Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

When Context Leads but Parametric Memory Follows in Large Language Models

Yufei Tao, Adam Hiatt, Erik Haake, Antonie J. Jetter, Ameeta Agrawal

TL;DR

Investigating how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios reveals consistent patterns across models, including a consistent reliance on both contextual and parametric knowledge, and a decrease in hallucinations with increasing context.

Abstract

Large language models (LLMs) have demonstrated remarkable progress in leveraging diverse knowledge sources. This study investigates how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios. We introduce a novel dataset, WikiAtomic, and systematically vary context sizes to analyze how LLMs prioritize and utilize the provided information and their parametric knowledge in knowledge-consistent scenarios. Additionally, we also study their tendency to hallucinate under varying context sizes. Our findings reveal consistent patterns across models, including a consistent reliance on both contextual (around 70%) and parametric (around 30%) knowledge, and a decrease in hallucinations with increasing context. These insights highlight the importance of more effective context organization and developing models that use input more deterministically for robust performance.

When Context Leads but Parametric Memory Follows in Large Language Models

TL;DR

Investigating how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios reveals consistent patterns across models, including a consistent reliance on both contextual and parametric knowledge, and a decrease in hallucinations with increasing context.

Abstract

Large language models (LLMs) have demonstrated remarkable progress in leveraging diverse knowledge sources. This study investigates how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios. We introduce a novel dataset, WikiAtomic, and systematically vary context sizes to analyze how LLMs prioritize and utilize the provided information and their parametric knowledge in knowledge-consistent scenarios. Additionally, we also study their tendency to hallucinate under varying context sizes. Our findings reveal consistent patterns across models, including a consistent reliance on both contextual (around 70%) and parametric (around 30%) knowledge, and a decrease in hallucinations with increasing context. These insights highlight the importance of more effective context organization and developing models that use input more deterministically for robust performance.
Paper Structure (29 sections, 25 figures, 1 table)

This paper contains 29 sections, 25 figures, 1 table.

Table of Contents

  1. Introduction
  2. Task and Terminology
  3. WikiAtomic Dataset
  4. Experiments and Evaluation
  5. Experiments
  6. Evaluation
  7. Models
  8. Results and Analysis
  9. In knowledge-consistent setting, how do models prioritize sources of knowledge?
  10. Which parts of context are used?
  11. How similar are various types of knowledge?
  12. In knowledge-consistent setting, (how much) do models hallucinate?
  13. Further Analyses
  14. Discussion
  15. Related Work
  16. ...and 14 more sections

Figures (25)

  • Figure 1: Overview of the dataset creation and model evaluation pipeline
  • Figure 2: An example of Context, Question, Model Response (GPT-4o) and the list of Atomic Response mapped to contextual knowledge and parametric knowledge
  • Figure 3: Knowledge-consistency between parametric knowledge and input context of WikiAtomic topics, computed using SBERT reimers2019sentencebert
  • Figure 4: Contextual (local), parametric (global), and total sentences in responses for (a) GPT-4o, (b) Claude Opus, (c) Llama 3 70B, and (d) Mistral 8x22B. On the $x$-axis, $k=0$ serves as the baseline when no context is provided.
  • Figure 5: Percentage of each quartile of context recalled in response (GPT-4o)
  • ...and 20 more figures