Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DGoT: Dynamic Graph of Thoughts for Scientific Abstract Generation

Xinyu Ning, Yutong Zhao, Yitong Liu, Hongwen Yang

TL;DR

Addressing high cost and hallucination risks in LLM-based scientific abstract generation, this work introduces Dynamic Graph of Thought (DGoT), which adaptively adjusts the graph structure during reasoning to balance quality and cost. The method splits training and reasoning, deriving two thresholding schemes—Simple Mean and Gumbel—to govern Dynamic Generate, Dynamic Aggregate, and Dynamic Improve steps, with a ranking module selecting top thoughts. On PubMedCite, DGoT achieves substantially better cost-effectiveness than fixed GoT/ToT/CoT baselines while maintaining competitive ROUGE scores, demonstrating practical efficiency for large-scale abstract generation. The authors provide open-source code to facilitate reproduction and further development.

Abstract

The method of training language models based on domain datasets has obtained significant achievements in the task of generating scientific paper abstracts. However, such models face problems of generalization and expensive training costs. The use of large language models (LLMs) to solve the task of generating paper abstracts saves the cost of model training. However, due to the hallucination problem of LLM, it is often necessary to improve the reliability of the results through multi-round query prompt approach such as Graph of Thoughts (GoT), which also brings additional reasoning costs. In this paper, we propose a Dynamic Graph of Thought (DGoT). It not only inherits the advantages of the existing GoT prompt approach, but also dynamically adjust the graph structure according to data characteristics while reducing model reasoning cost. Experimental results show that our method's cost-effectiveness in abstract generation tasks is only 43.7% to 56.4% of other multi-round query prompt approaches. Our code is available at https://github.com/JayceNing/DGoT.

DGoT: Dynamic Graph of Thoughts for Scientific Abstract Generation

TL;DR

Addressing high cost and hallucination risks in LLM-based scientific abstract generation, this work introduces Dynamic Graph of Thought (DGoT), which adaptively adjusts the graph structure during reasoning to balance quality and cost. The method splits training and reasoning, deriving two thresholding schemes—Simple Mean and Gumbel—to govern Dynamic Generate, Dynamic Aggregate, and Dynamic Improve steps, with a ranking module selecting top thoughts. On PubMedCite, DGoT achieves substantially better cost-effectiveness than fixed GoT/ToT/CoT baselines while maintaining competitive ROUGE scores, demonstrating practical efficiency for large-scale abstract generation. The authors provide open-source code to facilitate reproduction and further development.

Abstract

The method of training language models based on domain datasets has obtained significant achievements in the task of generating scientific paper abstracts. However, such models face problems of generalization and expensive training costs. The use of large language models (LLMs) to solve the task of generating paper abstracts saves the cost of model training. However, due to the hallucination problem of LLM, it is often necessary to improve the reliability of the results through multi-round query prompt approach such as Graph of Thoughts (GoT), which also brings additional reasoning costs. In this paper, we propose a Dynamic Graph of Thought (DGoT). It not only inherits the advantages of the existing GoT prompt approach, but also dynamically adjust the graph structure according to data characteristics while reducing model reasoning cost. Experimental results show that our method's cost-effectiveness in abstract generation tasks is only 43.7% to 56.4% of other multi-round query prompt approaches. Our code is available at https://github.com/JayceNing/DGoT.
Paper Structure (33 sections, 10 equations, 14 figures, 7 tables)

This paper contains 33 sections, 10 equations, 14 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Background & Notation
  3. Probability in LLM
  4. Prompt approaches
  5. Input-Output (IO)
  6. Chain-of-Thought (CoT)
  7. Tree-of-Thought (ToT)
  8. Graph-of-Thought (GoT)
  9. Method
  10. Procedure Overview
  11. Training Process
  12. Simple Mean Threshold
  13. Gumbel Threshold
  14. Reasoning Process
  15. Experiments
  16. ...and 18 more sections

Figures (14)

  • Figure 1: The overall process of our method, including the training process and reasoning process.
  • Figure 2: An example of a prompt framework.
  • Figure 3: Score distribution of training data. The dashed line represents the mean of the corresponding transformation score.
  • Figure 4: The distribution of the maximum score for each transformation. The solid line represents the kernel density estimation curve, while the dashed line represents the Gumbel distribution curve estimated according to the method in \ref{['subsec:gumbelthreshold']}. According to Equ. \ref{['GumbelThresh']}, taking the generation transformation as an example, the corresponding scores for confidence levels of 25%, 50%, and 75% are calculated.
  • Figure 5: Scores and costs of abstract generated by ChatGLM2-6B under different prompt approaches.
  • ...and 9 more figures