Logics-Parsing-Omni Technical Report

Xin An; Jingyi Cai; Xiangyang Chen; Huayao Liu; Peiting Liu; Peng Wang; Bei Yang; Xiuwen Zhu

cs.AIMar 10, 2026

Logics-Parsing-Omni Technical Report

Authors: Xin An, Jingyi Cai, Xiangyang Chen, Huayao Liu, Peiting Liu, Peng Wang, Bei Yang, Xiuwen Zhu,

Organizations: Alibaba Group

Abstract

Addressing the challenges of fragmented task definitions and the heterogeneity of unstructured data in multimodal parsing, this paper proposes the Omni Parsing framework. This framework establishes a Unified Taxonomy covering documents, images, and audio-visual streams, introducing a progressive parsing paradigm that bridges perception and cognition. Specifically, the framework integrates three hierarchical levels: 1) Holistic Detection, which achieves precise spatial-temporal grounding of objects or events to establish a geometric baseline for perception; 2) Fine-grained Recognition, which performs symbolization (e.g., OCR/ASR) and attribute extraction on localized objects to complete structured entity parsing; and 3) Multi-level Interpreting, which constructs a reasoning chain from local semantics to global logic. A pivotal advantage of this framework is its evidence anchoring mechanism, which enforces a strict alignment between high-level semantic descriptions and low-level facts. This enables ``evidence-based'' logical induction, transforming unstructured signals into standardized knowledge that is locatable, enumerable, and traceable. Building on this foundation, we constructed a standardized dataset and released the Logics-Parsing-Omni model, which successfully converts complex audio-visual signals into machine-readable structured knowledge. Experiments demonstrate that fine-grained perception and high-level cognition are synergistic, effectively enhancing model reliability. Furthermore, to quantitatively evaluate these capabilities, we introduce OmniParsingBench. Code, models and the benchmark are released at https://github.com/alibaba/Logics-Parsing/tree/master/Logics-Parsing-Omni.

Figures & tables

Figure 1: OmniParsingBench performance of Logics-Parsing-Omni.

Figure 2: Showcase of the multifaceted capabilities of Logics-Parsing-Omni.

Figure 3: The construction of unified multi-modal parsing corpus and training pipeline of our proposed Logics-Parsing-Omni.

Figure 4: Overview of the Omni Parsing Framework. The framework transforms multimodal raw data into unified structured training data via three progressive stages: (1) L1-Holistic Detection: Performs spatio-temporal grounding and coarse classification; (2) L2-Fine-grained Recognition: Extracts detailed text, symbols, knowledge, attributes, and speech content; (3) L3-Multi-level Interpreting: Synthesizes local semantics with global logical reasoning. The final output is a standardized JSON format containing all parsing results from L1 to L3.

Domain	Data Category	Task	# Samples
Document	General Document	Parsing (Structure)	342K
Document	General Document	Parsing (Structure&Semantic)	333K
Image	Natural Image	Caption (Single-Image)	360K
		Parsing (Single-Image)	411K
		Caption (Diff, Multi-Image)	354K
		Parsing (Diff, Multi-Image)	212K

Table 1: Detailed statistics of the constructed corpus.

Model	Natural Image			Graphics			Document	Audio			Natural Video			Text-Rich Video
Model	Ovr.	Perc.	Cog.	Ovr.	Perc.	Cog.	Perc.	Ovr.	Perc.	Cog.	Ovr.	Perc.	Cog.	Ovr.	Perc.	Cog.
Gemini-3-Pro	61.20	55.96	66.44	87.03	84.21	87.43	87.01	79.40	72.90	85.89	63.40	57.87	68.92	64.37	58.54	70.20
GPT-5.2	39.94	37.77	42.12	82.71	69.86	91.48	77.43	–	–	–	–	–	–	–	–	–
Qwen3.5-397B-A17B	57.40	56.95	57.85	82.81	73.77	83.13	81.09	–	–	–	–	–	–	–	–	–
Qwen3-VL-235B-A22B	58.61	56.23	60.99	79.49	71.51	83.46	84.47	–	–	–	–	–	–	–	–	–
Qwen3-VL-30B-A3B	50.92	48.91	52.94	73.25	65.71	79.36	78.94	–	–	–	–	–	–	–	–	–

Table 2: Performance comparison of various models on OmniParsingBench.

