What Breaks Under Pruning in Smart Homes, and When? Evaluating LLM Degradation Across Architectures and Task Complexity
Pruning can reduce the deployment cost of large language models (LLMs), but its impact on context-grounded tool calling remains poorly understood. We systematically study pruning-induced degradation in smart-home tool calling across four LLMs spanning dense AI architecture, dense hybrid, and mixture-of-experts (Mixture-of-Experts AI) architectures, together with depth, width, h
Reading in Plain English mode — toggle below to switch
Pruning can reduce the deployment cost of large language models (LLMs), but its impact on context-grounded tool calling remains poorly understood. We systematically study pruning-induced degradation in smart-home tool calling across four LLMs spanning dense AI architecture, dense hybrid, and mixture-of-experts (Mixture-of-Experts AI) architectures, together with depth, width, h
Understanding pruning's impact on LLMs, especially for critical applications like smart homes, is crucial for deploying efficient and reliable models. This research helps ML engineers make informed decisions about model compression strategies, balancing performance and cost.