{"id":"e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5","title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","argument":"In their paper, Jiaming Hu and coauthors introduce LRMoE, a system for training MoE models on memory-constrained nodes. They argue that uneven routing creates computation and communication bottlenecks, while copying experts to redistribute work competes with token-related buffers for limited GPU memory.\n\nLRMoE addresses this trade-off by planning how much work to retain and where to place experts, then selecting token-expert pairs according to their estimated contribution to expert updates. It distributes the retained pairs among eligible instances of their original experts and compacts them into smaller communication payloads and variable-size grouped GEMMs.\n\nThe authors report that, on Qwen3-30B-A3B and GLM4.5-106B-A12B, LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2–44.8% relative to SYMI. This result is limited to the models and comparison described in the abstract; it does not establish performance on other workloads or systems. How does LRMoE’s reported advantage change across workloads beyond these two models?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence provided is the arXiv abstract; it does not include experimental details needed to assess the reported results. The abstract reports results for Qwen3-30B-A3B and GLM4.5-106B-A12B relative to SYMI; it does not establish performance on other workloads or systems.","speaker":{"name":"Hu, Jiaming, Jiang, Zhida, Xing, Zhaolong, Li, Yu, Chen, Zhen, Zhang, Ke","roleAtSource":"Article author"},"source":{"title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.37720","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors say uneven routing in MoE models creates computation and communication bottlenecks, and that expert copies compete with token-related buffers for limited GPU memory.","excerpt":"uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory.","url":"https://arxiv.org/abs/2610.37720","relationship":"Source-stated"},{"text":"LRMoE plans retention quotas and expert placement under resource constraints, then selects token-expert pairs based on their estimated contribution to expert updates.","excerpt":"During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas","url":"https://arxiv.org/abs/2610.37720","relationship":"Source-stated"},{"text":"The authors report reduced cumulative MoE-layer training time to fixed-loss targets on two named models, compared with SYMI.","excerpt":"Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.","url":"https://arxiv.org/abs/2610.37720","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.37720","checkedAt":"2026-11-13T05:44:07.433Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does LRMoE’s reported advantage change across workloads beyond these two models?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","summary":"In their paper, Jiaming Hu and coauthors introduce LRMoE, a system for training MoE models on memory-constrained nodes. They argue that uneven routing creates computation and communication bottlenecks, while copying experts to redistribute work competes with token-related buffers for limited GPU memory.\n\nLRMoE addresses this trade-off by planning how much work to retain and where to place experts, then selecting token-expert pairs according to their estimated contribution to expert updates. It distributes the retained pairs among eligible instances of their original experts and compacts them into smaller communication payloads and variable-size grouped GEMMs.\n\nThe authors report that, on Qwen3-30B-A3B and GLM4.5-106B-A12B, LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2–44.8% relative to SYMI. This result is limited to the models and comparison described in the abstract; it does not establish performance on other workloads or systems. How does LRMoE’s reported advantage change across workloads beyond these two models?","question":"How does LRMoE’s reported advantage change across workloads beyond these two models?","claims":[{"text":"The authors say uneven routing in MoE models creates computation and communication bottlenecks, and that expert copies compete with token-related buffers for limited GPU memory.","excerpt":"uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory.","support":"source-stated"},{"text":"LRMoE plans retention quotas and expert placement under resource constraints, then selects token-expert pairs based on their estimated contribution to expert updates.","excerpt":"During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas","support":"source-stated"},{"text":"The authors report reduced cumulative MoE-layer training time to fixed-loss targets on two named models, compared with SYMI.","excerpt":"Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.","support":"source-stated"}],"limitations":["The evidence provided is the arXiv abstract; it does not include experimental details