{"id":"367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d","title":"arXiv cs.AI: Researchers propose a hybrid memory design for sequence models","argument":"Yannis Bendi-Ouis and Xavier Hinaut present the Dynamical Memory Transformer (DMT), which separates sequence encoding from information reading. Its memory units update recurrent states, and a decoder uses attention to read their outputs without going through token history. The authors say the design keeps recurrent states and the convolution cache fixed in size during autoregressive inference, while its recurrence allows an associative scan with O(T ) total work and O(log T ) parallel depth, with fixed dimensions.\n\nIn the authors’ comparisons on cognitive and language modeling tasks, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling. On selective copy, they report stronger length extrapolation than Mamba and Transformer. The work’s central trade-off is between Transformer’s explicit history, whose KV-Cache and new-token cost grow with context length, and State Space Models’ fixed-size recurrent state, which the authors say comes at the cost of reduced addressability. Their results suggest a different way to retain and retrieve sequence information, but the abstract does not establish how these task results translate to other settings.\n\nHow does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The reported comparisons are limited to cognitive and language modeling tasks, with a specific result on selective copy; the abstract does not establish performance in other settings. The abstract does not provide experiment details or results that would show how these task comparisons translate to practical deployments.","speaker":{"name":"Bendi-Ouis, Yannis, Hinaut, Xavier","roleAtSource":"Article author"},"source":{"title":"Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.41850","publishedAt":"2026-10-30","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors present DMT as a hybrid architecture that separates sequence encoding from information reading.","excerpt":"We present the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"The DMT decoder uses attention to read and combine memory-unit outputs without going through token history.","excerpt":"The decoder then uses attention to read and combine their outputs, without going through the token history.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"The authors state that recurrent states and the convolution cache remain fixed in size during autoregressive inference.","excerpt":"During autoregressive inference, the recurrent states and the convolution cache remain fixed in size.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"In the reported comparisons, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling.","excerpt":"DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"On selective copy, the authors report stronger length extrapolation for DMT than for Mamba and Transformer.","excerpt":"On selective copy, DMT also shows stronger length extrapolation than Mamba and Transformer.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"The authors describe a trade-off: Transformers retain explicit sequence history, while State Space Models compress the past into a fixed-size recurrent state at the cost of reduced addressability.","excerpt":"State Space Models compress the past into a fixed-size recurrent state, at the cost of reduced addressability.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.41850","checkedAt":"2026-11-11T00:58:10.251Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Researchers propose a hybrid memory design for sequence models","summary":"Yannis Bendi-Ouis and Xavier Hinaut present the Dynamical Memory Transformer (DMT), which separates sequence encoding from information reading. Its memory units update recurrent states, and a decoder uses attention to read their outputs without going through token history. The authors say the design keeps recurrent states and the convolution cache fixed in size during autoregressive inference, while its recurrence allows an associative scan with O(T ) total work and O(log T ) parallel depth, with fixed dimensions.\n\nIn the authors’ comparisons on cognitive and language modeling tasks, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling. On selective copy, they report stronger length extrapolation than Mamba and Transformer. The work’s central trade-off is between Transformer’s explicit history, whose KV-Cache and new-token cost grow with context length, and State Space Models’ fixed-size recurrent state, which the authors say comes at the cost of reduced addressability. Their results suggest a different way to retain and retrieve sequence information, but the abstract does not establish how these task results translate to other settings.\n\nHow does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?","question":"How does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?","claims":[{"text":"The authors present DMT as a hybrid architecture that separates sequence encoding from information reading.","excerpt":"We present the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading.","support":"source-stated"},{"text":"The DMT decoder uses attention to read and combine memory-unit outputs without going through token history.","excerpt":"The decoder then uses attention to read and combine their outputs, without going through the token history.","support":"source-stated"},{"text":"The authors state that recurrent states and the convolution cache remain fixed in size during autoregressive inference.","excerpt":"During autoregressive inference, the recurrent states and the convolution cache remain fixed in size.","support":"source-stated"},{"text":"In the reported comparisons, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling.","excerpt":"DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling.","support":"source-stated"},{"text":"On selective copy, the authors report stronger length extrapolation