Published on
Updated on
While many consumers are already using compressed AI foundation models on their phones, laptops and customized hardware, few realize it. “When you compress a model, you’re not just making it smaller,” Khaza Anuarul Hoque said. “You’re also changing how it behaves.”
Artificial intelligence (AI) is rapidly moving out of massive data centers and onto everyday devices — from smartphones and laptops to drones and autonomous vehicles.
But making that transition possible requires more than just shrinking today’s powerful models. It requires ensuring that, when they are compressed to run on limited hardware, they still behave the way we expect, especially in situations where safety is critical.
A new research effort led by Mizzou Engineering computer scientist Khaza Anuarul Hoque aims to do exactly that.
Hoque and his team are developing a framework designed to compress foundation models such as large language models (LLMs) and vision-language models (VLMs) to run on devices with limited processing power and energy while still providing formal, mathematical guarantees that key behaviors are preserved.
The project, which the researchers call VeriFAI, has drawn support from NVIDIA, which recently awarded the team 32,000 hours of GPU computing time to scale their approach.
“At a high level, we’re trying to answer a simple but critical question,” Hoque said. “Can we make these models smaller without introducing behaviors we don’t want — and mathematically prove that we haven’t?”