{ "profile": "wrs-published-research-view-1", "edition": "ratings-e039106f30dec177", "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS", "entry": { "id": "behrouz-et-al-nested-learning-hope-cms-2025", "title": "Nested Learning and Hope: memory components that update at different rates", "summary": "Some information should change quickly, while other knowledge should persist. Nested Learning organizes a model and its learning procedures as interacting optimization processes with different contexts and update rates. Its Continuum Memory System uses several memory components on different schedules. Hope combines that system with a self-referential extension of Titans, allowing parts of its memory processing to adapt while data arrive. The contribution here is the specified multi-timescale memory and Hope design with its reported computational demonstrations, beyond the earlier Titans and MIRAS families.", "boundary": "The authors explicitly do not claim that catastrophic forgetting is solved in general. Continual-learning experiments include substantial additional training; long-context results can require benchmark-specific fine-tuning. Transformers remain stronger on some short-recall tasks, and the multiscale optimizer has additional overhead. More memory levels do not establish unbounded knowledge retention. Biological analogies and clinical benchmarks are not separate discoveries in this record. The exact hosted source revision is retained; its 2025 conference family does not make every later revision identical.", "area": "Continual-learning architectures and optimization", "topic": "Artificial intelligence", "year": 2025, "lane": "HUMAN", "result": "ALGORITHM", "package_id": "behrouz-et-al-nested-learning-hope-cms-2025", "target": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "humans": [ "Ali Behrouz", "Meisam Razaviyayn", "Peilin Zhong", "Vahab Mirrokni" ], "authors_complete": true, "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.", "agents": [], "related": [ "behrouz-zhong-mirrokni-titans-testtime-memory-2024", "behrouz-et-al-miras-memory-objective-retention-2025" ], "subresults": [], "scientific_assessment": { "source_review": "SOURCE_CURATED", "scientific_review": "NOT_REVIEWED_BY_WRS", "formal_replay": "NOT_RUN", "provider_attestation": "NONE" }, "sources": [ { "id": "cur20261017-nested-learning-original", "title": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf", "date": "2025" } ], "points": "180000", "published_grade": { "id": "behrouz-et-al-nested-learning-hope-cms-2025", "entry_id": "behrouz-et-al-nested-learning-hope-cms-2025", "band": "A", "modifier": "+", "target": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "what_changed": "Some information should change quickly, while other knowledge should persist. Nested Learning organizes a model and its learning procedures as interacting optimization processes with different contexts and update rates. Its Continuum Memory System uses several memory components on different schedules. Hope combines that system with a self-referential extension of Titans, allowing parts of its memory processing to adapt while data arrive. The contribution here is the specified multi-timescale memory and Hope design with its reported computational demonstrations, beyond the earlier Titans and MIRAS families.", "scope": "Some information should change quickly, while other knowledge should persist. Nested Learning organizes a model and its learning procedures as interacting optimization processes with different contexts and update rates. Its Continuum Memory System uses several memory components on different schedules. Hope combines that system with a self-referential extension of Titans, allowing parts of its memory processing to adapt while data arrive. The contribution here is the specified multi-timescale memory and Hope design with its reported computational demonstrations, beyond the earlier Titans and MIRAS families. The authors explicitly do not claim that catastrophic forgetting is solved in general. Continual-learning experiments include substantial additional training; long-context results can require benchmark-specific fine-tuning. Transformers remain stronger on some short-recall tasks, and the multiscale optimizer has additional overhead. More memory levels do not establish unbounded knowledge retention. Biological analogies and clinical benchmarks are not separate discoveries in this record. The exact hosted source revision is retained; its 2025 conference family does not make every later revision identical.", "status": "PROVISIONAL_EDITORIAL", "scientific_correctness_certified": false }, "assessment_disposition": { "status": "ASSESSED", "label": "Provisionally assessed", "reason": "A current grade and separate reference score identify this exact contribution.", "target": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "edition": "ratings-e039106f30dec177" } }, "version": { "kind": "PUBLIC_RECORD_VIEW", "previous": null, "published_at": "2026-10-23T00:00:00Z", "publisher": { "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33", "role": "WRS_EDITORIAL_PUBLICATION" }, "record": { "contributors": [ { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Ali Behrouz", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Meisam Razaviyayn", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Peilin Zhong", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf" } }, { "attribution_basis": "SOURCE_CITED", "credit": "historical_scientific", "entity_kind": "historical_person", "name": "Vahab Mirrokni", "note": "Archive attribution; detailed role assignment remains source-bound", "roles": [ "formulator" ], "source": { "citation": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf" } } ], "evidence": [ { "description": "Bibliographic/stated-result source; not a WRS proof verification", "id": "source-0", "kind": "source_reference", "source": { "citation": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf" } } ], "origin": "HUMAN_LEGACY", "record_profile": "wrs-local-research-notes-0.1", "record_type": "research_note", "rrp_version": "0.1", "scope": { "description": "The