GISP

AI Frontier Lectures

Nov 13, 2026 · Artificial Intelligence

Train-Free Iterative Pruning (GISP) Outperforms Layer-Wise LLM Compression

This ACL 2026 outstanding paper introduces GISP, a train-free iterative structured pruning method that globally ranks attention heads and MLP channels by normalized first-order importance scores, achieving lower perplexity than layer-wise uniform pruning on Llama2-7B/13B at 40-50% sparsity without fine-tuning.