AI model compression technique reveals hidden tradeoff: removing unnecessary image data breaks low-precision math

Researchers found that two standard techniques for shrinking AI models work against each other. When you remove unnecessary image tokens from a multimodal AI model that's already been converted to low-precision math (to save memory), you accidentally discard the data points that keep the math stable — making the model less accurate. The fix: a method that checks both whether image data matters semantically and whether it matters numerically before pruning.

Multimodal AI models are too large and memory-intensive to run on phones, edge devices, or anything without a data center. Companies keep trying two separate tricks to shrink them: removing redundant image patches and converting math to lower precision. This paper shows those tricks interfere. Most deployments try both independently, which means they're probably leaving accuracy on the table. The researchers' co-optimized approach gets the same model size with measurably better accuracy — which translates directly to making these models practical for resource-constrained devices where they don't currently work.