Theori announced that a research paper authored by one of its researchers as first author has been accepted for presentation at ACM/SIGAPP SAC 2026 (The 41st ACM SIGAPP Symposium on Applied Computing).

ACM/SIGAPP SAC is a globally recognized international conference that evaluates both academic excellence and real-world industrial impact in computer science and artificial intelligence. 

The accepted paper was led by Sanghyeon Park, a researcher on Theori’s LLM security solution αprism team, as first author. Titled “FRAIN to Train: A Fast-and-Reliable Solution for Decentralized Asynchronous Federated Learning,” the study demonstrates, both empirically and mathematically, that AI model merging can remain stable even in highly challenging conditions. These include environments with malicious participants, extreme data imbalance, and severe network latency occurring simultaneously.

The research offers meaningful insights into improving the reliability of AI models in decentralized and asynchronous settings. In particular, its findings have been applied to advance Theori’s LLM security solution, αprism. The checkpoint merging technique currently used in αprism’s Identifier model has been refined based on this research, enabling more consistent performance and more stable decision-making even in incomplete or adversarial environments.

The paper will be presented at SAC 2026, which will take place from March 23 to 27, 2026, in Thessaloniki, Greece.

“This research addresses real-world challenges that arise when developing AI models in practice,” said Sanghyeon Park, researcher at Theori. “We hope these results will serve as a useful reference for future LLM security research and practical discussions.”