Improving Large Language model reasoning — On Neurosymbolic training utilizing Family Tree problem

Thumbnail Image

Files

master_Leuze_Federico_2025.pdf (2.86 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Electrical Engineering | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2025-04-15

Department

Major/Subject

Speech and Language Technology

Mcode

Degree programme

Master's Programme in Computer, Communication and Information Sciences

Language

en

Pages

29

Series

Abstract

This thesis examines the potential of employing neurosymbolic methodologies to enhance large language models (LLMs) with the objective of improving their reasoning capabilities, which remain a significant limitation of current LLMs. Enhancing the reasoning abilities of LLMs could substantially improve their performance on out-of-distribution tasks, thereby increasing their capacity to solve novel problems, a capability they currently lack. The central hypothesis explored in this study is that the inadequacy of reasoning demonstrated by LLMs is partly due to the absence of explicit reasoning steps in their training data. By introducing such reasoning steps, it is hypothesized that the models would exhibit improved performance in reasoning-related tasks. To date, this approach has not been implemented programmatically. Therefore, this thesis proposes a novel methodology that leverages neurosymbolic systems to enhance the training of LLMs. To investigate this hypothesis, experiments were conducted using the CLEVR dataset. A baseline model was trained without incorporating chained explanations, while the proposed model included them. Extensive experimentation revealed no performance improvements in the proposed model compared to the baseline. On the contrary, the results indicated a decline in performance for the proposed model. These findings suggest that fine-tuning alone may not be the most effective approach for achieving the desired improvements in reasoning capabilities. Despite these results, the exploration of neurosymbolic training for LLMs remains a promising area for future research, with the potential to uncover more effective strategies for addressing this limitation.

Description

Supervisor

Bäckström, Tom

Keywords

large language models, neurosymbolic, reasoning, family tree, CLUTRR, teacher student method

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202505203865

Collections

[dipl] Sähkötekniikan korkeakoulu / ELEC