AINeutralarXiv – CS AI · 7h ago6/10
🧠
ProbeScale: Probing Analysis to Optimize Neural Scaling Laws for Efficient Small Language Model Inference
Researchers introduce ProbScale, a framework that combines neural scaling laws with probing analysis to identify parameter-efficient subnetworks in Small Language Models. The method achieves 5-10x parameter reduction while maintaining 95-98% performance on downstream tasks, addressing deployment challenges for resource-constrained environments.