for DMT than for Mamba and Transformer.","excerpt":"On selective copy, DMT also shows stronger length extrapolation than Mamba and Transformer.","support":"source-stated"},{"text":"The authors describe a trade-off: Transformers retain explicit sequence history, while State Space Models compress the past into a fixed-size recurrent state at the cost of reduced addressability.","excerpt":"State Space Models compress the past into a fixed-size recurrent state, at the cost of reduced addressability.","support":"source-stated"}],"limitations":["The reported comparisons are limited to cognitive and language modeling tasks, with a specific result on selective copy; the abstract does not establish performance in other settings.","The abstract does not provide experiment details or results that would show how these task comparisons translate to practical deployments."],"mediaId":"b72e835264ed1b9d4a434b4b71c481c3f2b66132c885ceeb848a9f06ae023304","correctionReason":null,"id":"367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d","publicationId":"a8723b51-9d83-4def-861c-59e457a2f896","version":1,"sourceUrl":"https://arxiv.org/abs/2610.41850","sourceTitle":"Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-30","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-02T05:01:34.768Z","createdAt":"2026-11-11T00:26:54.104Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b72e835264ed1b9d4a434b4b71c481c3f2b66132c885ceeb848a9f06ae023304","kind":"article","url":"https://arxiv.org/abs/2610.41850","sourceUrl":"https://arxiv.org/abs/2610.41850","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-02T10:17:34.889Z","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/2610.41850","checkedAt":"2026-11-11T00:58:10.251Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"b72e835264ed1b9d4a434b4b71c481c3f2b66132c885ceeb848a9f06ae023304","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"a9ebf30396636dd29b8c8f40f9dca6631327644d25bffe3b89faf468c8bd728c","decidedAt":"2026-11-02T10:17:36.595Z"},"provenance":{"textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observationId":"2a4fd5954388db55ab069e22506b351de9e2830d395088060a915fcdc2645bf1","retrievedAt":"2026-11-02T10:17:34.644Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"44c9ba5a-2b0d-43a6-85a0-0a2013e8ba72"},"reviewer":{"model":"gpt-6-luna","runId":"183a0a6f-3749-4c42-82fe-96d39aa64971"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T00:26:54.104Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Bendi-Ouis, Yannis","url":null,"kind":"unknown"},{"name":"Hinaut, Xavier","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Bendi-Ouis, Yannis","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Bendi-Ouis, Yannis","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.41850","locator":"head/meta[name=author|citation_author]"},"assertionId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T00:58:10.251Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd","sourceUrl":"https://arxiv.org/abs/2610.41850","contentSha256":"39b296970252e81d4a93a79593043cd5d3ed07fc2829b1787a283784a3059f80","textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observedAt":"2026-11-11T00:58:10.251Z"}},{"actor":{"name":"Hinaut, Xavier","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hinaut, Xavier","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.41850","locator":"head/meta[name=author|citation_author]"},"assertionId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T00:58:10.251Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd","sourceUrl":"https://arxiv.org/abs/2610.41850","contentSha256":"39b296970252e81d4a93a79593043cd5d3ed07fc2829b1787a283784a3059f80","textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observedAt":"2026-11-11T00:58:10.251Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.41850","snapshotId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd","contentSha256":"39b296970252e81d4a93a79593043cd5d3ed07fc2829b1787a283784a3059f80","textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observedAt":"2026-11-11T00:58:10.251Z","basis":"publisher-page-metadata","sourceLinks":38,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-cs-ai-researchers-propose-a-hybrid-memory-design-for-367213a1","formats":{"html":"/readings/arxiv-cs-ai-researchers-propose-a-hybrid-memory-design-for-367213a1","markdown":"/readings/arxiv-cs-ai-researchers-propose-a-hybrid-memory-design-for-367213a1.md","json":"/readings/arxiv-cs-ai-researchers-propose-a-hybrid-memory-design-for-367213a1.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-context-window","name":"Context window","explanation":"How much text a model can take in and consider at one time.","referenceUrl":"https://developers.google.com/machine-learning/glossary#context-window"}],"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"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d","revision":"a8723b51-9d83-4def-861c-59e457a2f896","revisionKind":"publication-id","title":"arXiv cs.AI: Researchers propose a hybrid memory design for sequence models","url":"https://escapemeridian.com/readings/arxiv-cs-ai-researchers-propose-a-hybrid-memory-design-for-367213a1"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.41850","revision":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","revisionKind":"sha256-source-text","title":"Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention","url":"https://arxiv.org/abs/2610.41850"},"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":"367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d","title":"arXiv cs.AI: Researchers propose a hybrid memory design for sequence models","argument":"Yannis Bendi-Ouis and Xavier Hinaut present the Dynamical Memory Transformer (DMT), which separates sequence encoding from information reading. Its memory units update recurrent states, and a decoder uses attention to read their outputs without going through token history. The authors say the design keeps recurrent states and the convolution cache fixed in size during autoregressive inference, while its recurrence allows an associative scan with O(T ) total work and O(log T ) parallel depth, with fixed dimensions.