Paper accepted at ACL 2026!

Our paper OCP: Outlier-Centric Probing for Dynamic Structured Pruning of LLMs has been accepted at ACL 2026 (CCF-A). This work proposes an input-adaptive LLM pruning framework that dynamically allocates layer-wise sparsity based on critical tokens, achieving up to 25% perplexity reduction compared to SOTA methods at 1.6× speedup.