Reasoning beyond scale: Structured inference for small language models

Recent progress in large language models (LLMs), such as GPT-4 [1], PaLM [2], and LLaMA [3], has substantially advanced the field of natural language processing (NLP), particularly in tasks requiring reasoning, such as multi-hop question answering (QA) and claim verification [4, 5]. Despite these achievements, such models require significant computational and financial resources, limiting their real-world accessibility [6, 7]. This has motivated a growing interest in small language models (SLMs), typically with fewer than 8 billion parameters, which offer more efficient alternatives. However, SLMs face major challenges when performing complex reasoning under long-context and distractor-rich scenarios, often exhibiting difficulties with factual consistency and multi-step inference [8, 9, 10]. This thesis investigates multi-hop reasoning in small models through the lens of structured knowledge integration, focusing on whether small models can remain grounded in relevant context while avoiding hallucination and distractor interference. Multi-hop reasoning requires the model to chain together intermediate facts and establish logical links across multiple documents or sentences [11]. A key question explored is whether explicit knowledge representations, such as knowledge graphs (KGs), can act as a scaffold to improve logical consistency, contextual grounding, and interpretability in reasoning tasks. The study further examines how SLMs manage cross-document coreference, semantic role tracking, and inference reliability when guided by structured signals. This thesis attempts to identify the linguistic and representational bottlenecks that hinder SLMs from attaining robust reasoning, going beyond surface-level performance. It analyzes the conditions under which models succeed or fail to maintain coherence across reasoning chains, particularly when processing lengthy, noisy input. The thesis also explores how models respond to adversarial distractors and the extent to which structured inputs reduce hallucination rates [12, 13]. Overall, the research contributes to a broader understanding of how reasoning, factuality, and context grounding can be enabled in models – an essential step toward deploying capable and trustworthy NLP systems in real-world environments.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Aakriti, -

Contributors dc:contributor

• Han, Jiawei

• Small Language Models
• Knowledge Graphs
• Question Answering
• Claim Verification
• Reasoning

Statement dc:rights

• Copyright 2025 - Aakriti

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/129549

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F129549