Model	Overall ¶	Perception					Cognition
		Location	Content ∗			Avg.	General	Knowledge	Avg.
		Location	Entity	Knowledge	Text	Avg.	General	Knowledge	Avg.
Gemini-3-Pro 13 13 13 https://aistudio.google.com/models/gemini-3	61.20	74.33	97.35	59.44	70.51	55.96	74.14	58.74	66.44
GPT-5.2 † 9 9 9 https://openai.com/zh-Hans-CN/index/introducing-gpt-5/	39.94	52.91	93.08	–	44.78	37.77	66.46	17.77	42.12
Qwen3.5-397B-A17B 16 16 16 https://github.com/QwenLM/Qwen3.5	57.40	81.98	97.60	51.80	63.56	56.95	66.41	49.28	57.85
Qwen3-VL-235B-A22B Bai et al. (2025a)	58.61	77.99	97.41	58.68	57.89	56.23	68.26	53.71	60.99

Table 3: Evaluation results on the Natural Image module of OmniParsingBench.

Model	Overall	Perception					Cognition
		Location	Content			Avg.	Element	Relation	Reasoning	Avg.
		Location	Element	Relation	Chart	Avg.	Element	Relation	Reasoning	Avg.
Gemini-3-Pro	87.03	78.87	86.21	83.95	87.80	84.21	89.37	89.83	83.11	87.43
GPT-5.2	82.71	45.62	83.30	66.97	83.56	69.86	93.18	87.30	93.95	91.48
Qwen3.5-397B-A17B	82.81	46.48	82.15	74.55	91.91	73.77	86.21	89.22	73.97	83.13
Qwen3-VL-235B-A22B	79.49	52.56	83.20	66.28	83.99	71.51	85.68	82.97	81.74	83.46

Table 4: Evaluation results on the Graphics Module of OmniParsingBench.

Model	Overall	Perception
Model		Text Edit $\downarrow$	Formula CDM	Table TEDs	Table ${}^{TEDS_{s}}$	Reading Order Edit $\downarrow$
Pipeline
MinerU2-pipeline Wang et al. (2024a)		77.93	0.1013	75.89	68.03	76.14	0.2157
MinerU2.5 Niu et al. (2025)	86.30	0.0754	85.07	81.36	84.48	0.1334
PP-StructureV3 Cui et al. (2025c)	70.26	0.1155	78.20	44.12	74.65	0.1891
PaddleOCR-VL Cui et al. (2025a)	85.09	0.0581	83.80	77.27	81.60	0.1404

Table 5: Comprehensive evaluation of document reading on LogicsDocBench.

Model	Overall	Perception				Cognition
Model	Overall	Time	Key	ASR $\downarrow$	Avg.	Info. Ext.	Rec.	Avg.
Gemini-3-Pro	79.4	64.89	76.62	0.2280	72.90	94.29	77.49	85.89
Qwen3-Omni-30B-A3B	75.17	59.12	65.39	0.3813	62.13	95.92	80.52	88.22
Logics-Parsing-Omni(Ours)	79.63	59.81	65.48	0.1747	69.27	94.69	85.28	89.99

Table 6: Evaluation results on the Audio Module of OmniParsingBench

Figure 5: Qualitative examples illustrating the comprehensive audio parsing capability of Logics-Parsing-Omni.

Model	Overall	Perception						Cognition
		Time	Key	ASR $\downarrow$	Visual	Camera	Avg.	Audio			Visual			Audio-Visual		Avg.
		Time	Key	ASR $\downarrow$	Visual	Camera	Avg.	Count	Rec.	Info. Ext.	Attr.& States	Event	Camera	Emot.	Caus. Rsn.	Avg.
Gemini-3-Pro	63.40	79.71	83.17	0.1631	7.99	34.81	57.87	48.29	76.82	46.15	84.38	81.2	41.92	88.95	83.67	68.92
Qwen3-Omni-30B-A3B	45.23	32.01	45.29	0.5114	9.98	34.62	34.15	30.73	57.51	20.98	80.63	64.66	50.38	74.21	71.43	56.32
Logics-Parsing-Omni(Ours)	61.12	60.18	64.85	0.1698	6.82	65.58	56.09	39.51	68.67	43.36	81.88	68.42	68.65	84.74	73.98	66.15

Table 7: Evaluation results on the Natural Video Module of OmniParsingBench, integrating both General Videos and Camera-aware Videos metrics.

Figure 6: Qualitative examples illustrating the comprehensive camera-aware video parsing capability of Logics-Parsing-Omni.