needed to assess the reported results.","The abstract reports results for Qwen3-30B-A3B and GLM4.5-106B-A12B relative to SYMI; it does not establish performance on other workloads or systems."],"mediaId":"2e338f4975c6ee042842bab23dc6a4873694452bbccec4edaa0cb395ae90f91d","correctionReason":null,"id":"e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5","publicationId":"e5d380e7-f028-4737-8a41-85a0a41941c4","version":1,"sourceUrl":"https://arxiv.org/abs/2610.37720","sourceTitle":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-30T06:45:25.392Z","createdAt":"2026-11-13T03:36:26.567Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"2e338f4975c6ee042842bab23dc6a4873694452bbccec4edaa0cb395ae90f91d","kind":"article","url":"https://arxiv.org/abs/2610.37720","sourceUrl":"https://arxiv.org/abs/2610.37720","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-31T11:48:29.215Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.37720","checkedAt":"2026-11-13T05:44:07.433Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"2e338f4975c6ee042842bab23dc6a4873694452bbccec4edaa0cb395ae90f91d","confidence":0.85,"provider":"typesafe","model":"jev-1.13.0","inputHash":"a88051524c15e8995879ff07abd00849743cc2201d060f907229c4ac00e6e43f","decidedAt":"2026-10-31T11:48:31.125Z"},"provenance":{"textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observationId":"689a0633edb907d9ecef7ba3af1ca08c2a4d7502b9041b9834f67a23c9543280","retrievedAt":"2026-10-31T11:48:28.975Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"e755a17c-a9cb-4dcb-85db-44feab0ca8db"},"reviewer":{"model":"gpt-6-luna","runId":"14c0c9e0-4c42-41f4-8119-f8f52e2d6404"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-13T03:36:26.567Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Hu, Jiaming","url":null,"kind":"unknown"},{"name":"Jiang, Zhida","url":null,"kind":"unknown"},{"name":"Xing, Zhaolong","url":null,"kind":"unknown"},{"name":"Li, Yu","url":null,"kind":"unknown"},{"name":"Chen, Zhen","url":null,"kind":"unknown"},{"name":"Zhang, Ke","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Hu, Jiaming","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hu, Jiaming","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Jiang, Zhida","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Jiang, Zhida","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Xing, Zhaolong","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Xing, Zhaolong","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Li, Yu","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Li, Yu","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Chen, Zhen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chen, Zhen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Zhang, Ke","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhang, Ke","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.37720","snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z","basis":"publisher-page-metadata","sourceLinks":46,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5/source"}},"canonicalUrl":"https://escapemeridian.com/readings/lrmoe-demand-side-load-re-calibration-for-efficient-moe-e410f37b","formats":{"html":"/readings/lrmoe-demand-side-load-re-calibration-for-efficient-moe-e410f37b","markdown":"/readings/lrmoe-demand-side-load-re-calibration-for-efficient-moe-e410f37b.md","json":"/readings/lrmoe-demand-side-load-re-calibration-for-efficient-moe-e410f37b.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-mixture-of-experts","name":"Mixture of experts","explanation":"A model design that sends each input to a few specialized sub-networks instead of the whole model, which saves computation.","referenceUrl":"https://developers.google.com/machine-learning/glossary#mixture-of-experts"},{"id":"term-gpu","name":"Graphics processing unit","explanation":"A chip that performs many calculations at once; the main hardware used to train and run AI models.","referenceUrl":"https://en.wikipedia.org/wiki/Graphics_processing_unit"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5","revision":"e5d380e7-f028-4737-8a41-85a0a41941c4","revisionKind":"publication-id","title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","url":"https://escapemeridian.com/readings/lrmoe-demand-side-load-re-calibration-for-efficient-moe-e410f37b"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.37720","revision":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","revisionKind":"sha256-source-text","title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","url":"https://arxiv.org/abs/2610.37720"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5","title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","argument":"In their paper, Jiaming Hu and coauthors introduce LRMoE, a system for training MoE models on memory-constrained nodes. They argue that uneven routing creates computation and communication bottlenecks, while copying experts to redistribute work competes with token-related buffers for limited GPU memory.