Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLMs as Repositories of Factual Knowledge: Limitations and Solutions

Seyed Mahed Mousavi, Simone Alghisi, Giuseppe Riccardi

TL;DR

The paper investigates LLMs as repositories of time-sensitive factual knowledge, highlighting that knowledge snapshots collected over time lead to outdated or inconsistent outputs. It introduces DyKnow, a dynamic benchmark that uses Wikidata to continuously assess accuracy and consistency of 24 LLMs under prompt perturbations, revealing substantial reliability gaps. To address these issues, the authors compare existing approaches (ROME, MEMIT, SERAC, IKE, RAG) and propose ENtity-Aware Fine-tuning (ENAF), a soft neurosymbolic method that injects structured entity representations to stabilize recall across varied prompts. Empirically, while editing and retrieval methods show partial gains, ENAF demonstrates potential by distributing recall across multiple layers and linking lexical variants to unified symbolic references, suggesting dynamic benchmarking combined with structured fine-tuning as a robust path toward reliable time-sensitive knowledge in LLMs.

Abstract

LLMs' sources of knowledge are data snapshots containing factual information about entities collected at different timestamps and from different media types (e.g. wikis, social media, etc.). Such unstructured knowledge is subject to change due to updates through time from past to present. Equally important are the inconsistencies and inaccuracies occurring in different information sources. Consequently, the model's knowledge about an entity may be perturbed while training over the sequence of snapshots or at inference time, resulting in inconsistent and inaccurate model performance. In this work, we study the appropriateness of Large Language Models (LLMs) as repositories of factual knowledge. We consider twenty-four state-of-the-art LLMs that are either closed-, partially (weights), or fully (weight and training data) open-source. We evaluate their reliability in responding to time-sensitive factual questions in terms of accuracy and consistency when prompts are perturbed. We further evaluate the effectiveness of state-of-the-art methods to improve LLMs' accuracy and consistency. We then propose "ENtity-Aware Fine-tuning" (ENAF), a soft neurosymbolic approach aimed at providing a structured representation of entities during fine-tuning to improve the model's performance.

LLMs as Repositories of Factual Knowledge: Limitations and Solutions

TL;DR

The paper investigates LLMs as repositories of time-sensitive factual knowledge, highlighting that knowledge snapshots collected over time lead to outdated or inconsistent outputs. It introduces DyKnow, a dynamic benchmark that uses Wikidata to continuously assess accuracy and consistency of 24 LLMs under prompt perturbations, revealing substantial reliability gaps. To address these issues, the authors compare existing approaches (ROME, MEMIT, SERAC, IKE, RAG) and propose ENtity-Aware Fine-tuning (ENAF), a soft neurosymbolic method that injects structured entity representations to stabilize recall across varied prompts. Empirically, while editing and retrieval methods show partial gains, ENAF demonstrates potential by distributing recall across multiple layers and linking lexical variants to unified symbolic references, suggesting dynamic benchmarking combined with structured fine-tuning as a robust path toward reliable time-sensitive knowledge in LLMs.

Abstract

LLMs' sources of knowledge are data snapshots containing factual information about entities collected at different timestamps and from different media types (e.g. wikis, social media, etc.). Such unstructured knowledge is subject to change due to updates through time from past to present. Equally important are the inconsistencies and inaccuracies occurring in different information sources. Consequently, the model's knowledge about an entity may be perturbed while training over the sequence of snapshots or at inference time, resulting in inconsistent and inaccurate model performance. In this work, we study the appropriateness of Large Language Models (LLMs) as repositories of factual knowledge. We consider twenty-four state-of-the-art LLMs that are either closed-, partially (weights), or fully (weight and training data) open-source. We evaluate their reliability in responding to time-sensitive factual questions in terms of accuracy and consistency when prompts are perturbed. We further evaluate the effectiveness of state-of-the-art methods to improve LLMs' accuracy and consistency. We then propose "ENtity-Aware Fine-tuning" (ENAF), a soft neurosymbolic approach aimed at providing a structured representation of entities during fine-tuning to improve the model's performance.
Paper Structure (11 sections, 4 figures, 5 tables)

This paper contains 11 sections, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Assessment of LLMs' Reliability (RQ1)
  4. DyKnow: Dynamic Knowledge Benchmark
  5. Evaluating LLMs' Accuracy
  6. Evaluating LLMs' Consistency
  7. Results
  8. Enhancing LLMs' Reliability (RQ2)
  9. State-of-the-Art Methods
  10. ENtity-Aware Finetuning (ENAF)
  11. Conclusion

Figures (4)

  • Figure 1: Output consistency across LLMs when prompted with subject perturbations. The x-axis represents the release year of each model, and the y-axis shows the level of output consistency when prompted with varied lexicalizations of the same subject. Subscripts ${I.}$ and ${C.}$ stand for Instruct and Chat, respectively. Instruction-tuned models demonstrate a comparatively higher prompt agreement.
  • Figure 2: Output consistency across LLMs when prompted with property perturbations. The x-axis represents the release year of each model, and the y-axis shows the level of output consistency when prompted with varied lexicalizations of the property. Subscripts ${I.}$ and ${C.}$ stand for Instruct and Chat, respectively. Instruction-tuned models demonstrate a comparatively higher prompt agreement.
  • Figure 3: The layer-wise distribution of recall contributions for the subject entity Cristiano Ronaldo associated with relevant attributes (i.e. affiliated clubs) across different fine-tuning and knowledge-editing strategies in GPT-2. "(A) Pre-Trained Model" shows the contributions scattered across mid-to-upper layers. "(B) Vanilla Fine-Tuning"skews the distribution towards a single attribute. Knowledge editing algorithms "(C) ROME" and "(D) MEMIT" concentrate the recall in the top layers for an edited association. ENtity-Aware Fine-tuning with "(E) Selected Named Entity Tags" distributes recall contributions more evenly across layers, enhancing factual consistency. While "(F) ID Tags + Selected Named Entity Tags" shifts contributions to the lower layers (responsible for factual retrieval for subject entity) and higher level, minimizing the impact of mid-layer contributions.
  • Figure 4: The layer-wise distribution of recall contributions for subject entity LeBron James associated with affiliated clubs in GPT-2. "(A) Pre-Trained Model" shows scattered contributions across layers. While "(F) ID Tags + Selected Named Entity Tags" shifts the contributions to the lower layers (responsible for factual retrieval for subject entity) and higher levels, minimizing the impact of mid-layer contributions.