Model	Overall	Perception				Cognition
		Time	Visual	Audio $\downarrow$	Avg.	Audio	Visual				Audio-Visual		In-Depth Str. Cap.	Avg.
		Time	Visual	Audio $\downarrow$	Avg.	Info. Ext.	Object	OCR	Attr.&States	Event	Info. Syn.	Caus. Rsn.	In-Depth Str. Cap.	Avg.
Gemini-3-Pro	64.37	24.29	57.89	0.0655	58.54	61.83	23.64	37.09	18.67	43.85	63.38	85.12	92.75	70.20
Qwen3-Omni-30B-A3B	26.86	4.83	25.84	1.0242	10.22	6.22	1.82	5.63	2.67	6.92	8.92	9.30	81.08	43.50
Logics-Parsing-Omni(Ours)	69.12	32.73	57.83	0.1839	57.39	77.59	57.27	74.18	54.67	69.23	82.16	84.65	90.30	80.85

Table 8: Evaluation results on the Text-Rich Video Module of OmniParsingBench

Model	Overall	Category			Video Duration
Model	Overall	Core Summary	Structured Outline	Deep Content Mining	Short ( $<5$ min.)	Medium (5–15 min.)	Long (15–30 min.)
Gemini-3-Pro	92.75	99.23	93.24	90.64	92.87	93.42	91.59
Qwen3-Omni-30B-A3B	81.08	94.98	82.43	76.25	89.99	82.51	68.57
Logics-Parsing-Omni(Ours)	90.30	97.30	91.51	87.36	93.69	91.56	84.44

Table 9: Detailed evaluation results for the Cognition (In-Depth Str. Cap.) dimension on the Text-Rich Video Module of OmniParsingBench

Model Type	Methods	Overall $\uparrow$	Text Edit $\downarrow$	Formula CDM $\uparrow$	Table TEDS $\uparrow$	Table TEDS-S $\uparrow$	ReadOrder $\downarrow$
Pipeline Tools	Marker-1.8.2 4 4 4 https://github.com/datalab-to/marker	71.3	0.206	76.66	57.88	71.17	0.250
	MinerU2-Pipe Wang et al. (2024a)	75.51	0.209	76.55	70.90	79.11	0.225
	PP-StructureV3 Cui et al. (2025c)	86.73	0.073	85.79	81.68	89.48	0.073
General MLLMs	GPT-4o 10 10 10 https://mistral.ai/news/mistral-ocr	75.02	0.217	79.70	67.07	76.09	0.148
	InternVL3 Zhu et al. (2025)	80.33	0.131	83.42	70.64	77.74	0.113
	InternVL3.5 Wang et al. (2025c)	82.67	0.142	87.23	75.00	81.28	0.125

Table 10: Comparisons with State-of-the-art methods on OmniDocBench-v1.5.

Model	Perception		Cognition
Model	Precision ( $\uparrow$ )	Recall ( $\uparrow$ )	Precision ( $\uparrow$ )	Irre. ( $\downarrow$ )	Recall ( $\uparrow$ )
Gemini-3-Pro	0.417	0.349	0.573	0.331	0.460
GPT-5.2	0.123	0.169	0.671	0.414	0.490
Qwen3.5-397B-A17B	0.383	0.328	0.621	0.376	0.461
Qwen3-VL-235B-A22B	0.356	0.293	0.653	0.339	0.518
Qwen3-VL-30B-A3B	0.242	0.315	0.496	0.402	0.365

Table 11: Evaluation results on the Natural-Image Difference Module of OmniParsingBench

Figure 7: Qualitative examples illustrating the natural image difference parsing capability of Logics-Parsing-Omni.

Model	Overall	Perception							Cognition
		Diff Detection		Diff Identification					Accuracy		Plausibility
		Positioning		Existence			Relation		Object	Operation	Logic	Halluc.
		Coord.	Radius	Point	Line/Arc	Circle	Geo.	Quant.	Object	Operation	Logic	Halluc.
Gemini-3-Pro	27.35	26.45	0.00	32.23	1.05	1.59	4.71	2.71	37.16	33.92	34.74	37.42
GPT-5.2	36.43	7.46	0.47	58.55	9.02	18.98	6.18	5.76	50.16	43.74	46.98	47.94
Qwen3.5-397B-A17B	33.68	11.92	0.87	44.54	2.60	6.54	6.34	4.94	46.84	43.22	44.62	45.88

Table 12: Evaluation results on the Geo-Image Difference Module of OmniParsingBench.

Figure 8: Qualitative examples illustrating the graphic image difference parsing capability of Logics-Parsing-Omni.

