External Knowledge Integration in Large Language Models: A Survey on Methods, Challenges, and Future Directions

Submitted by Sirko Schindler on 11/20/2025 - 12:58

Tracking #: 3976-5190

Authors: 
Itisha Yadav
Sirko Schindler
Diana Peters
Roman Klinger

Responsible editor: 
Guest Editors 2025 LLM GenAI KGs

Submission type: 
Survey Article
Abstract: 
Large Language Models (LLMs) have shown effectiveness in various natural language understanding (NLU) tasks. However, they face notable limitations like hallucinations, a lack of contextual knowledge, and outdated or incomplete knowledge when applied across knowledge-intensive domains such as scientific research, biomedical sciences, finance, law, and others. These challenges commonly arise from the scarcity and under-representation of domain-specific data during the training and model alignment phases, the latter being synonymous with reinforcement learning from human feedback (RLHF). Furthermore, LLMs struggle to provide nuanced expertise, as their internal knowledge remains static and generalized, hindering their ability to reason accurately or deliver context-aware results in specialized tasks. This survey investigates the integration of external knowledge into LLMs to address these limitations. The focus is on decoder-based LLMs, i.e., autoregressive models that generate text sequentially. By investigating parametric and non-parametric approaches, this work discusses methods to enhance model reasoning capabilities, factual accuracy, and adaptability for domain-specific and knowledge-intensive tasks. Additionally, it highlights the potential of integrating external knowledge to improve explainability and ensure more trustworthy outputs. This survey supports software developers and natural language processing (NLP) researchers in designing natural language understanding systems for specialized domains by leveraging pre-trained LLMs. Additionally, the work provides a foundation for advancing LLM-based NLU systems with insights into future research areas.
Full PDF Version: 
PDF icon swj3976.pdf
Previous Version: 
External Knowledge Integration in Large Language Models: A Survey on Methods, Challenges, and Future Directions
Tags: 
Reviewed
Long-term Stable Link to Resources: 
https://doi.org/10.5281/zenodo.15064852

Decision/Status: 
Accept

Solicited Reviews:
Click to Expand/Collapse
Review #1
By Célian Ringwald submitted on 02/Feb/2026
Suggestion:
Accept
Review Comment:

I really appreciated reading the new version of the proposed journal article, and I would like to thank the authors for the great improvement they made on all the aspects outlined in the reviews. The results are a high-quality survey with a detailed and clear methodology. The tables added fill many of the clarifications required. And I want to underline the significant enrichment made to the RAG section, which is a direction that is followed today, and this new version really helps to appreciate the current state of the art on these questions.
I think the article could be accepted in it current version, but if the authors have to update their file I would like to add two minors comments, that may easily checked:
I regret that all figures were deleted in this new version; from my point of view, Fig. 1 is still interesting and could support the new Table 2.
2.Concerning the constraint-decoding in LLM, i appreciate the small clarification added and answering to reviewer 3. Nonetheless, the author introduced it, but limitations of these approaches are still underdiscussed. Moreover the authors only consider it relativly to the ontology satisfaction perspective where more globally the syntax of the generated output is also a first layer of constraint that need to follow a LLM, when extracting structured content. From my current understanding, constrained decoding may have a significant impact on generation cost: https://arxiv.org/pdf/2305.13971. I’m not sure that the JSON extraction mode of ChatGPT solves all the problems, since many libraries, Grammar/Dot.txt/Instructor, are also trying to answer this question.

Review #2
By George Hannah submitted on 02/Feb/2026
Suggestion:
Accept
Review Comment:

I thank the authors for their comments and subsequent actions on the comments of my initial review. With the changes made to the paper, the work much more suitably addresses the target audience it defines when compared to the initial version of the paper. The authors place the work within the field of research in relation to existing surveys in the area, and increase the comprehensiveness of the survey within its self-defined scope.

The identification of limitations in the surveying methods used suitably address my comments in the original review.

The inclusion of the prompt that led to the given example of a hallucination in section 3.1 is a welcome addition as it providing readers with limited domain knowledge a clearer understanding as to why a specific response would be considered a hallucination. In addition to this, the table added in section 4.1, effectively summarises the different approaches of knowledge integration.

Considering my initial review, the author's responses it, and their alterations made to this version of the paper. I recommend that this paper to be accepted.