Graph Counselor: Adaptive Graph Exploration via Multi-Agent Synergy to Enhance LLM Reasoning

Graph Counselor is a multi-agent GraphRAG framework that uses collaborative Planning, Thought, and Execution agents with adaptive graph information extraction and self-reflection to dynamically improve LLM reasoning over complex graph-structured knowledge. It addresses the rigidity of single-agent, fixed-pattern graph retrieval systems by enabling adaptive reasoning depth and semantic correction.

Problem Statement

Existing GraphRAG methods rely on a single agent with fixed iterative patterns, making them unable to adaptively capture multi-level textual, structural, and degree information in graph data. Additionally, preset reasoning schemes prevent dynamic adjustment of reasoning depth and precise semantic correction, limiting factual accuracy in specialized domains. These limitations reduce LLM performance on complex graph reasoning tasks requiring multi-hop inference and nuanced knowledge integration.

Key Novelty

Adaptive Graph Information Extraction Module (AGIEM): A three-agent collaborative system (Planning, Thought, Execution) that dynamically models complex graph structures and adjusts multi-level information extraction strategies at runtime
Self-Reflection with Multiple Perspectives (SR) module: Combines self-reflection with backward reasoning to correct semantic inconsistencies and improve accuracy of final reasoning outputs
Multi-agent synergy for GraphRAG: Replaces monolithic single-agent graph traversal with a specialized, role-separated multi-agent pipeline that adapts reasoning depth to query complexity

Evaluation Highlights

Graph Counselor outperforms existing GraphRAG baselines across multiple graph reasoning benchmarks, demonstrating higher reasoning accuracy on tasks requiring multi-level dependency modeling
Improved generalization ability across diverse graph reasoning tasks compared to fixed-pattern single-agent methods, validating the adaptive extraction and reflection mechanisms

Breakthrough Assessment

6/10 Graph Counselor presents a solid and practically motivated contribution by combining multi-agent collaboration with adaptive graph exploration and self-reflection for GraphRAG. However, the core ideas (multi-agent LLM pipelines, self-reflection, RAG over graphs) are evolutionary rather than paradigm-shifting, making this a strong applied contribution rather than a foundational advance.

Methodology

Step 1 - Adaptive Graph Information Extraction (AGIEM): The Planning Agent decomposes the input query into sub-goals and determines which graph elements to explore; the Thought Agent reasons about multi-level graph structure (textual, structural, degree); the Execution Agent retrieves and aggregates relevant subgraph information accordingly
Step 2 - Dynamic Reasoning Depth Adjustment: Based on intermediate outputs from the three agents, the system adaptively decides whether further graph traversal is needed, avoiding fixed-depth limitations of prior methods
Step 3 - Self-Reflection with Multiple Perspectives (SR): The system applies backward reasoning over the generated answer, checking semantic consistency and factual alignment with retrieved graph evidence, then refines the output iteratively until convergence

System Components

Planning Agent

Decomposes user queries into structured sub-goals and determines the scope and strategy for graph exploration

Thought Agent

Reasons over multi-level graph properties including textual node/edge content, structural topology, and node degree to guide meaningful information extraction

Execution Agent

Executes the retrieval and aggregation of graph substructures based on the plans and thoughts provided by the other agents

AGIEM (Adaptive Graph Information Extraction Module)

The unified orchestration module combining the three agents to enable dynamic, multi-level graph traversal and information extraction

SR (Self-Reflection with Multiple Perspectives) Module

Post-generation module that uses backward reasoning and multi-perspective evaluation to detect and correct semantic errors in reasoning outputs

Results

Metric/Benchmark	Best Baseline	Graph Counselor	Delta
Graph reasoning accuracy (multi-hop)	Competitive single-agent GraphRAG	Higher accuracy	Positive improvement
Generalization across graph tasks	Fixed-pattern methods	Superior generalization	Qualitatively better
Semantic consistency of outputs	Preset reasoning schemes	Improved via SR module	Reduced semantic errors

Key Takeaways

For GraphRAG applications: Replacing single-agent fixed-depth traversal with a role-specialized multi-agent pipeline (Plan→Think→Execute) is a practical architecture pattern for improving knowledge graph reasoning in LLM systems
For LLM reasoning pipelines: Adding a backward reasoning self-reflection step after generation is a lightweight but effective mechanism to catch and correct semantic inconsistencies, especially in knowledge-intensive tasks
For ML practitioners: The multi-agent decomposition approach generalizes well across graph reasoning benchmarks, suggesting that task decomposition by information type (textual, structural, degree) is a useful inductive bias when designing retrieval systems over heterogeneous graph data

Abstract

Graph Retrieval Augmented Generation (GraphRAG) effectively enhances external knowledge integration capabilities by explicitly modeling knowledge relationships, thereby improving the factual accuracy and generation quality of Large Language Models (LLMs) in specialized domains. However, existing methods suffer from two inherent limitations: 1) Inefficient Information Aggregation: They rely on a single agent and fixed iterative patterns, making it difficult to adaptively capture multi-level textual, structural, and degree information within graph data. 2) Rigid Reasoning Mechanism: They employ preset reasoning schemes, which cannot dynamically adjust reasoning depth nor achieve precise semantic correction. To overcome these limitations, we propose Graph Counselor, an GraphRAG method based on multi-agent collaboration. This method uses the Adaptive Graph Information Extraction Module (AGIEM), where Planning, Thought, and Execution Agents work together to precisely model complex graph structures and dynamically adjust information extraction strategies, addressing the challenges of multi-level dependency modeling and adaptive reasoning depth. Additionally, the Self-Reflection with Multiple Perspectives (SR) module improves the accuracy and semantic consistency of reasoning results through self-reflection and backward reasoning mechanisms. Experiments demonstrate that Graph Counselor outperforms existing methods in multiple graph reasoning tasks, exhibiting higher reasoning accuracy and generalization ability. Our code is available at https://github.com/gjq100/Graph-Counselor.git.