Model	Overall	Chart				Geometry
Model	Overall	OCR & Data	Visual Elements	Logical Reasoning	Avg.	Basic Elements	Geometric Relations	Quantitative Relations	Avg.
Qwen3-Omni-30B-A3B	79.55	82.33	86.51	73.97	81.40	80.16	71.67	80.41	77.70
+ Graphics (Caption Only)	79.70	81.80	87.91	68.04	80.10	84.10	70.65	75.26	79.30
+ Graphics (Parsing & Caption)	91.80	97.17	95.81	90.87	95.50	90.16	82.94	90.72	88.10

Table 13: Ablation results on the Graphics Cognition.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Entity			Knowledge-Aware Entity			Text				Average
Model	Loc. Recall	Sem. Acc. ∗	Score	Loc. Recall	Sem. Acc. ∗	Score	Loc. Recall	OCR Acc. ∗	Lang. Acc. ∗	Score	Average
Gemini 3 Pro	75.90	97.35	73.89	88.01	59.44	52.32	59.09	74.49	88.46	41.67	55.96
GPT 5.2 †	58.30	93.08	54.26	-	-	-	47.52	49.25	86.57	21.28	37.77
Qwen3.5-397B-A17B	73.54	97.60	71.77	89.30	51.80	46.26	83.10	74.58	84.70	52.82	56.95
Qwen3-VL-235B-A22B-Instruct	68.56	97.41	66.79	89.30	58.68	52.41	76.11	69.30	85.96	49.50	56.23
Qwen3-VL-30B-A3B-Instruct	63.08	97.65	61.60	83.15	44.77	37.23	88.03	64.80	83.20	47.89	48.91

Appendix

Table 14: Detailed Evaluation Results on the Natural Image module of OmniParsingBench - Perception

Model	Overall Metrics			Caption Length	Detailed Evaluation Metrics
Model	Overall	Answerable	Precision	Caption Length	Theme	Emotion	Scene	Count	Attribute	Location	Text	Style	Composition
Direct Dense Caption
Gemini 3 Pro	86.38	93.40	92.48	1,613	95.43	86.27	87.54	80.55	90.06	83.87	90.16	84.72	77.55
GPT 5.2	73.52	81.83	89.86	1,726	89.12	49.02	79.44	63.64	75.22	73.94	59.39	78.12	68.54
Qwen3.5-397B-A17B	82.48	90.32	91.33	1,060	95.23	88.24	86.20	71.50	86.72	80.07	79.19	80.06	74.10
Qwen3-VL-235B-A22B-Instruct	79.86	88.75	89.98	1,234	94.40	89.22	86.76	68.47	82.94	75.44	80.71	78.67	69.82

Appendix

Table 15: Detailed Evaluation Results on the Natural Image module of OmniParsingBench - Cognition.

Model	Overall	Perception			Cognition
Model	Overall	Stat. Charts (F1)	Struct. Diagrams (Sim.)	Avg.	OCR & Data	Visual Elements	Logical Reasoning	Avg. Acc.
Gemini-3-Pro 13 13 footnotemark: 13	88.00	87.43	89.60	87.80	89.22	90.70	83.11	88.20
GPT-5.2 9 9 footnotemark: 9	88.28	79.63	87.49	83.56	94.88	87.21	93.95	93.00
Qwen3.5-397B-A17B 16 16 footnotemark: 16	87.41	92.28	90.12	91.91	83.57	90.23	73.97	82.90
Qwen3-VL-235B-A22B Bai et al. (2025a)	86.05	84.51	81.46	83.99	89.22	91.63	81.74	88.10
Qwen3-VL-30B-A3B Bai et al. (2025a)	75.90	63.22	87.76	67.40	86.75	88.37	74.43	84.40

Appendix

Table 16: Detailed Evaluation Results on the Graphics Module of OmniParsingBench - Chart.

Model	Overall	Perception								Cognition
		Positioning		Existence			Relation		Avg.	Basic Elements	Geometric Relations	Quantitative Relations	Avg.
		Coord.	Radius	Point	Line/Arc	Circle	Geo.	Quant.	Avg.	Basic Elements	Geometric Relations	Quantitative Relations	Avg.
Gemini-3-Pro 13 13 footnotemark: 13	86.05	72.53	85.20	92.63	69.04	96.97	77.88	90.02	83.10	89.51	87.03	91.75	89.00
GPT-5.2 9 9 footnotemark: 9	77.13	5.26	85.98	89.63	62.93	97.33	58.45	75.49	65.46	91.48	81.91	92.78	88.80
Qwen3.5-397B-A17B 16 16 footnotemark: 16	78.21	7.30	85.66	88.87	61.17	96.40	68.95	80.15	68.41	88.85	84.64	92.78	88.00
Qwen3-VL-235B-A22B Bai et al. (2025a)	72.92	16.75	88.37	89.89	62.91	96.80	50.77	81.79	67.24	82.13	67.58	89.69	78.60

Appendix

Table 17: Detailed evaluation results on the Graphics Module of OmniParsingBench - Geometrics.

