Mathematical Reasoning Failures in Large Language Models: A Systematic Review and Novel Solution Framework
Year: 2025
Type: article
Abstract: Large language models (LLMs) exhibit a systematic paradox in mathematical reasoning capabilities: achieving high performance on standardized mathematical benchmarks while demonstrating consistent failures in basic arithmetic operations. This systematic review examines mathematical reasoning limitations in current state-of-the-art LLMs through analysis of 47 peer-reviewed studies published between 2022-2025. We identify three primary failure modes: (1) tokenization-induced numerical processing errors with frequency rates of 23-31%, (2) pattern matching over algorithmic reasoning evidenced by 15-65% performance degradation under numerical modifications, and (3) absence of verification mechanisms leading to 34% reasoning inconsistency rates. Our analysis reveals that current solutions—including tool augmentation, chain-of-thought prompting, and neuro-symbolic approaches—address symptoms rather than fundamental architectural limitations. We propose a novel Hierarchical Mathematical Reasoning Architecture (HMRA) that integrates dedicated mathematical working memory, continuous symbol grounding, and built-in verification systems. Theoretical analysis demonstrates that HMRA addresses identified communication complexity barriers while maintaining computational efficiency. This framework provides a systematic pathway toward reliable mathematical reasoning in artificial intelligence systems with implications for scientific computing, educational applications, and automated mathematical discovery.
Author: Mustafa, Waleed
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6912308223
FWCI:
Cites:
Cited by:
Related to: