{"paper":{"title":"DECO: Sparse Mixture-of-Experts with Dense-Comparable Performance on End-Side Devices","license":"http://creativecommons.org/licenses/by/4.0/","headline":"DECO sparse MoE matches dense Transformer performance while activating only 20% of experts.","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Chaojun Xiao, Chenyang Song, Weilin Zhao, Xu Han, Yingfa Chen, Zhiyuan Liu","submitted_at":"2026-05-11T17:58:28Z","abstract_excerpt":"While Mixture-of-Experts (MoE) scales model capacity without proportionally increasing computation, its massive total parameter footprint creates significant storage and memory-access bottlenecks, which hinder efficient end-side deployment that simultaneously requires high performance, low computational cost, and small storage overhead. To achieve these properties, we present DECO, a sparse MoE architecture designed to match the performance of dense Transformers under identical total parameter budgets and training tokens. DECO utilizes the differentiable and flexible ReLU-based routing enhance"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"DECO, activating only 20% of experts, matches dense performance and outperforms established MoE baselines.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the learned expert-wise scaling and NormSiLU will continue to produce stable sparsity and matching performance when model size, data distribution, or hardware change substantially.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"DECO matches dense model performance at 20% expert activation via ReLU-based routing with learnable scaling and the NormSiLU activation, plus a 3x real-hardware speedup.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"DECO sparse MoE matches dense Transformer performance while activating only 20% of experts.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"fb0d46a45ae445bdec59bbd573a2ac45d8e9b9807b9a5e97c2abb680d0d9a919"},"source":{"id":"2605.10933","kind":"arxiv","version":3},"verdict":{"id":"b57856c1-ac6a-44b7-9016-87aeeeded385","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T07:25:41.711109Z","strongest_claim":"DECO, activating only 20% of experts, matches dense performance and outperforms established MoE baselines.","one_line_summary":"DECO matches dense model performance at 20% expert activation via ReLU-based routing with learnable scaling and the NormSiLU activation, plus a 3x real-hardware speedup.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the learned expert-wise scaling and NormSiLU will continue to produce stable sparsity and matching performance when model size, data distribution, or hardware change substantially.","pith_extraction_headline":"DECO sparse MoE matches dense Transformer performance while activating only 20% of experts."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.10933/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"claim_evidence","ran_at":"2026-05-20T05:02:00.913621Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T13:36:55.079203Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T10:31:17.132399Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T08:53:45.739118Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"96fadb87ebbca95e4bdf70d6c14a464c9c1764e752eea7a100432d28aacfb521"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"b3f92bf15c7f14242316b7185ec75741ceb26ba35c9d21eee0ef878cf7d36495"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}