💬 NLP

$G^2$ -Reader: Dual Evolving Graphs for Multimodal Document QA

$G^2$ -Reader addresses the brittleness of retrieval-augmented generation in multimodal long-document QA by introducing a dual-graph system that evolves a Content Graph to preserve native document structure and a Planning Graph to guide iterative evidence gathering, achieving state-of-the-art performance on VisDoMBench.

Original authors: Yaxin Du, Junru Song, Yifan Zhou, Cheng Wang, Jiahao Gu, Zimeng Chen, Menglan Chen, Wen Yao, Yang Yang, Ying Wen, Siheng Chen

Published 2026-01-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yaxin Du, Junru Song, Yifan Zhou, Cheng Wang, Jiahao Gu, Zimeng Chen, Menglan Chen, Wen Yao, Yang Yang, Ying Wen, Siheng Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a massive, 500-page mystery novel that is written in a mix of text, charts, photos, and handwritten notes. You are an expert detective (the AI), but you have a very short memory: you can only hold about 10 pages in your head at once.

This is the problem the paper calls Multimodal Document QA (Question Answering). Current AI systems try to solve this by chopping the book into tiny, flat pieces (like cutting a puzzle into individual squares) and searching for the right pieces. The problem is, this destroys the picture. A chart gets separated from its caption; a graph gets separated from the paragraph explaining it. The AI ends up with "semantic fragments"—pieces of a puzzle that don't make sense on their own.

Furthermore, if the AI asks a question and gets a partial answer, it often gets stuck in a loop, searching the same irrelevant pages over and over, or drifting off into unrelated sections because it has no "big picture" map of where it has been.

G2-Reader is a new system designed to fix this. The authors propose a solution using two evolving graphs (think of them as two different types of maps) that work together.

1. The Content Graph: The "Living Encyclopedia"

Instead of cutting the document into flat pieces, G2-Reader builds a Content Graph.

The Analogy: Imagine the document isn't a stack of paper, but a city.
- Nodes (The Buildings): Each paragraph, table, or figure is a building.
- Edges (The Roads): The connections between them are roads. A caption is a road connecting to a picture; a reference is a road connecting to a previous paragraph.
The Magic (Evolution): In old systems, these roads are just based on physical proximity (what's next to what). In G2-Reader, the system acts like a city planner who keeps walking the streets. It asks, "Does this building actually relate to that one?"
- If a chart in Chapter 1 explains a graph in Chapter 5, the planner builds a new, invisible road connecting them.
- It updates the "signs" on the buildings (the summaries) to include context from their neighbors.
- Result: The AI doesn't just see a floating chart; it sees a chart connected to the story it tells. This preserves the "native structure" of the document.

2. The Planning Graph: The "Agentic Detective's Notebook"

Once the city map is built, the AI needs to solve the specific question. This is where the Planning Graph comes in.

The Analogy: Imagine a detective trying to solve a complex case. Instead of just guessing, they write down a flowchart of sub-questions in a notebook.
- The Root: The main question ("Who stole the cookie?").
- The Branches: Smaller questions ("Did the butler enter the room?" "Was the cookie jar open?").
The Magic (Dynamic Replanning):
- The AI acts as an agent. It follows the flowchart, checking the "Content Graph" (the city) for evidence for each sub-question.
- The "Evidence Checker": After gathering clues, a special part of the system acts like a quality control inspector. It looks at the notebook and asks, "Do we have enough proof to solve the main case?"
- The "Replan": If the inspector says, "No, we are missing data about the butler's alibi," the system doesn't just guess. It rewrites the flowchart. It adds a new branch to the notebook specifically to find that missing alibi.
- Result: The AI never gets stuck in a loop. It has a persistent memory of what it knows and what it still needs to find, guiding it step-by-step to the answer.

How They Work Together

Think of the Content Graph as the Library (where the books are organized perfectly with cross-references) and the Planning Graph as the Librarian's Strategy (a step-by-step plan to find the right books).

The Librarian (Planning Graph) breaks the big question into small tasks.
For each task, they go to the Library (Content Graph). Because the Library is organized with "living" connections, they find the exact cluster of related text, tables, and images instantly.
If the Librarian realizes they missed a clue, they don't just wander aimlessly; they update their strategy map and go back to the Library to find the specific missing piece.

The Results

The authors tested this system on a benchmark called VisDoMBench, which involves tricky questions about slides, scientific papers, and tables.

They used an open-source AI model (Qwen3-VL-32B) as the "brain."
The Outcome: G2-Reader achieved 66.21% accuracy.
The Comparison: This beat the best open-source models trying to do it alone (29.90%) and even beat a massive, proprietary "super-model" (GPT-5) which only scored 53.08%.

The Takeaway

The paper claims that for complex, multi-page documents with pictures and tables, structure is more important than raw power. By giving the AI a structured map of the document (Content Graph) and a dynamic plan for how to search it (Planning Graph), a smaller, open-source AI can outperform much larger, "dumber" systems that just try to read everything at once.

The paper does not claim this is ready for medical diagnosis or legal courtrooms yet; it simply proves that this "dual-graph" architecture is a superior way to read and reason over complex documents.

