Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Large Language Models as Evaluators for Scientific Synthesis

Julia Evans, Jennifer D'Souza, Sören Auer

TL;DR

This work examines whether cutting-edge LLMs can serve as scalable evaluators of scientific syntheses, comparing GPT-4 Turbo and Mistral to human judgments on the CORE-GPT dataset. Using a structured prompt, the models rate syntheses on comprehensiveness, trust, and utility and justify their scores in JSON format. Findings show that although LLMs produce logically consistent assessments and agree with each other, their ratings poorly track human judgments, indicating current limitations in using LLMs as direct evaluators of scientific syntheses. The results suggest potential for LLM-based evaluation as a supplementary tool, but emphasize the need for larger datasets, robust prompting, and validation against human standards to ensure reliability for research and deployment.

Abstract

Our study explores how well the state-of-the-art Large Language Models (LLMs), like GPT-4 and Mistral, can assess the quality of scientific summaries or, more fittingly, scientific syntheses, comparing their evaluations to those of human annotators. We used a dataset of 100 research questions and their syntheses made by GPT-4 from abstracts of five related papers, checked against human quality ratings. The study evaluates both the closed-source GPT-4 and the open-source Mistral model's ability to rate these summaries and provide reasons for their judgments. Preliminary results show that LLMs can offer logical explanations that somewhat match the quality ratings, yet a deeper statistical analysis shows a weak correlation between LLM and human ratings, suggesting the potential and current limitations of LLMs in scientific synthesis evaluation.

Large Language Models as Evaluators for Scientific Synthesis

TL;DR

This work examines whether cutting-edge LLMs can serve as scalable evaluators of scientific syntheses, comparing GPT-4 Turbo and Mistral to human judgments on the CORE-GPT dataset. Using a structured prompt, the models rate syntheses on comprehensiveness, trust, and utility and justify their scores in JSON format. Findings show that although LLMs produce logically consistent assessments and agree with each other, their ratings poorly track human judgments, indicating current limitations in using LLMs as direct evaluators of scientific syntheses. The results suggest potential for LLM-based evaluation as a supplementary tool, but emphasize the need for larger datasets, robust prompting, and validation against human standards to ensure reliability for research and deployment.

Abstract

Our study explores how well the state-of-the-art Large Language Models (LLMs), like GPT-4 and Mistral, can assess the quality of scientific summaries or, more fittingly, scientific syntheses, comparing their evaluations to those of human annotators. We used a dataset of 100 research questions and their syntheses made by GPT-4 from abstracts of five related papers, checked against human quality ratings. The study evaluates both the closed-source GPT-4 and the open-source Mistral model's ability to rate these summaries and provide reasons for their judgments. Preliminary results show that LLMs can offer logical explanations that somewhat match the quality ratings, yet a deeper statistical analysis shows a weak correlation between LLM and human ratings, suggesting the potential and current limitations of LLMs in scientific synthesis evaluation.
Paper Structure (32 sections, 2 figures, 15 tables)

This paper contains 32 sections, 2 figures, 15 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Evaluation Metrics for Text Summarization.
  4. LLMs for Text Evaluation.
  5. LLMs for the Scientific Synthesis Evaluation Task
  6. LLM Synthesis Evaluation Output
  7. Qualitative Evaluations
  8. Correlation
  9. Discussion and Conclusion
  10. Prompt
  11. Main Evaluation Prompt
  12. Additional Mistral System Prompt
  13. Example GPT-4 Turbo Output
  14. Example Mistral Output
  15. Trust Evaluation
  16. ...and 17 more sections

Figures (2)

  • Figure 1: Example output from GPT-4 Turbo.
  • Figure 2: Example output from Mistral. The text under the separator line was returned outside of the JSON response.