\n\nLRMoE addresses this trade-off by planning how much work to retain and where to place experts, then selecting token-expert pairs according to their estimated contribution to expert updates. It distributes the retained pairs among eligible instances of their original experts and compacts them into smaller communication payloads and variable-size grouped GEMMs.\n\nThe authors report that, on Qwen3-30B-A3B and GLM4.5-106B-A12B, LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2–44.8% relative to SYMI. This result is limited to the models and comparison described in the abstract; it does not establish performance on other workloads or systems. How does LRMoE’s reported advantage change across workloads beyond these two models?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence provided is the arXiv abstract; it does not include experimental details needed to assess the reported results. The abstract reports results for Qwen3-30B-A3B and GLM4.5-106B-A12B relative to SYMI; it does not establish performance on other workloads or systems.","speaker":{"name":"Hu, Jiaming, Jiang, Zhida, Xing, Zhaolong, Li, Yu, Chen, Zhen, Zhang, Ke","roleAtSource":"Article author"},"source":{"title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.37720","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors say uneven routing in MoE models creates computation and communication bottlenecks, and that expert copies compete with token-related buffers for limited GPU memory.","excerpt":"uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory.","url":"https://arxiv.org/abs/2610.37720","relationship":"Source-stated"},{"text":"LRMoE plans retention quotas and expert placement under resource constraints, then selects token-expert pairs based on their estimated contribution to expert updates.","excerpt":"During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas","url":"https://arxiv.org/abs/2610.37720","relationship":"Source-stated"},{"text":"The authors report reduced cumulative MoE-layer training time to fixed-loss targets on two named models, compared with SYMI.","excerpt":"Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.","url":"https://arxiv.org/abs/2610.37720","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.37720","checkedAt":"2026-11-13T05:44:07.433Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does LRMoE’s reported advantage change across workloads beyond these two models?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","summary":"In their paper, Jiaming Hu and coauthors introduce LRMoE, a system for training MoE models on memory-constrained nodes. They argue that uneven routing creates computation and communication bottlenecks, while copying experts to redistribute work competes with token-related buffers for limited GPU memory.\n\nLRMoE addresses this trade-off by planning how much work to retain and where to place experts, then selecting token-expert pairs according to their estimated contribution to expert updates. It distributes the retained pairs among eligible instances of their original experts and compacts them into smaller communication payloads and variable-size grouped GEMMs.\n\nThe authors report that, on Qwen3-30B-A3B and GLM4.5-106B-A12B, LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2–44.8% relative to SYMI. This result is limited to the models and comparison described in the abstract; it does not establish performance on other workloads or systems. How does LRMoE’s reported advantage change across workloads beyond these two models?","question":"How does LRMoE’s reported advantage change across workloads beyond these two models?","claims":[{"text":"The authors say uneven routing in MoE models creates computation and communication bottlenecks, and that expert copies compete with token-related buffers for limited GPU memory.","excerpt":"uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory.","support":"source-stated"},{"text":"LRMoE plans retention quotas and expert placement under resource constraints, then selects token-expert pairs based on their estimated contribution to expert updates.","excerpt":"During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas","support":"source-stated"},{"text":"The authors report reduced cumulative MoE-layer training time to fixed-loss targets on two named models, compared with SYMI.","excerpt":"Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.","support":"source-stated"}],"limitations":["The evidence provided is the arXiv abstract; it does not include experimental details needed to assess the reported results.","The abstract reports results for Qwen3-30B-A3B and GLM4.5-106B-A12B relative to SYMI; it does not establish performance on other workloads or