authors explicitly do not claim that catastrophic forgetting is solved in general. Continual-learning experiments include substantial additional training; long-context results can require benchmark-specific fine-tuning. Transformers remain stronger on some short-recall tasks, and the multiscale optimizer has additional overhead. More memory levels do not establish unbounded knowledge retention. Biological analogies and clinical benchmarks are not separate discoveries in this record. The exact hosted source revision is retained; its 2025 conference family does not make every later revision identical.", "domain": "Continual-learning architectures and optimization" }, "sources": [ { "citation": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf" } ], "statement": "# Nested Learning and Hope: memory components that update at different rates\n\n## A project needs more than one clock\n\nA team may revise its next action every minute, its weekly plan each Friday and its long-term goals only occasionally. Nested Learning uses different update schedules inside a computational model. The analogy concerns schedules, not human thought. Its Continuum Memory System places several memory components on different clocks so that rapid adaptation and more persistent knowledge do not have to use exactly the same update rule.\n\n## What Hope adds to Titans\n\nTitans already changes a neural memory while processing input. Hope extends that design with components that adapt parts of the memory-processing procedure and with several memory timescales. A fast component can react to recent input while a slower component changes less often. The outer training process still matters: these components need suitable initial states and learning behavior. Attaching a label called slow memory does not make a randomly assembled model useful.\n\n## Count the updates, then count the cost\n\nIn a classroom scheduling example, a stream contains sixteen steps. One component updates every step and another updates every four steps. They make sixteen and four updates, respectively, for twenty in total. Two components that both update every step would make thirty-two. This is only an update count: the real cost also depends on each component’s size, its forward computations and the update algorithm. Twenty large updates could cost more than thirty-two small ones.\n\n## Slower change does not mean no forgetting\n\nThe paper explicitly says catastrophic forgetting is not solved in general. Its demonstrations include additional training and task-specific conditions. Some conventional Transformer results remain better on short recall, and the multiscale optimizer adds overhead. Ask your agent to draw a timeline for the sixteen-step example, then identify which experimental comparison in the paper actually isolates a memory schedule. A useful continuation states the tested conditions rather than promising a model that learns everything forever.\n\n## Leave the schedule visible\n\nYour small timeline should name both update intervals and show every update position. Record what the count excludes: component size, forward work and training. Another reader can then change an interval and check the result. Do not describe fewer counted updates as a measured speedup or a solution to forgetting.\n\n## Scientific scope\n\nThe authors explicitly do not claim that catastrophic forgetting is solved in general. Continual-learning experiments include substantial additional training; long-context results can require benchmark-specific fine-tuning. Transformers remain stronger on some short-recall tasks, and the multiscale optimizer has additional overhead. More memory levels do not establish unbounded knowledge retention. Biological analogies and clinical benchmarks are not separate discoveries in this record. The exact hosted source revision is retained; its 2025 conference family does not make every later revision identical.\n\n## Sources for this explanation\n\n- [Nested Learning: The Illusion of Deep Learning Architecture](https://abehrouz.github.io/files/NL.pdf): Pinned hosted revision: update-frequency definitions, sections7–8 CMS and Hope, section9 experiments and final explicit catastrophic-forgetting limitation; not an independent reconstruction of optimizer proofs.\n", "summary": "Some information should change quickly, while other knowledge should persist. Nested Learning organizes a model and its learning procedures as interacting optimization processes with different contexts and update rates. Its Continuum Memory System uses several memory components on different schedules. Hope combines that system with a self-referential extension of Titans, allowing parts of its memory processing to adapt while data arrive. The contribution here is the specified multi-timescale memory and Hope design with its reported computational demonstrations, beyond the earlier Titans and MIRAS families.", "title": "Nested Learning and Hope: memory components that update at different rates" }, "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "source_ref": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "export_view": "wrs-public-editorial-view-1", "public_view_hash": "sha256:6534e122fd7ca481cb6c4cfbabf9967819b4bc28f10f1272e115b16c9f46cd34" }, "target": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "published_grade": { "profile": "wrs-published-grade-1", "edition": "2026-10-28.1", "id": "behrouz-et-al-nested-learning-hope-cms-2025", "entry_id": "behrouz-et-al-nested-learning-hope-cms-2025", "revision": 1, "title": "Nested Learning and Hope: memory components that update at different rates", "context": "Continual-learning architectures and optimization", "track": "Reusable scientific method, experimental result or historical framework", "band": "A", "modifier": "+", "status": "PROVISIONAL_EDITORIAL", "assessed_on": "2026-10-23", "evidence_cutoff": "2026-10-23", "achievement_state": "SUPPORTED_SCOPED_RESULT", "what_changed": "Some information should change quickly, while other knowledge should persist. Nested Learning organizes a model and its learning procedures as interacting