{"id":"5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1","title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","argument":"Wonseok Shin and Songkuk Kim argue that a score suited to far-OOD detection can miss near-OOD inputs: these inputs activate class-characteristic feature directions much like in-distribution samples. They propose PLUS, a signal designed to complement existing scores rather than replace them. For each class, PLUS selects the most class-active neurons from a sparse autoencoder, forms a low-dimensional subspace from their decoder directions, and scores an input using its per-class whitened squared Mahalanobis distance within the subspace of its predicted class. The authors add this signal to a calibrated base score at one fixed global weight, with calibration statistics estimated from ID validation data alone. They report improved near-OOD AUROC across 126 detector–backbone combinations using 18 existing detectors, while far-OOD performance was broadly maintained or improved, with marginal declines where the base score was near-optimal. The proposed approach matters because it aims to add sensitivity to near-OOD inputs without discarding detectors already performing well on far-OOD. The abstract reports these evaluation results but does not identify the datasets, so their scope cannot be assessed from this evidence alone.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the arXiv abstract; it does not identify the datasets used in the reported evaluations, limiting assessment of their scope. The reported evaluation results are claims made by the paper's authors, not an independent assessment.","speaker":{"name":"Shin, Wonseok, Kim, Songkuk","roleAtSource":"Article author"},"source":{"title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.06688","publishedAt":"2026-11-04","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors argue that far-OOD and near-OOD differ in how much they activate class-characteristic feature directions, making a score optimized for one setting misaligned with the other.","excerpt":"far-OOD inputs largely fail to activate those of their predicted class, whereas near-OOD inputs activate them much as in-distribution samples do.","url":"https://arxiv.org/abs/2611.06688","relationship":"Source-stated"},{"text":"The authors propose PLUS as a complement to existing scores, using class-specific subspaces derived from a sparse autoencoder and an input's per-class whitened squared Mahalanobis distance.","excerpt":"We propose PLUS (Per-class Low-dimensional sUbspace Signal). For each class, PLUS selects the $m$ most class-active neurons from a sparse autoencoder","url":"https://arxiv.org/abs/2611.06688","relationship":"Source-stated"},{"text":"The authors report improved near-OOD AUROC in 126 detector–backbone combinations and broadly maintained or improved far-OOD performance, with some marginal declines.","excerpt":"consistently improves near-OOD AUROC in all 126 detector--backbone combinations. Far-OOD performance is broadly maintained or improved","url":"https://arxiv.org/abs/2611.06688","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.06688","checkedAt":"2026-11-09T06:42:06.462Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does PLUS perform on near-OOD and far-OOD datasets beyond those used in the reported evaluations?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","summary":"Wonseok Shin and Songkuk Kim argue that a score suited to far-OOD detection can miss near-OOD inputs: these inputs activate class-characteristic feature directions much like in-distribution samples. They propose PLUS, a signal designed to complement existing scores rather than replace them. For each class, PLUS selects the most class-active neurons from a sparse autoencoder, forms a low-dimensional subspace from their decoder directions, and scores an input using its per-class whitened squared Mahalanobis distance within the subspace of its predicted class. The authors add this signal to a calibrated base score at one fixed global weight, with calibration statistics estimated from ID validation data alone. They report improved near-OOD AUROC across 126 detector–backbone combinations using 18 existing detectors, while far-OOD performance was broadly maintained or improved, with marginal declines where the base score was near-optimal. The proposed approach matters because it aims to add sensitivity to near-OOD inputs without discarding detectors already performing well on far-OOD. The abstract reports these evaluation results but does not identify the datasets, so their scope cannot be assessed from this evidence alone.","question":"How does PLUS perform on near-OOD and far-OOD datasets beyond those used in the reported evaluations?","claims":[{"text":"The authors argue that far-OOD and near-OOD differ in how much they activate