging conditions. Four fields are highlighted: the fee-detail table, the total amount ( 51206.15 ), the medical-insurance payout ( 23069.33 ), and the payee-unit number ( 1600206215 ). FinixDoc parses all four correctly. Youtu-Parsing reads the total as 161206.15 , the insurance payout as 230893.33 , drops the payee-unit number, and produces a heavily garbled fee table. Kimi-K2.5 gets the numeric fields right but its fee table is missing most of the 14 items present in the original.
Figure 13: FinixPhoto Sample 2. Per-field comparison on a Chongqing outpatient medical receipt. Four fields are highlighted: the invoice number ( 0559817 ), the Chinese-capital total (肆佰叁拾叁圆捌角整), the payee (龙邱渝), and a synthetic “back-page area” box at the bottom strip of the receipt. FinixDoc gets the invoice number, the Chinese total, and the payee correct, and crucially does not hallucinate any content beyond what is visible on the front of the receipt. Youtu-Parsing renders the Chinese total as a plain “ 433.80 ” and drops the payee. Both Youtu-Parsing and Kimi-K2.5 hallucinate extra back-page items (e.g., CT, 重症监护, 中医治疗) that are not visible in the input image.
Figure 14: FinixPhoto Sample 3. Per-field comparison on a Hubei outpatient medical invoice. Four fields are highlighted: the receipt code ( 1420601 ), the unified social-credit code ( 00000507 ), the issue date ( 2020-08-12 09:52:59 ), and the cashier code ( 1890266 ). FinixDoc gets all four correct. Youtu-Parsing truncates the receipt code to 20601 , prefixes the social-credit code with an extraneous 123 , drops the issue date, and reads the cashier code as 1890260 . Kimi-K2.5 reads 0000507 (one leading zero missing) for the social-credit code and 2020 06 12 09 52 59 (wrong month, lost separators) for the issue date.
In this paper, we propose LingDT-VL-OCR, a document parsing system tailored to financial-domain documents, transforming ultra-long financial PDFs into semantically consistent, highly accurate, structured outputs with auditing-grade provenance. To address finance-specific challenges such as complex layouts, cross-page structural discontinuities, and cell-level referencing capability, LingDT-VL-OCR combines (1) a Cross-page Contents Consolidation algorithm to restore continuity across pages and a Document-level Heading Hierarchy Reconstruction (DHR) module to build a globally consistent Table of Contents (TOC) tree for structure-aware retrieval, and (2) a difficulty-adaptive curriculum learning training strategy for table parsing, together with a CellBBoxRegressor module that uses structural anchor tokens to localize table cells from decoder hidden states without external detectors. Experiments demonstrate that our model shows high performance on the Overall metric of OmniDocBench. To enable realistic evaluation in the financial vertical, we further introduce FinDocBench, a benchmark that includes six financial document categories with expert-verified annotations and evaluation metrics including Table of Contents edit-distance-based similarity (TocEDS), cross-page concatenated TEDS, and Table Cell Intersection over Union (C-IoU). We evaluate a wide range of state-of-the-art models on FinDocBench to assess their capabilities and remaining limitations on financial documents. Overall, LingDT-VL-OCR and FinDocBench provide a practical foundation for reliable downstream financial document applications.
AI agents are changing the requirements for document parsing. What matters is semantic correctness: parsed output must preserve the structure and meaning needed for autonomous decisions, including correct table structure, precise chart data, semantically meaningful formatting, and visual grounding. Existing benchmarks do not fully capture this setting for enterprise automation, relying on narrow document distributions and text-similarity metrics that miss agent-critical failures. We introduce ParseBench, a benchmark of ∼2,000 human-verified pages from enterprise documents spanning insurance, finance, and government, organized around five capability dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. Across 14 methods spanning vision-language models, specialized document parsers, and LlamaParse, the benchmark reveals a fragmented capability landscape: no method is consistently strong across all five dimensions. LlamaParse Agentic achieves the highest overall score at 84.9%, and the benchmark highlights the remaining capability gaps across current systems. Dataset and evaluation code are available on https://huggingface.co/datasets/llamaindex/ParseBench and https://github.com/run-llama/ParseBench.
Boyang Zhang, Sebastián G. Acosta, Preston Carlson +4
Document parsing models are improving rapidly, but existing benchmarks are still insufficient for a comprehensive evaluation of their capabilities. We introduce PureDocBench, a benchmark with 10 domains, 66 subcategories, and 1,475 pages. Each page has matched Clean, Digital, and Real versions, covering clean renders, simulated degradation, and real capture or sharing conditions. The resulting 4,425 images have source-derived annotations for text, formulas, tables, and reading order, checked through automated validation and human review. Keeping page content and annotations fixed allows direct comparisons across image conditions. We evaluate 58 models, including pipeline-based parsers, end-to-end parsers, and general vision-language models. The best three-track average score is 79.69 out of 100, and the mean across models is 63.84. Scores drop by 3.34 points on Digital and 11.61 points on Real on average, relative to Clean. TeleOCR leads on Clean, GLM-5.3-Flash on Digital, and Gemini 3.6 Flash on Real. The benchmark dataset, source documents, generation tools, and evaluation code are publicly available at https://github.com/zhihengli-casia/PureDocBench.
Zhiheng Li, Zongyang Ma, Jiaxian Chen +12
CASIA · UCAS · Beijing Key Laboratory of Super Intelligent Security of Multi-Modal Information +6