{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:AWRVZGGYQ4WTIPW3335OIQDHNM","short_pith_number":"pith:AWRVZGGY","schema_version":"1.0","canonical_sha256":"05a35c98d8872d343edbdefae440676b0842cccfd4e2cb485a5d7898d3470bd6","source":{"kind":"arxiv","id":"2604.01622","version":2},"attestation_state":"computed","paper":{"title":"Expert-Choice Routing Enables Adaptive Computation in Diffusion Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Caspian Zhuang, Chihan Cui, Fred Zhangzhi Peng, Guanhua Chen, Haoyue Bai, Ming Liu, Shuibai Zhang, Yang Zhou, Yanxin Zhang, Zack Jia, Zhihan Yang","submitted_at":"2026-04-02T05:01:36Z","abstract_excerpt":"Diffusion language models (DLMs) enable parallel, non-autoregressive text generation, yet existing DLM mixture-of-experts (MoE) models inherit token-choice (TC) routing from autoregressive systems, leading to load imbalance and rigid computation allocation. We show that expert-choice (EC) routing is a better fit for DLMs: it provides deterministic load balancing by design, yielding higher throughput and faster convergence than TC. Building on the property that EC capacity is externally controllable, we introduce timestep-dependent expert capacity, which varies expert allocation according to th"},"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":"2604.01622","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.LG","submitted_at":"2026-04-02T05:01:36Z","cross_cats_sorted":["cs.CL"],"title_canon_sha256":"765e01f6558571bdb1e94304382bcbbc2ee88c1d58e9ffcf97c7a65a377c7f93","abstract_canon_sha256":"7a08b53a320ba71509b675527322a15861d77647e1fa36547ef0833bc6e98c81"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-08-04T02:40:31.943427Z","signature_b64":"cJ01+4K87OFE5FbYCpVCfO3WsiLSgvPr4D8KkDpK7T8Mxo1ZLccrnkdqVy1bcuqCVxqIyP4ZDxAYJsgBLBTSDg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"05a35c98d8872d343edbdefae440676b0842cccfd4e2cb485a5d7898d3470bd6","last_reissued_at":"2026-08-04T02:40:31.940861Z","signature_status":"signed_v1","first_computed_at":"2026-08-04T02:40:31.940861Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Expert-Choice Routing Enables Adaptive Computation in Diffusion Language Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":["cs.CL"],"primary_cat":"cs.LG","authors_text":"Caspian Zhuang, Chihan Cui, Fred Zhangzhi Peng, Guanhua Chen, Haoyue Bai, Ming Liu, Shuibai Zhang, Yang Zhou, Yanxin Zhang, Zack Jia, Zhihan Yang","submitted_at":"2026-04-02T05:01:36Z","abstract_excerpt":"Diffusion language models (DLMs) enable parallel, non-autoregressive text generation, yet existing DLM mixture-of-experts (MoE) models inherit token-choice (TC) routing from autoregressive systems, leading to load imbalance and rigid computation allocation. We show that expert-choice (EC) routing is a better fit for DLMs: it provides deterministic load balancing by design, yielding higher throughput and faster convergence than TC. Building on the property that EC capacity is externally controllable, we introduce timestep-dependent expert capacity, which varies expert allocation according to th"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2604.01622","kind":"arxiv","version":2},"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/2604.01622/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":"2604.01622","created_at":"2026-08-04T02:40:31.943119+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.01622v2","created_at":"2026-08-04T02:40:31.943119+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.01622","created_at":"2026-08-04T02:40:31.943119+00:00"},{"alias_kind":"pith_short_12","alias_value":"AWRVZGGYQ4WT","created_at":"2026-08-04T02:40:31.943119+00:00"},{"alias_kind":"pith_short_16","alias_value":"AWRVZGGYQ4WTIPW3","created_at":"2026-08-04T02:40:31.943119+00:00"},{"alias_kind":"pith_short_8","alias_value":"AWRVZGGY","created_at":"2026-08-04T02:40:31.943119+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2605.31215","citing_title":"Fixed-Point Masked Generative Modeling","ref_index":75,"is_internal_anchor":true},{"citing_arxiv_id":"2605.30876","citing_title":"dMoE: dLLMs with Learnable Block Experts","ref_index":31,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM","json":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM.json","graph_json":"https://pith.science/api/pith-number/AWRVZGGYQ4WTIPW3335OIQDHNM/graph.json","events_json":"https://pith.science/api/pith-number/AWRVZGGYQ4WTIPW3335OIQDHNM/events.json","paper":"https://pith.science/paper/AWRVZGGY"},"agent_actions":{"view_html":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM","download_json":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM.json","view_paper":"https://pith.science/paper/AWRVZGGY","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.01622&json=true","fetch_graph":"https://pith.science/api/pith-number/AWRVZGGYQ4WTIPW3335OIQDHNM/graph.json","fetch_events":"https://pith.science/api/pith-number/AWRVZGGYQ4WTIPW3335OIQDHNM/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM/action/timestamp_anchor","attest_storage":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM/action/storage_attestation","attest_author":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM/action/author_attestation","sign_citation":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM/action/citation_signature","submit_replication":"https://pith.science/pith/AWRVZGGYQ4WTIPW3335OIQDHNM/action/replication_record"}},"created_at":"2026-08-04T02:40:31.943119+00:00","updated_at":"2026-08-04T02:40:31.943119+00:00"}