class-characteristic feature directions, making a score optimized for one setting misaligned with the other.","excerpt":"far-OOD inputs largely fail to activate those of their predicted class, whereas near-OOD inputs activate them much as in-distribution samples do.","support":"source-stated"},{"text":"The authors propose PLUS as a complement to existing scores, using class-specific subspaces derived from a sparse autoencoder and an input's per-class whitened squared Mahalanobis distance.","excerpt":"We propose PLUS (Per-class Low-dimensional sUbspace Signal). For each class, PLUS selects the $m$ most class-active neurons from a sparse autoencoder","support":"source-stated"},{"text":"The authors report improved near-OOD AUROC in 126 detector–backbone combinations and broadly maintained or improved far-OOD performance, with some marginal declines.","excerpt":"consistently improves near-OOD AUROC in all 126 detector--backbone combinations. Far-OOD performance is broadly maintained or improved","support":"source-stated"}],"limitations":["The supplied evidence is the arXiv abstract; it does not identify the datasets used in the reported evaluations, limiting assessment of their scope.","The reported evaluation results are claims made by the paper's authors, not an independent assessment."],"mediaId":"b9324470af05aec39997243d8659bec00d17005bdee26c5fbd11f63fd1f40a7d","correctionReason":null,"id":"5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1","publicationId":"884cc872-9cfc-49fe-8662-d5bf022061a4","version":1,"sourceUrl":"https://arxiv.org/abs/2611.06688","sourceTitle":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-04","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-05T05:09:25.183Z","createdAt":"2026-11-09T06:10:31.462Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b9324470af05aec39997243d8659bec00d17005bdee26c5fbd11f63fd1f40a7d","kind":"article","url":"https://arxiv.org/abs/2611.06688","sourceUrl":"https://arxiv.org/abs/2611.06688","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T06:09:45.624Z","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/2611.06688","checkedAt":"2026-11-09T06:42:06.462Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"b9324470af05aec39997243d8659bec00d17005bdee26c5fbd11f63fd1f40a7d","confidence":0.8300000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"7aa9f6f73f837032bbe9034b50b895f1ca6cfd2229c7d5367cdfe35496f0b004","decidedAt":"2026-11-09T06:09:46.574Z"},"provenance":{"textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observationId":"dac8dc6178258746f8b03baa29d74429c83dd51ed523a0036db12005c15c6f2c","retrievedAt":"2026-11-09T06:09:45.474Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"8bf7b311-9b8d-4929-8468-be10055f841a"},"reviewer":{"model":"gpt-6-luna","runId":"73e43a78-6ff3-443c-8910-465c2788a0e7"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T06:10:31.462Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Shin, Wonseok","url":null,"kind":"unknown"},{"name":"Kim, Songkuk","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Shin, Wonseok","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shin, Wonseok","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06688","locator":"head/meta[name=author|citation_author]"},"assertionId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:42:06.462Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf","sourceUrl":"https://arxiv.org/abs/2611.06688","contentSha256":"fba1184cb591d90c06ec29c4fc68bc8726ec371929452c56e8f6db76ae7f8acc","textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observedAt":"2026-11-09T06:42:06.462Z"}},{"actor":{"name":"Kim, Songkuk","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kim, Songkuk","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06688","locator":"head/meta[name=author|citation_author]"},"assertionId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:42:06.462Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf","sourceUrl":"https://arxiv.org/abs/2611.06688","contentSha256":"fba1184cb591d90c06ec29c4fc68bc8726ec371929452c56e8f6db76ae7f8acc","textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observedAt":"2026-11-09T06:42:06.462Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.06688","snapshotId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf","contentSha256":"fba1184cb591d90c06ec29c4fc68bc8726ec371929452c56e8f6db76ae7f8acc","textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observedAt":"2026-11-09T06:42:06.462Z","basis":"publisher-page-metadata","sourceLinks":42,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1/source"}},"canonicalUrl":"https://escapemeridian.com/readings/don-t-replace-complement-a-detector-agnostic-subspace-signal-5e00286c","formats":{"html":"/readings/don-t-replace-complement-a-detector-agnostic-subspace-signal-5e00286c","markdown":"/readings/don-t-replace-complement-a-detector-agnostic-subspace-signal-5e00286c.md","json":"/readings/don-t-replace-complement-a-detector-agnostic-subspace-signal-5e00286c.