systems."],"mediaId":"2e338f4975c6ee042842bab23dc6a4873694452bbccec4edaa0cb395ae90f91d","correctionReason":null,"id":"e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5","publicationId":"e5d380e7-f028-4737-8a41-85a0a41941c4","version":1,"sourceUrl":"https://arxiv.org/abs/2610.37720","sourceTitle":"LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-30T06:45:25.392Z","createdAt":"2026-11-13T03:36:26.567Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"2e338f4975c6ee042842bab23dc6a4873694452bbccec4edaa0cb395ae90f91d","kind":"article","url":"https://arxiv.org/abs/2610.37720","sourceUrl":"https://arxiv.org/abs/2610.37720","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-31T11:48:29.215Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.37720","checkedAt":"2026-11-13T05:44:07.433Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"2e338f4975c6ee042842bab23dc6a4873694452bbccec4edaa0cb395ae90f91d","confidence":0.85,"provider":"typesafe","model":"jev-1.13.0","inputHash":"a88051524c15e8995879ff07abd00849743cc2201d060f907229c4ac00e6e43f","decidedAt":"2026-10-31T11:48:31.125Z"},"provenance":{"textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observationId":"689a0633edb907d9ecef7ba3af1ca08c2a4d7502b9041b9834f67a23c9543280","retrievedAt":"2026-10-31T11:48:28.975Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"e755a17c-a9cb-4dcb-85db-44feab0ca8db"},"reviewer":{"model":"gpt-6-luna","runId":"14c0c9e0-4c42-41f4-8119-f8f52e2d6404"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-13T03:36:26.567Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Hu, Jiaming","url":null,"kind":"unknown"},{"name":"Jiang, Zhida","url":null,"kind":"unknown"},{"name":"Xing, Zhaolong","url":null,"kind":"unknown"},{"name":"Li, Yu","url":null,"kind":"unknown"},{"name":"Chen, Zhen","url":null,"kind":"unknown"},{"name":"Zhang, Ke","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Hu, Jiaming","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hu, Jiaming","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Jiang, Zhida","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Jiang, Zhida","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Xing, Zhaolong","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Xing, Zhaolong","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Li, Yu","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Li, Yu","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Chen, Zhen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chen, Zhen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}},{"actor":{"name":"Zhang, Ke","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhang, Ke","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37720","locator":"head/meta[name=author|citation_author]"},"assertionId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T05:44:07.433Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","sourceUrl":"https://arxiv.org/abs/2610.37720","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.37720","snapshotId":"484b905816e85c0377f7837bb6bb272a2227311a98ceea1403e8f39e21563ab9","contentSha256":"9fd902cf0ad04836f44212eb57c2e4191263aa89f2276504479ea44f57ef10f2","textHash":"5922419a972500953e306d3dea6a71c84fcf3d2dbb6599025d3eaac3290802fb","observedAt":"2026-11-13T05:44:07.433Z","basis":"publisher-page-metadata","sourceLinks":46,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/e410f37b8d7cfa31521511f8956caca9a85a47418b435b7f530ebbedf103c5f5/source"}}}],"audits":[{"publicationId":"e5d380e7-f028-4737-8a41-85a0a41941c4","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly credits the work to Jiaming Hu and coauthors on arXiv, and it frames the results as the authors' reports rather than as established facts.","excerpt":"Authors:Jiaming Hu, Zhida Jiang, Zhaolong Xing, Yu Li, Zhen Chen, Ke Zhang"},{"dimension":"summary","status":"uncertain","explanation":"A non-specialist cannot follow the central terms. MoE is never expanded to Mixture-of-Experts, and experts, routing, token-expert pairs, grouped GEMMs, fixed-loss targets and SYMI are not explained, so tooltips or brief glosses are needed. The draft also drops a key condition: all routed computations are kept in the forward pass, and token-expert pairs are selected only during backpropagation. Smallest repair: add that every forward computation is preserved and that selection trims only backward (update) work, which cuts traffic, computation and buffer memory.","excerpt":"plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation"},{"dimension":"date","status":"pass","explanation":"The originalPublished date, 2026-10-27, matches the v1 arXiv submission date. The later feed date does not affect this.","excerpt":"[Submitted on 27 Oct 2026]"},{"dimension":"support","status":"pass","explanation":"The claims, the 21.2–44.8% figure, the comparison with SYMI and the two named models all match the abstract. The draft adds no unsupported benefits or costs, and it correctly limits the results to these models and this baseline.","excerpt":"LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to reach the source.","excerpt":null}],"createdAt":"2026-11-13T03:36:26.567Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}