Logics-Parsing-Omni Technical Report

TL;DR

Omni Parsing framework standardizes multimodal data parsing via a unified taxonomy and progressive parsing paradigm.

cs.AI 🔴 Advanced 2026-03-10 29 views
Xin An Jingyi Cai Xiangyang Chen Huayao Liu Peiting Liu Peng Wang Bei Yang Xiuwen Zhu Yongfan Chen Yan Gao Yuan Gao Baoyu Hou Guangzheng Hu Shuzhao Li Weixu Qiao Weidong Ren Yanan Wang Boyu Yang Fan Yang Jiangtao Zhang Lixin Zhang Lin Qu Hu Wei Xiaoxiao Xu Bing Zhao
multimodal parsing unified taxonomy progressive parsing evidence anchoring knowledge transformation

Key Findings

Methodology

The Omni Parsing framework bridges perception and cognition through three hierarchical levels: Holistic Detection, Fine-grained Recognition, and Multi-level Interpreting. Holistic Detection provides spatiotemporal grounding, Fine-grained Recognition performs symbolization and attribute extraction, and Multi-level Interpreting constructs reasoning chains from local semantics to global logic.

Key Results

  • On OmniParsingBench, the Logics-Parsing-Omni model shows consistent improvements across all modalities, demonstrating a strong balance between structural fidelity and semantic interpretation.
  • The model significantly enhances entity-rich reasoning on knowledge-intensive image samples.
  • In the video domain, annotations for fine-grained shot analysis and long educational content are optimized, ensuring accurate temporal localization and narrative understanding.

Significance

This study introduces the Omni Parsing framework, addressing challenges of fragmented task definitions and heterogeneity in unstructured data, providing a method to convert complex audio-visual signals into machine-readable structured knowledge, significantly enhancing model reliability.

Technical Contribution

Proposes a unified parsing framework integrating perception and cognition, providing strict alignment from low-level signals to high-level semantics, enabling evidence-based logical induction.

Novelty

This framework is the first to transform multimodal signals into locatable, enumerable, and traceable standardized knowledge, significantly differing from existing methods.

Limitations

  • The model may encounter performance bottlenecks when handling extremely complex audio-visual signals, particularly in real-time applications.
  • Parsing of unstructured data still requires optimization to improve efficiency.

Future Work

Future research directions include optimizing real-time parsing capabilities, expanding the framework to support more types of unstructured data, and validating its effectiveness in more practical applications.

AI Executive Summary

Multimodal parsing has long faced challenges of fragmented task definitions and heterogeneity in unstructured data. Existing methods often lack layout fidelity and fine-grained grounding when dealing with visually rich documents and long educational videos. To address this, the Omni Parsing framework is proposed, combining perception and cognition through a unified taxonomy and progressive parsing paradigm. This framework includes three levels: Holistic Detection, Fine-grained Recognition, and Multi-level Interpreting, ensuring strict alignment between high-level semantic descriptions and low-level facts.

Experimental results show that the Logics-Parsing-Omni model demonstrates consistent improvements across all modalities on OmniParsingBench, proving the synergy between fine-grained perception and high-level cognition. This framework not only enhances model reliability but also achieves logical induction through an evidence anchoring mechanism, transforming complex audio-visual signals into machine-readable structured knowledge.

However, the framework still faces limitations when handling extremely complex signals, especially in real-time applications. Future research directions include optimizing real-time parsing capabilities, expanding the framework to support more types of unstructured data, and validating its effectiveness in more practical applications.

Deep Analysis

Background

The field of multimodal parsing has seen significant advancements, particularly with the rise of large-scale language models. However, existing methods often lack layout fidelity and fine-grained grounding when dealing with visually rich documents and long educational videos, leading to the loss of critical information and making deep retrieval and automated document conversion challenging.

Core Problem

The core problem in multimodal parsing is integrating fine-grained content parsing with deep semantic understanding without losing semantic value. Existing methods struggle to achieve a unified representation when dealing with complex layouts and diverse non-text visual content.

Innovation

The Omni Parsing framework combines perception and cognition through a unified taxonomy and progressive parsing paradigm. It includes three levels: Holistic Detection, Fine-grained Recognition, and Multi-level Interpreting, ensuring strict alignment between high-level semantic descriptions and low-level facts.

Methodology

  • �� Holistic Detection: Achieves precise spatiotemporal grounding of objects or events to establish a geometric baseline for perception.
  • �� Fine-grained Recognition: Performs symbolization (e.g., OCR/ASR) and attribute extraction on localized objects to complete structured entity parsing.
  • �� Multi-level Interpreting: Constructs a reasoning chain from local semantics to global logic.

Experiments

The experimental design includes evaluations on OmniParsingBench, covering document, image, and audio-video content. The model is trained using knowledge-intensive image samples and optimized video annotations to ensure a balance between structural fidelity and semantic interpretation.

Results

Results show that the Logics-Parsing-Omni model demonstrates consistent improvements across all modalities, particularly enhancing entity-rich reasoning on knowledge-intensive image samples.

Applications

The framework can be applied to downstream tasks such as Retrieval-Augmented Generation, Question Answering, and intelligent tutoring, especially in scenarios requiring fine-grained parsing and deep semantic understanding.

Limitations & Outlook

While the framework performs well in multimodal parsing, it may encounter performance bottlenecks when handling extremely complex audio-visual signals, particularly in real-time applications.

Plain Language Accessible to non-experts

Imagine you're in a library with various books, images, and videos. The Omni Parsing framework acts like a super librarian, not only helping you find the books you need but also summarizing each book's main content, detailing the images, and highlighting key scenes in videos. It organizes all this information into an easy-to-understand format, like translating all the books and videos into a language you can read.

ELI14 Explained like you're 14

Imagine you're playing a super complex game with many levels, each with different tasks and challenges. Omni Parsing is like a super helper that quickly finds key points in each level, tells you how to pass them, and even organizes a guide for you to excel in the game! Isn't that cool?

Glossary

Omni Parsing

A multimodal parsing framework that integrates perception and cognition through a unified taxonomy and progressive parsing paradigm.

Used to transform unstructured signals into locatable, enumerable, and traceable standardized knowledge.

Holistic Detection

Achieves precise spatiotemporal grounding of objects or events to establish a geometric baseline for perception.

First level of the Omni Parsing framework.

Fine-grained Recognition

Performs symbolization and attribute extraction on localized objects to complete structured entity parsing.

Second level of the Omni Parsing framework.

Multi-level Interpreting

Constructs a reasoning chain from local semantics to global logic.

Third level of the Omni Parsing framework.

Evidence Anchoring

Ensures strict alignment between high-level semantic descriptions and low-level facts.

Used to achieve evidence-based logical induction.

Open Questions Unanswered questions from this research

  • 1 How to improve parsing efficiency in real-time applications, especially when handling extremely complex audio-visual signals.
  • 2 How to further optimize parsing of unstructured data to enhance overall model performance.

Applications

Immediate Applications

Intelligent Document Parsing

Transforms complex documents into structured data using the Omni Parsing framework, enhancing information retrieval and automated processing capabilities.

Multimodal Question Answering Systems

Utilizes the framework's fine-grained parsing capabilities to enhance the accuracy and response speed of question answering systems.

Long-term Vision

Fully Automated Content Generation

By further optimizing parsing capabilities, achieve a fully automated process from multimodal data to content generation.

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.

cs.AI