\n\nIn the authors’ comparisons on cognitive and language modeling tasks, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling. On selective copy, they report stronger length extrapolation than Mamba and Transformer. The work’s central trade-off is between Transformer’s explicit history, whose KV-Cache and new-token cost grow with context length, and State Space Models’ fixed-size recurrent state, which the authors say comes at the cost of reduced addressability. Their results suggest a different way to retain and retrieve sequence information, but the abstract does not establish how these task results translate to other settings.\n\nHow does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The reported comparisons are limited to cognitive and language modeling tasks, with a specific result on selective copy; the abstract does not establish performance in other settings. The abstract does not provide experiment details or results that would show how these task comparisons translate to practical deployments.","speaker":{"name":"Bendi-Ouis, Yannis, Hinaut, Xavier","roleAtSource":"Article author"},"source":{"title":"Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.41850","publishedAt":"2026-10-30","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors present DMT as a hybrid architecture that separates sequence encoding from information reading.","excerpt":"We present the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"The DMT decoder uses attention to read and combine memory-unit outputs without going through token history.","excerpt":"The decoder then uses attention to read and combine their outputs, without going through the token history.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"The authors state that recurrent states and the convolution cache remain fixed in size during autoregressive inference.","excerpt":"During autoregressive inference, the recurrent states and the convolution cache remain fixed in size.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"In the reported comparisons, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling.","excerpt":"DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"On selective copy, the authors report stronger length extrapolation for DMT than for Mamba and Transformer.","excerpt":"On selective copy, DMT also shows stronger length extrapolation than Mamba and Transformer.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"},{"text":"The authors describe a trade-off: Transformers retain explicit sequence history, while State Space Models compress the past into a fixed-size recurrent state at the cost of reduced addressability.","excerpt":"State Space Models compress the past into a fixed-size recurrent state, at the cost of reduced addressability.","url":"https://arxiv.org/abs/2610.41850","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.41850","checkedAt":"2026-11-11T00:58:10.251Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Researchers propose a hybrid memory design for sequence models","summary":"Yannis Bendi-Ouis and Xavier Hinaut present the Dynamical Memory Transformer (DMT), which separates sequence encoding from information reading. Its memory units update recurrent states, and a decoder uses attention to read their outputs without going through token history. The authors say the design keeps recurrent states and the convolution cache fixed in size during autoregressive inference, while its recurrence allows an associative scan with O(T ) total work and O(log T ) parallel depth, with fixed dimensions.\n\nIn the authors’ comparisons on cognitive and language modeling tasks, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling. On selective copy, they report stronger length extrapolation than Mamba and Transformer. The work’s central trade-off is between Transformer’s explicit history, whose KV-Cache and new-token cost grow with context length, and State Space Models’ fixed-size recurrent state, which the authors say comes at the cost of reduced addressability. Their results suggest a different way to retain and retrieve sequence information, but the abstract does not establish how these task results translate to other settings.\n\nHow does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?","question":"How does DMT perform beyond the cognitive, language modeling, and selective copy tasks reported here?","claims":[{"text":"The authors present DMT as a hybrid architecture that separates sequence encoding from information reading.","excerpt":"We present the Dynamical Memory Transformer (DMT), a hybrid architecture that separates sequence encoding from information reading.","support":"source-stated"},{"text":"The DMT decoder uses attention to read and combine memory-unit outputs without going through token history.","excerpt":"The decoder then uses attention to read and combine their outputs, without going through the token history.","support":"source-stated"},{"text":"The authors state that recurrent states and the convolution cache remain fixed in size during autoregressive inference.","excerpt":"During autoregressive inference, the recurrent states and the convolution cache remain fixed in size.","support":"source-stated"},{"text":"In the reported comparisons, DMT performs best on several cognitive tasks and has language modeling performance close to Transformer, but remains behind Mamba in language modeling.","excerpt":"DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling.","support":"source-stated"},{"text":"On selective copy, the authors report stronger length extrapolation for DMT than for Mamba and Transformer.","excerpt":"On selective copy, DMT also shows stronger length extrapolation than Mamba and Transformer.","support":"source-stated"},{"text":"The authors describe a trade-off: Transformers retain explicit sequence history, while State Space Models