Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal Graphs Meet Thoughts: Enhancing Complex Reasoning in Graph-Augmented LLMs

Hang Luo, Jian Zhang, Chujun Li

TL;DR

This work tackles the challenge of knowledge-intensive reasoning under high-stakes constraints by integrating causal reasoning with graph-based retrieval in a Graph-RAG framework. It introduces CGMT, a causal-first, CoT-aligned retrieval pipeline that filters large knowledge graphs to emphasize cause–effect edges, aligns retrieval steps with the model’s chain-of-thought, and employs multi-stage path processing with re-injection for coherence. Across medical QA datasets MedMCQA and MedQA, CGMT demonstrates consistent accuracy gains across GPT-4o, GPT-4, and GPT-4o-mini, with up to about 10% absolute improvement and clear ablation-supported evidence of each component’s contribution. The findings suggest that prioritizing causal structure and synchronizing retrieval with the model’s reasoning steps yields more interpretable and reliable solutions, with potential for extending to legal and scientific QA domains.

Abstract

In knowledge-intensive tasks, especially in high-stakes domains like medicine and law, it is critical not only to retrieve relevant information but also to provide causal reasoning and explainability. Large language models (LLMs) have achieved remarkable performance in natural language understanding and generation tasks. However, they often suffer from limitations such as difficulty in incorporating new knowledge, generating hallucinations, and explaining their reasoning process. To address these challenges, integrating knowledge graphs with Graph Retrieval-Augmented Generation (Graph RAG) has emerged as an effective solution. Traditional Graph RAG methods often rely on simple graph traversal or semantic similarity, which do not capture causal relationships or align well with the model's internal reasoning steps. This paper proposes a novel pipeline that filters large knowledge graphs to emphasize cause-effect edges, aligns the retrieval process with the model's chain-of-thought (CoT), and enhances reasoning through multi-stage path improvements. Experiments on medical question-answering tasks show consistent gains, with up to a 10\% absolute improvement across multiple large language models (LLMs). This approach demonstrates the value of combining causal reasoning with stepwise retrieval, leading to more interpretable and logically grounded solutions for complex queries.

Causal Graphs Meet Thoughts: Enhancing Complex Reasoning in Graph-Augmented LLMs

TL;DR

This work tackles the challenge of knowledge-intensive reasoning under high-stakes constraints by integrating causal reasoning with graph-based retrieval in a Graph-RAG framework. It introduces CGMT, a causal-first, CoT-aligned retrieval pipeline that filters large knowledge graphs to emphasize cause–effect edges, aligns retrieval steps with the model’s chain-of-thought, and employs multi-stage path processing with re-injection for coherence. Across medical QA datasets MedMCQA and MedQA, CGMT demonstrates consistent accuracy gains across GPT-4o, GPT-4, and GPT-4o-mini, with up to about 10% absolute improvement and clear ablation-supported evidence of each component’s contribution. The findings suggest that prioritizing causal structure and synchronizing retrieval with the model’s reasoning steps yields more interpretable and reliable solutions, with potential for extending to legal and scientific QA domains.

Abstract

In knowledge-intensive tasks, especially in high-stakes domains like medicine and law, it is critical not only to retrieve relevant information but also to provide causal reasoning and explainability. Large language models (LLMs) have achieved remarkable performance in natural language understanding and generation tasks. However, they often suffer from limitations such as difficulty in incorporating new knowledge, generating hallucinations, and explaining their reasoning process. To address these challenges, integrating knowledge graphs with Graph Retrieval-Augmented Generation (Graph RAG) has emerged as an effective solution. Traditional Graph RAG methods often rely on simple graph traversal or semantic similarity, which do not capture causal relationships or align well with the model's internal reasoning steps. This paper proposes a novel pipeline that filters large knowledge graphs to emphasize cause-effect edges, aligns the retrieval process with the model's chain-of-thought (CoT), and enhances reasoning through multi-stage path improvements. Experiments on medical question-answering tasks show consistent gains, with up to a 10\% absolute improvement across multiple large language models (LLMs). This approach demonstrates the value of combining causal reasoning with stepwise retrieval, leading to more interpretable and logically grounded solutions for complex queries.
Paper Structure (43 sections, 10 equations, 3 figures, 3 tables)

This paper contains 43 sections, 10 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Retrieval Challenges in Knowledge-Intensive Queries
  3. Key Limitations in Existing Graph-RAG Approaches
  4. Significance and Motivation:
  5. Contributions and Paper Organization.
  6. Related Works
  7. Retrieval-Augmented Generation and Graph-Based Knowledge Discovery
  8. Causality in Knowledge-Intensive Natural Language Processing
  9. Chain-of-Thought Prompts and Reasoning Transparency
  10. Challenges in Aligning Causality, Graph Retrieval, and CoT Reasoning
  11. Methodology
  12. Overall Framework
  13. Causal Graph Construction and Updating
  14. Filtering the Knowledge Graph for Causality.
  15. Continuous Updates via Causal Mining.
  16. ...and 28 more sections

Figures (3)

  • Figure 1: Overall framework in three stages: (1) building/updating the causal graph, (2) chain-of-thought-driven retrieval, (3) multi-stage path processing.
  • Figure 2: An example of chain-of-thought-driven retrieval. Each CoT segment triggers a partial query to the causal subgraph, producing candidate paths that are later fused.
  • Figure 3: Path scoring and pruning. Each path is evaluated based on CUI overlap, semantic overlap, and length heuristics.