Cincinnati Machine Learning Meetup

Topic: Inside IJCNN 2026: When LLMs Meet Malware and Neural Networks Meet Noise

What happens when you train an LLM on real malware binaries? What if you could build neural networks that learn accurately even when half your training labels are wrong?
These aren't theoretical questions—they're live problems being solved at the frontier of AI research.
At IJCNN 2026 in Maastricht, Netherlands, researchers presented breakthrough work on two critical challenges: using generative AI to automatically decompile malicious code,
and building robust learning systems that thrive under label noise. The results are striking: one approach achieved 86% functional correctness in generated code (versus 15% baselines),
while another delivered best-in-class accuracy across dozens of noisy real-world datasets.
In this talk, Sarjoun will walk through the technical innovations and practical lessons from two cutting-edge papers he reviewed for the conference.
You'll see how domain-adaptive pretraining transforms LLMs from generic code generators into specialized malware analysts. You'll discover why bidirectional
translation frameworks (Assembly ↔ Source) paired with malware-aware training unlock capabilities that general-purpose models can't match. And you'll learn
how kernel-based risk-sensitive learning lets neural networks distinguish signal from noise in corrupted training data.
But beyond the specific techniques, this talk explores fundamental questions about how we build AI systems: When does domain specialization beat general-purpose scale?
Why does functional correctness (can it run?) matter more than semantic similarity (does it look right?) for code generation? How do you evaluate AI systems when your
ground truth itself is noisy or approximate? And what makes some AI approaches transformative versus incremental?

Whether you're building AI systems, working with noisy data, or exploring how LLMs can tackle specialized domains, this talk offers concrete techniques and strategic insights from the forefront of computational intelligence research.

About the speaker:
Our speaker is Sarjoun Doumit, and he is passionate about technology and computer programming. His interests and expertise include machine learning, cyber security and
complex systems. He holds a PhD in Machine Learning and serves as a reviewer for major AI conferences including IJCNN (International Joint Conference on Neural Networks).
His work spans deep learning, robust learning under uncertainty, and AI-driven code analysis, contributing to advancing computational intelligence research in industrial
applications. He presents at technical meetups on emerging ML techniques and conference insights.