International Journal of Scientific Research and Engineering Development

Fuzzing is established as the de facto standard for software vulnerability testing; however, its efficacy is largely compromised by “corpus bloat”, where redundant inputs degrade the feedback loop. This review aims to analyze the evolution of corpus minimization from the Algorithmic Era to the modern-day Generative Renaissance. This paper demonstrates that while the traditional set-cover algorithms such as aflcmin, MoonLight and OptiMin provide mathematical efficiency, they suffer from “semantic blindness”, discarding the valuable seeds that fail to find immediate coverage gains.Furthermore, the review paper examines the failure of the early Deep Learning integrations in corpusminimization, which were stifled by the “Vanishing Gradient” problem and the “Latency Wall”. This paper identifies a critical pathology and terms the “Semantic Bottleneck”: while Large Language Models (LLMs) generate high-fidelity inputs, the industry still relies on legacy heuristic algorithms for corpus minimization. This creates a bottleneck where the diversity gains of the generative agents are pruned by “dumb” filters. The review concludes that the “ML Gap” that was first identified in 2019 remains unbridged. Future research should focus on Lightweight Supervised Learning to achieve efficient intelligence by aligning semantic awareness with the velocity of modern Continuous Integration (CI) pipelines.

Kekuluwala Pathiranage Navodya Lakshani, Kiruparetnam Jude Myuran,"Applications of Machine Learning in Fuzzer Corpus Minimization: A Systematic Literature Review" International Journal of Scientific Research and Engineering Development, V9(2): Page(1739-1745) Mar-Apr 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.