Improving reasoning capabilities of large language models
Dixit, Tanay
This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.
Permalink
https://hdl.handle.net/2142/129521
Description
Title
Improving reasoning capabilities of large language models
Author(s)
Dixit, Tanay
Issue Date
2025-04-11
Director of Research (if dissertation) or Advisor (if thesis)
Han, Jiawei
Department of Study
Siebel School Comp & Data Sci
Discipline
Computer Science
Degree Granting Institution
University of Illinois Urbana-Champaign
Degree Name
M.S.
Degree Level
Thesis
Keyword(s)
Natural Language Processing
Large Language Models
Abstract
Large Language Models (LLMs) have succeeded at solving several wide range of tasks like mathematical problem solving, code generation, common-sense reasoning, etc. The recent success of these models is largely attributed to scaling in both model size and training data, which often includes massive amounts of web-scale and synthetic data — raising important questions about their true generalization capabilities. Several studies have highlighted critical failure cases in the reasoning abilities of LLMs, such as token biases in logical problem-solving and sensitivity to the order of premises, indicating a reliance on surface-level cues rather than true logical understanding. Additionally, these reasoning abilities ares hown to emerge only when models are trained on extremely large datasets. This technique of learning to reason deviates from how humans learn to reason and think. Humans learn to solve problems by first understanding and acquiring the fundamental principles involved in reasoning, and then learn to apply these principles to new tasks, rather than directly learning to solve hundreds of complex problems. Inspired by this, we aim to train LLMs to learn to reason with the help of axioms - fundamental principles of reasoning, in particular causal axioms. Causal axioms lay the crucks of causal inference which humans use in making decisions or inferences in several scenarios. The influence of causal axioms on the reasoning abilities of LLMs remains underexplored; in this work, we demonstrate that causal axiomatic training can enhance LLM performance across a broad range of reasoning tasks, even those not directly related to causality. Our extensive evaluation results across 16 benchmarks, shows that LLMs fine-tuned using our axiomatic data show stronger gains compared to baseline approaches on most tasks.
Use this login method if you
don't
have an
@illinois.edu
email address.
(Oops, I do have one)
IDEALS migrated to a new platform on June 23, 2022. If you created
your account prior to this date, you will have to reset your password
using the forgot-password link below.