Model	Overall	Perception			Cognition
Model	Overall	Joint Acc.	L1 Acc.	L2 Acc.	Acc.
Gemini-3-Pro	38.37	34.81	65.19	34.81	41.92
Qwen3-Omni-30B-A3B	42.50	34.62	63.08	36.54	50.38
Logics-Parsing-Omni	67.12	65.58	88.65	65.58	68.65

Appendix

Table 18: Detailed evaluation results on the Natural Videos Module of OmniParsingBench - Camera.

Panel of Figure 9 — Figure 9: Qualitative examples illustrating the natural image parsing capability of Logics-Parsing-Omni.

Panel of Figure 10 — Figure 10: Qualitative examples illustrating the graphics image parsing capability of Logics-Parsing-Omni.

Figure 11: Qualitative examples illustrating the graphics image difference parsing capability of Logics-Parsing-Omni.

Figure 12: Qualitative examples illustrating the natural image difference parsing capability of Logics-Parsing-Omni.

Figure 13: Qualitative examples illustrating the natural video parsing capability of Logics-Parsing-Omni.

Figure 14: Qualitative example illustrating the comprehensive text-rich video parsing capability of Logics-Parsing-Omni.

Figure 15: Qualitative examples illustrating the text-rich video in-depth caption capability of Logics-Parsing-Omni.

Figure 16: Qualitative example illustrating the comprehensive document structure and semantic parsing capability of Logics-Parsing-Omni on the geometric illustration.

Figure 17: Qualitative example illustrating the comprehensive document structure and semantic parsing capability of Logics-Parsing-Omni on the graphics and natrual illustrations.

Multimodal Understanding

Multimodal understanding studies how AI systems interpret and reason across text, images, video, audio, and other signals.

See all 479 papers

Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training
Mar 25, 2026Gengluo Li, Pengyuan Lyu, Chengquan Zhang +7Document Parsing Multimodal Large Language Models
Institute of Information Engineering, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Tencent +2
SceneParser: Hierarchical Scene Parsing for Visual Semantics Understanding
May 14, 2026Pengxin Xu, Xincheng Lin, Luping Xiao +5Fine-Grained Visual Perception Multimodal Understanding
Harbin Institute of Technology, Shenzhen · Zhongguancun Academy · Huazhong University of Science and Technology +4
Multimodal OCR: Parse Anything from Documents
Mar 13, 2026Handong Zheng, Yumeng Li, Kaile Zhang +22Document Parsing Document Image Understanding
Huazhong University of Science and Technology · hi lab, Xiaohongshu Inc

Mar 25, 2026J/K move · Enter open · S save

Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training

Gengluo Li, Pengyuan Lyu, Chengquan Zhang, Huawen Shen+6

Institute of Information Engineering, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Tencent+2

Document Parsing Multimodal Large Language Models Synthetic Data Generation

Open paper AskPDF arXiv

Document parsing has recently advanced with multimodal large language models (MLLMs) that directly map document images to structured outputs. Traditional cascaded pipelines depend on precise layout analysis and often fail under casually captured or non-standard conditions. Although end-to-end approaches mitigate this dependency, they still exhibit repetitive, hallucinated, and structurally inconsistent predictions - primarily due to the scarcity of large-scale, high-quality full-page (document-level) end-to-end parsing data and the lack of structure-aware training strategies. To address these challenges, we propose a data-training co-design framework for robust end-to-end document parsing. A Realistic Scene Synthesis strategy constructs large-scale, structurally diverse full-page end-to-end supervision by composing layout templates with rich document elements, while a Document-Aware Training Recipe introduces progressive learning and structure-token optimization to enhance structural fidelity and decoding stability. We further build Wild-OmniDocBench, a benchmark derived from real-world captured documents for robustness evaluation. Integrated into a 1B-parameter MLLM, our method achieves superior accuracy and robustness across both scanned/digital and real-world captured scenarios. All models, data synthesis pipelines, and benchmarks will be publicly released to advance future research in document understanding.

Abstract

Multimodal Understanding

Explore similar work

Towards Real-World Document Parsing via Realistic Scene Synthesis and Document-Aware Training

SceneParser: Hierarchical Scene Parsing for Visual Semantics Understanding

Multimodal OCR: Parse Anything from Documents