EvoGraph-R1: Self-Evolving Multimodal Knowledge Hypergraphs for Agentic Retrieval

CVPR 2026

Why EvoGraph-R1

From Static Retrieval to Knowledge Evolution

The graph is no longer a frozen index. It becomes an environment that the agent can inspect, extend, and repair while reasoning.

Comparison of RAG-based, graph-based, and EvoGraph-R1 retrieval paradigms on a memorial question.
Motivation. Flat RAG lacks explicit cross-modal alignment and structural modeling, while static GraphRAG cannot repair missing or noisy facts during inference. EvoGraph-R1 closes the loop between retrieval, graph editing, and answer generation.
01

Abstract

02

Method

One Agent, Four Actions, An Evolving Graph

EvoGraph-R1 method overview showing user question, agent policy, action pool, graph edit, domain knowledge extraction, reward modeling, and refined dynamic hypergraph.

Main Results

Stronger Across Text and Multimodal QA

03

Analysis

What Actually Drives the Gain

The ablations show that the multimodal hypergraph, graph edits, and external evidence are all necessary. Graph evolution also produces a more coherent structure, not just a better final score.

Ablation study table and answer quality radar chart for EvoGraph-R1.
Ablations isolate the contribution of each graph-evolution component. The answer-quality radar compares comprehensiveness, factuality, logical coherence, diversity, relevance, correctness, and knowledgeability.
Comparison of reasoning turns across methods.
Reasoning efficiency. Comparison of retrieval and reasoning turns.
Comparison of response length across methods.
Response profile. Answer length across methods and settings.

Graph Refinement

From a Sparse Graph to a Coherent Knowledge State

The subgraph centered on “NARRATIVE” becomes denser and more coherent after refinement, exposing clearer thematic structure and deeper reasoning paths.

Hypergraph centered on narrative before refinement, with sparse and fragmented thematic associations.
Before refinement. Shallow, fragmented associations around “NARRATIVE”.
Hypergraph centered on narrative after refinement, with denser thematic clusters and deeper paths.
After refinement. Denser thematic clusters and deeper reasoning paths.
04

Citation

Cite EvoGraph-R1