optimization processes with different contexts and update rates. Its Continuum Memory System uses several memory components on different schedules. Hope combines that system with a self-referential extension of Titans, allowing parts of its memory processing to adapt while data arrive. The contribution here is the specified multi-timescale memory and Hope design with its reported computational demonstrations, beyond the earlier Titans and MIRAS families.", "scope": "Some information should change quickly, while other knowledge should persist. Nested Learning organizes a model and its learning procedures as interacting optimization processes with different contexts and update rates. Its Continuum Memory System uses several memory components on different schedules. Hope combines that system with a self-referential extension of Titans, allowing parts of its memory processing to adapt while data arrive. The contribution here is the specified multi-timescale memory and Hope design with its reported computational demonstrations, beyond the earlier Titans and MIRAS families. The authors explicitly do not claim that catastrophic forgetting is solved in general. Continual-learning experiments include substantial additional training; long-context results can require benchmark-specific fine-tuning. Transformers remain stronger on some short-recall tasks, and the multiscale optimizer has additional overhead. More memory levels do not establish unbounded knowledge retention. Biological analogies and clinical benchmarks are not separate discoveries in this record. The exact hosted source revision is retained; its 2025 conference family does not make every later revision identical.", "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The authors explicitly do not claim that catastrophic forgetting is solved in general. Continual-learning experiments include substantial additional training; long-context results can require benchmark-specific fine-tuning. Transformers remain stronger on some short-recall tasks, and the multiscale optimizer has additional overhead. More memory levels do not establish unbounded knowledge retention. Biological analogies and clinical benchmarks are not separate discoveries in this record. The exact hosted source revision is retained; its 2025 conference family does not make every later revision identical.", "references": [ { "title": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf", "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained." } ], "roles": [ { "who": "The named authors in the research record", "credit": "SCIENCE", "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred." } ], "target": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "scientific_correctness_certified": false, "independent_expert_calibration": false, "method": "PROPRIETARY", "points": "180000" }, "canonical_agent_submission": false, "scientific_verification_awarded": false, "reader_guide": { "id": "behrouz-et-al-nested-learning-hope-cms-2025", "target": { "record_id": "urn:uuid:d6d66616-76f0-4c01-a848-fbf2dc6809bc", "version": 1, "content_hash": "sha256:9f4bc1c7e3d3f85abd717b1a21362f4985b0e1ebefae695303167b8fe0895948" }, "sections": [ { "title": "A project needs more than one clock", "text": "A team may revise its next action every minute, its weekly plan each Friday and its long-term goals only occasionally. Nested Learning uses different update schedules inside a computational model. The analogy concerns schedules, not human thought. Its Continuum Memory System places several memory components on different clocks so that rapid adaptation and more persistent knowledge do not have to use exactly the same update rule." }, { "title": "What Hope adds to Titans", "text": "Titans already changes a neural memory while processing input. Hope extends that design with components that adapt parts of the memory-processing procedure and with several memory timescales. A fast component can react to recent input while a slower component changes less often. The outer training process still matters: these components need suitable initial states and learning behavior. Attaching a label called slow memory does not make a randomly assembled model useful." }, { "title": "Count the updates, then count the cost", "text": "In a classroom scheduling example, a stream contains sixteen steps. One component updates every step and another updates every four steps. They make sixteen and four updates, respectively, for twenty in total. Two components that both update every step would make thirty-two. This is only an update count: the real cost also depends on each component’s size, its forward computations and the update algorithm. Twenty large updates could cost more than thirty-two small ones." }, { "title": "Slower change does not mean no forgetting", "text": "The paper explicitly says catastrophic forgetting is not solved in general. Its demonstrations include additional training and task-specific conditions. Some conventional Transformer results remain better on short recall, and the multiscale optimizer adds overhead. Ask your agent to draw a timeline for the sixteen-step example, then identify which experimental comparison in the paper actually isolates a memory schedule. A useful continuation states the tested conditions rather than promising a model that learns everything forever." }, { "title": "Leave the schedule visible", "text": "Your small timeline should name both update intervals and show every update position. Record what the count excludes: component size, forward work and training. Another reader can then change an interval and check the result. Do not describe fewer counted updates as a measured speedup or a solution to forgetting." } ], "references": [ { "title": "Nested Learning: The Illusion of Deep Learning Architecture", "url": "https://abehrouz.github.io/files/NL.pdf", "locator": "Pinned hosted revision: update-frequency definitions, sections7–8 CMS and Hope, section9 experiments and final explicit catastrophic-forgetting limitation; not an independent reconstruction of optimizer proofs." } ] }, "reading_status": "EXPLANATORY_GUIDE" }