Technical Summary: G2-Reader: Dual Evolving Graphs for Multimodal Document QA

Problem Statement

Retrieval-Augmented Generation (RAG) has emerged as a dominant paradigm for question answering over long documents. However, current RAG systems face significant limitations when applied to multimodal long documents (e.g., academic papers, slide decks, textbooks) where text, tables, and figures are interleaved across many pages. The paper identifies two coupled crises in existing approaches:

The Representation Challenge: Standard RAG pipelines flatten documents into independent text chunks. This "flat chunking" breaks the document's native structure and cross-modal alignment (e.g., separating a figure from its caption or referring text). Consequently, retrieved contexts become semantic fragments that are difficult to interpret in isolation.
The Retrieval Challenge: Traditional RAG is often one-shot and brittle in long contexts. While iterative retrieval was introduced to refine queries step-by-step, it frequently fails in long multimodal settings by looping on partial evidence or drifting into irrelevant sections as noise accumulates. This occurs because each step is guided only by the current snippet without a persistent global search state.

Methodology: G2-Reader

To address these issues, the authors propose G2-Reader, a dual-graph system that advances both evidence representation and retrieval-reasoning. The system operates through two tightly coupled, evolving graphs:

1. Content Graph ( $G_C$ ): Structured Multimodal Representation

Instead of flat chunking, G2-Reader constructs a heterogeneous multimodal Content Graph to preserve document-native layouts and cross-modal semantics.

Construction: Documents are parsed into atomic units (paragraphs, tables, figures with captions) forming the node set $V$ . Edges are initialized based on reading order (sliding windows).
Joint Evolution: The graph undergoes an iterative evolution process ( $T$ $T$ steps) driven by a Vision-Language Model (VLM). In each step, the VLM performs a joint update of node attributes (summaries and keywords) and graph topology (edges).
- Attribute Update: The VLM enriches node summaries with context from neighbors, resolves implicit references, and emphasizes distinguishing features.
- Topology Update: The VLM identifies meaningful logical links (e.g., explanation, causality, cross-modal grounding) beyond simple similarity, creating a sparse set of substantive dependencies.
Retrieval: Evidence is retrieved via a structure-aware readout operator. Given a query, nodes are ranked by similarity, and the top nodes are expanded to include their immediate neighbors in the graph, ensuring context-aware evidence assembly.

2. Planning Graph ( $G_P$ ): Agentic Reasoning State

The Planning Graph is a Directed Acyclic Graph (DAG) that explicitly anchors the reasoning process, serving as a dynamic state tracker for evidence assembly.

Initialization: A decomposer agent breaks the user query into an initial DAG of atomic sub-questions, encoding dependency relations.
Execution & Refinement: The system executes sub-questions in topological order. For each sub-question, a worker agent retrieves a local evidence subgraph from $G_C$ and generates an intermediate answer.
Verification & Replanning: An evidence checker evaluates whether the accumulated intermediate answers are sufficient to answer the original query.
- If sufficient, the process terminates.
- If insufficient, the checker identifies specific gaps (missing data, ambiguities). A decomposer then refines the Planning Graph by adding, updating, or removing nodes to target these gaps.
Synthesis: Once sufficiency is met, a reasoner agent synthesizes the final answer based on the original query and the verified, structured evidence set.

Key Contributions

The paper claims three primary contributions:

Problem Identification: It formally identifies the coupled limitations of fragmented evidence representation and unstable long-context retrieval in current multimodal RAG systems.
Dual-Graph Architecture: It introduces the Content Graph to preserve semantic structures within multimodal documents and the Planning Graph to maintain a reasoning state that guides stepwise navigation for evidence completion.
Empirical Superiority: On the VisDoMBench benchmark across five multimodal domains (slides, web pages, academic papers, textbooks, and scientific graphs), G2-Reader achieves state-of-the-art performance.

Experimental Results

The system was evaluated on VisDoMBench using the open-source Qwen3-VL-32B-Instruct as the backbone.

Overall Performance: G2-Reader achieved an average accuracy of 66.21%.
Comparison with Baselines:
- It significantly outperformed the proprietary frontier model GPT-5 (53.08%) by a margin of 13.13 percentage points.
- It surpassed strong open-source baselines, including VisDoMRAG (65.01%) and specialized graph-based systems like GraphRAG (50.86%).
- It demonstrated a 121% relative improvement over the base Qwen3-VL-32B model used in a zero-shot setting (29.90%).
Ablation Studies:
- Multimodality: Including visual evidence improved average accuracy by 8.4%.
- Component Synergy: The combination of the Content Graph and Planning Graph yielded a performance gain (+4.8%) that exceeded the sum of their individual contributions, indicating strong complementarity.
- Iterative Refinement: Moderate refinement of the Planning Graph (up to 3 iterations) improved performance, while over-refinement led to diminishing returns.
- Graph Evolution: Iterative evolution of the Content Graph (up to 3 rounds) improved node representations, with performance dropping at higher iterations due to over-smoothing.

Significance and Claims

The paper posits that G2-Reader validates the hypothesis that structured architecture can empower open-source models to outperform powerful closed-source frontiers. By providing a structured foundation for complex evidence assembly, the dual-graph approach effectively compensates for the long-context reasoning limitations of medium-sized models.

The authors emphasize that their method shifts the paradigm from "documents-as-text" to "documents-as-knowledge," turning visually rich documents into explicit, reusable external knowledge structures. This approach facilitates more controllable knowledge management and robust reasoning over multimodal long documents, addressing the brittleness of current RAG systems in real-world scenarios. The work suggests that sophisticated evidence organization and iterative planning are more effective than raw model scaling or proprietary training for complex multimodal reasoning tasks.