compress the past into a fixed-size recurrent state at the cost of reduced addressability.","excerpt":"State Space Models compress the past into a fixed-size recurrent state, at the cost of reduced addressability.","support":"source-stated"}],"limitations":["The reported comparisons are limited to cognitive and language modeling tasks, with a specific result on selective copy; the abstract does not establish performance in other settings.","The abstract does not provide experiment details or results that would show how these task comparisons translate to practical deployments."],"mediaId":"b72e835264ed1b9d4a434b4b71c481c3f2b66132c885ceeb848a9f06ae023304","correctionReason":null,"id":"367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d","publicationId":"a8723b51-9d83-4def-861c-59e457a2f896","version":1,"sourceUrl":"https://arxiv.org/abs/2610.41850","sourceTitle":"Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-30","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-02T05:01:34.768Z","createdAt":"2026-11-11T00:26:54.104Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b72e835264ed1b9d4a434b4b71c481c3f2b66132c885ceeb848a9f06ae023304","kind":"article","url":"https://arxiv.org/abs/2610.41850","sourceUrl":"https://arxiv.org/abs/2610.41850","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-02T10:17:34.889Z","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/2610.41850","checkedAt":"2026-11-11T00:58:10.251Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"b72e835264ed1b9d4a434b4b71c481c3f2b66132c885ceeb848a9f06ae023304","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"a9ebf30396636dd29b8c8f40f9dca6631327644d25bffe3b89faf468c8bd728c","decidedAt":"2026-11-02T10:17:36.595Z"},"provenance":{"textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observationId":"2a4fd5954388db55ab069e22506b351de9e2830d395088060a915fcdc2645bf1","retrievedAt":"2026-11-02T10:17:34.644Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"44c9ba5a-2b0d-43a6-85a0-0a2013e8ba72"},"reviewer":{"model":"gpt-6-luna","runId":"183a0a6f-3749-4c42-82fe-96d39aa64971"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T00:26:54.104Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Bendi-Ouis, Yannis","url":null,"kind":"unknown"},{"name":"Hinaut, Xavier","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Bendi-Ouis, Yannis","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Bendi-Ouis, Yannis","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.41850","locator":"head/meta[name=author|citation_author]"},"assertionId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T00:58:10.251Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd","sourceUrl":"https://arxiv.org/abs/2610.41850","contentSha256":"39b296970252e81d4a93a79593043cd5d3ed07fc2829b1787a283784a3059f80","textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observedAt":"2026-11-11T00:58:10.251Z"}},{"actor":{"name":"Hinaut, Xavier","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hinaut, Xavier","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.41850","locator":"head/meta[name=author|citation_author]"},"assertionId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T00:58:10.251Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd","sourceUrl":"https://arxiv.org/abs/2610.41850","contentSha256":"39b296970252e81d4a93a79593043cd5d3ed07fc2829b1787a283784a3059f80","textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observedAt":"2026-11-11T00:58:10.251Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.41850","snapshotId":"4ada2e85d36c18cf9451a843c62fd3ab744843df0f3850149d1d9e6b0b3739bd","contentSha256":"39b296970252e81d4a93a79593043cd5d3ed07fc2829b1787a283784a3059f80","textHash":"4532edd599ee56e41b3a071bdba3687216ed4c0b185e2f53c45d0cf76a6d9fb2","observedAt":"2026-11-11T00:58:10.251Z","basis":"publisher-page-metadata","sourceLinks":38,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/367213a1f5676dc7e2bf158346ec9217b6dae025535665bb074b0637c6d88a1d/source"}}}],"audits":[{"publicationId":"a8723b51-9d83-4def-861c-59e457a2f896","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly credits Yannis Bendi-Ouis and Xavier Hinaut on arXiv cs.AI. Performance results are framed as the authors' reported comparisons, not as established fact.","excerpt":"Authors:Yannis Bendi-Ouis (Mnemosyne), Xavier Hinaut (Mnemosyne)"},{"dimension":"summary","status":"uncertain","explanation":"The draft is faithful, but a non-specialist would struggle with it. Several central terms are not explained in plain language: KV-Cache, State Space Models, Mamba, autoregressive inference, associative scan, O(T)/O(log T), selective copy and length extrapolation. The draft also never states the practical stake: a Transformer's memory and per-token cost grow with context, while DMT is designed to keep its state fixed in size. Smallest repair: a plain sentence on that cost and memory point, plus tooltips for the terms. The draft also softens the authors' 'demonstrate ... promising' to 'suggest'; that is acceptable caution, but it should note that this is the authors' claim.","excerpt":"However, the size of the KV-Cache and the cost of a new token increase with the context length."},{"dimension":"date","status":"pass","explanation":"The originalPublished date of 2026-10-30 matches the arXiv v1 submission date. The feed date of 2026-11-02 is later and is correctly not used.","excerpt":"[Submitted on 30 Oct 2026]"},{"dimension":"support","status":"pass","explanation":"Every claim, including the mixed results (behind Mamba in language modeling), is quoted from the abstract. The limitations and the question stay within the evidence, and the draft adds no unsupported benefits or costs.","excerpt":"DMT achieves the best results on several cognitive tasks and language modeling performance close to that of Transformer, but remains behind Mamba in language modeling."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-11T00:26:54.104Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}