LaMoC: Loss-Aware Modular Compression for LLMs

Authors/institution/venue: Mohanad Odema and Jacob Song; affiliations not specified on the abstract page; arXiv preprint submitted 31 August. Problem: pruning/compression can damage language-model quality. Contribution: combines activation and empirical-Fisher statistics to rank modules; authors report average 2.5% lower perplexity and 1% relative task-accuracy improvement on 4–8B models. Significance: possible edge and cost optimisation. Limitations: preprint, smaller-model scope and no independent reproduction. Paper