{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:QH7BB27ABTKNLM6BUUU3IED56O","short_pith_number":"pith:QH7BB27A","schema_version":"1.0","canonical_sha256":"81fe10ebe00cd4d5b3c1a529b4107df3967ab5ca8ecaec9b2d22dfab21ec1ddd","source":{"kind":"arxiv","id":"2606.10722","version":1},"attestation_state":"computed","paper":{"title":"Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hantao Huang, Hao Zeng, Ian En-Hsu Yen, Jinyuan Shi, Minghui Yu, Ruixuan Huang, Yifan Huang, Ziyi Guan","submitted_at":"2026-06-09T11:32:58Z","abstract_excerpt":"We study dense-to-sparse continual training as a way to construct channel-sparse large language models from dense checkpoints. Starting from a Qwen2.5-8B dense backbone, we continue training at 32K context and introduce a predictor-gated sparse SwiGLU FFN in the 32K stage. For each token and layer, we use a low-rank predictor to produce FFN-channel routing logits. We then apply a bank-wise top-k rule to retain 16 channels in every 64-channel bank, yielding 4x sparsity in the FFN intermediate activation. Unlike post-hoc sparse inference methods, the routing module is placed on the main language"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2606.10722","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CL","submitted_at":"2026-06-09T11:32:58Z","cross_cats_sorted":[],"title_canon_sha256":"75b29590d8c2ebd24bdc01b3c86f32d2fde8cac5a6e558815bbc4260e06c6adc","abstract_canon_sha256":"d6393d1fba881d6e5c3ac998aab94f6ef6b1d4ccc2b8adb629f89eab1f1a52e1"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-06-10T01:10:36.487005Z","signature_b64":"OUgsH6mIihsFFxYXvieABDuwD34z2WYrJbV7jD7bNEpmrCsHOEw2Yw/xS4Cohk/HWZ4pSdl91me5ChNrAhRzCw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"81fe10ebe00cd4d5b3c1a529b4107df3967ab5ca8ecaec9b2d22dfab21ec1ddd","last_reissued_at":"2026-06-10T01:10:36.486196Z","signature_status":"signed_v1","first_computed_at":"2026-06-10T01:10:36.486196Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Hantao Huang, Hao Zeng, Ian En-Hsu Yen, Jinyuan Shi, Minghui Yu, Ruixuan Huang, Yifan Huang, Ziyi Guan","submitted_at":"2026-06-09T11:32:58Z","abstract_excerpt":"We study dense-to-sparse continual training as a way to construct channel-sparse large language models from dense checkpoints. Starting from a Qwen2.5-8B dense backbone, we continue training at 32K context and introduce a predictor-gated sparse SwiGLU FFN in the 32K stage. For each token and layer, we use a low-rank predictor to produce FFN-channel routing logits. We then apply a bank-wise top-k rule to retain 16 channels in every 64-channel bank, yielding 4x sparsity in the FFN intermediate activation. Unlike post-hoc sparse inference methods, the routing module is placed on the main language"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2606.10722","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2606.10722/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2606.10722","created_at":"2026-06-10T01:10:36.486332+00:00"},{"alias_kind":"arxiv_version","alias_value":"2606.10722v1","created_at":"2026-06-10T01:10:36.486332+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2606.10722","created_at":"2026-06-10T01:10:36.486332+00:00"},{"alias_kind":"pith_short_12","alias_value":"QH7BB27ABTKN","created_at":"2026-06-10T01:10:36.486332+00:00"},{"alias_kind":"pith_short_16","alias_value":"QH7BB27ABTKNLM6B","created_at":"2026-06-10T01:10:36.486332+00:00"},{"alias_kind":"pith_short_8","alias_value":"QH7BB27A","created_at":"2026-06-10T01:10:36.486332+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O","json":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O.json","graph_json":"https://pith.science/api/pith-number/QH7BB27ABTKNLM6BUUU3IED56O/graph.json","events_json":"https://pith.science/api/pith-number/QH7BB27ABTKNLM6BUUU3IED56O/events.json","paper":"https://pith.science/paper/QH7BB27A"},"agent_actions":{"view_html":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O","download_json":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O.json","view_paper":"https://pith.science/paper/QH7BB27A","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2606.10722&json=true","fetch_graph":"https://pith.science/api/pith-number/QH7BB27ABTKNLM6BUUU3IED56O/graph.json","fetch_events":"https://pith.science/api/pith-number/QH7BB27ABTKNLM6BUUU3IED56O/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O/action/timestamp_anchor","attest_storage":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O/action/storage_attestation","attest_author":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O/action/author_attestation","sign_citation":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O/action/citation_signature","submit_replication":"https://pith.science/pith/QH7BB27ABTKNLM6BUUU3IED56O/action/replication_record"}},"created_at":"2026-06-10T01:10:36.486332+00:00","updated_at":"2026-06-10T01:10:36.486332+00:00"}