Abstract
Large language models (LLMs) encounter challenges such as hallucination and factual errors in knowledge-intensive tasks. On the one hand, LLMs sometimes struggle to generate reliable answers based on the black-box parametric knowledge, due to the lack of responsible knowledge. Moreover, fragmented knowledge facts extracted by knowledge retrievers fail to provide explicit and coherent reasoning paths for improving LLM reasoning. To address these challenges, we propose KG-CoT, a novel knowledge-augmented paradigm that leverages a small-scale step-by-step graph reasoning model to reason over knowledge graphs (KGs) and utilizes a reasoning path generation method to generate chains of reasoning with high confidence for large-scale LLMs. Extensive experiments demonstrate that our KG-CoT significantly improves the performance of LLMs on knowledge-intensive question answering tasks, such as multi-hop, single-hop, and open-domain question answering benchmarks, without fine-tuning LLMs. KG-CoT outperforms the CoT prompting as well as prior retrieval-augmented and knowledge base question answering baselines. Moreover, KG-CoT can reduce the number of API calls and cost and generalize to various LLM backbones in a lightweight plug-and-play manner.
Introduction
Large Language Models (LLMs) like GPT-4 have revolutionized natural language processing, yet they remain prone to two significant issues in knowledge-intensive tasks: hallucination and the cognition gap. Hallucination occurs when models rely on outdated or incorrect "parametric knowledge" stored in their weights. The cognition gap refers to the difficulty LLMs face when trying to connect fragmented external facts—often provided via standard Retrieval-Augmented Generation (RAG)—into a coherent reasoning chain for complex, multi-hop questions.
Existing solutions often involve iterative querying of external databases by the LLM, which is computationally expensive, slow, and increases API costs. KG-CoT addresses these issues by introducing a "Large + Small" paradigm that offloads graph traversal to a specialized, lightweight model.
Methodology: The KG-CoT Framework
The KG-CoT (Chain-of-Thought over Knowledge Graphs) framework operates in three distinct stages to bridge the gap between structured knowledge and natural language reasoning.
- Step-by-Step Graph Reasoning: A small-scale model analyzes the input question and the Knowledge Graph (KG). It calculates relation scores to identify the most probable paths leading to the answer, constructing a transition matrix for each reasoning "hop."
- Reasoning Path Generation: The system retraces these high-confidence steps to produce human-readable reasoning chains. For instance, for the question "Whence is the artist nominated for awards for 'Live from Paris'?", the path might be:
Live from Paris → music.artist.album_reversed → Shakira → music.artist.origin → Colombia. - Joint Reasoning: These serialized paths are integrated into the LLM's prompt. The LLM then synthesizes this structured evidence to generate a final, factually grounded answer.
This "plug-and-play" approach is highly efficient, requiring only a single API call to the LLM per question, compared to the multiple calls required by iterative search methods.
Experimental Evaluation
The researchers evaluated KG-CoT across four diverse benchmarks: WebQSP and CompWebQ (multi-hop), SimpleQuestions (single-hop), and WebQuestions (open-domain).
Key Results
KG-CoT consistently outperformed standard prompting and previous state-of-the-art (SOTA) retrieval-augmented methods. As shown in the results table below, ChatGPT combined with KG-CoT achieved an accuracy of 82.1% on WebQSP, a significant improvement over the 63.3% achieved with standard IO prompts.
The framework proved to be model-agnostic, providing substantial gains for Llama2-7B, Llama2-13B, and GPT-4. Furthermore, KG-CoT demonstrated superior retrieval capabilities, achieving a 91.4% hit rate on WebQSP, far surpassing traditional retrievers like BM25 and DPR.
Conclusion and Future Work
KG-CoT demonstrates that providing a "chain" of knowledge is significantly more effective than providing a "bag" of fragmented facts. By leveraging a small model for structured search and a large model for synthesis, the framework achieves high accuracy with minimal latency and cost.
However, the system's performance is inherently tied to the quality and coverage of the underlying Knowledge Graph. Future research will focus on improving the graph reasoning model's ability to handle highly complex logical constraints and enhancing its generalizability across different KGs without the need for extensive retraining.
Summary updated