SCENIC: Semantic-Conditioned Edge-Aware Neural Framework for Structured IoT Command Generation

The paper introduces SCENIC, a Semantic-Conditioned Edge-Aware Neural Framework designed for structured command generation in edge IoT environments. It utilizes sub-0.2B-scale transformer backbones, achieving a 99.0% exact match rate on the Smart Home Instruct-Bench with a pruned INT8 encoder-decoder model that reduces size by 25.38% while maintaining 91.0% EM@1 accuracy. This framework is significant for practitioners as it enables efficient deployment of language models on resource-constrained edge devices, enhancing smart-home command processing while addressing memory and latency challenges.