Announcement_29_reap_iclr

Mike Lasby’s collaborative work with Cerebras, “REAP the experts: Why pruning prevails for one-shot moe compression” (Lasby et al., 2026), has been accepted at the International Conference on Learning Representations (ICLR), 2026. This work explores the compression of Sparse Mixture of Experts (SMoE) models through expert compression techniques, demonstrating that REAP (Router-weighted Expert Activation Pruning) outperforms existing expert merging and pruning methods in terms of compressed model quality retention.