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":[],"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":"5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1","revision":"884cc872-9cfc-49fe-8662-d5bf022061a4","revisionKind":"publication-id","title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","url":"https://escapemeridian.com/readings/don-t-replace-complement-a-detector-agnostic-subspace-signal-5e00286c"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.06688","revision":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","revisionKind":"sha256-source-text","title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","url":"https://arxiv.org/abs/2611.06688"},"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":"5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1","title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","argument":"Wonseok Shin and Songkuk Kim argue that a score suited to far-OOD detection can miss near-OOD inputs: these inputs activate class-characteristic feature directions much like in-distribution samples. They propose PLUS, a signal designed to complement existing scores rather than replace them. For each class, PLUS selects the most class-active neurons from a sparse autoencoder, forms a low-dimensional subspace from their decoder directions, and scores an input using its per-class whitened squared Mahalanobis distance within the subspace of its predicted class. The authors add this signal to a calibrated base score at one fixed global weight, with calibration statistics estimated from ID validation data alone. They report improved near-OOD AUROC across 126 detector–backbone combinations using 18 existing detectors, while far-OOD performance was broadly maintained or improved, with marginal declines where the base score was near-optimal. The proposed approach matters because it aims to add sensitivity to near-OOD inputs without discarding detectors already performing well on far-OOD. The abstract reports these evaluation results but does not identify the datasets, so their scope cannot be assessed from this evidence alone.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the arXiv abstract; it does not identify the datasets used in the reported evaluations, limiting assessment of their scope. The reported evaluation results are claims made by the paper's authors, not an independent assessment.","speaker":{"name":"Shin, Wonseok, Kim, Songkuk","roleAtSource":"Article author"},"source":{"title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.06688","publishedAt":"2026-11-04","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors argue that far-OOD and near-OOD differ in how much they activate class-characteristic feature directions, making a score optimized for one setting misaligned with the other.","excerpt":"far-OOD inputs largely fail to activate those of their predicted class, whereas near-OOD inputs activate them much as in-distribution samples do.","url":"https://arxiv.org/abs/2611.06688","relationship":"Source-stated"},{"text":"The authors propose PLUS as a complement to existing scores, using class-specific subspaces derived from a sparse autoencoder and an input's per-class whitened squared Mahalanobis distance.","excerpt":"We propose PLUS (Per-class Low-dimensional sUbspace Signal). For each class, PLUS selects the $m$ most class-active neurons from a sparse autoencoder","url":"https://arxiv.org/abs/2611.06688","relationship":"Source-stated"},{"text":"The authors report improved near-OOD AUROC in 126 detector–backbone combinations and broadly maintained or improved far-OOD performance, with some marginal declines.","excerpt":"consistently improves near-OOD AUROC in all 126 detector--backbone combinations. Far-OOD performance is broadly maintained or improved","url":"https://arxiv.org/abs/2611.06688","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.06688","checkedAt":"2026-11-09T06:42:06.462Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How does PLUS perform on near-OOD and far-OOD datasets beyond those used in the reported evaluations?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","summary":"Wonseok Shin and Songkuk Kim argue that a score suited to far-OOD detection can miss near-OOD inputs: these inputs activate class-characteristic feature directions much like in-distribution samples. They propose PLUS, a signal designed to complement existing scores rather than replace them. For each class, PLUS selects the most class-active neurons from a sparse autoencoder, forms a low-dimensional subspace from their decoder directions, and scores an input using its per-class whitened squared Mahalanobis distance within the subspace of its predicted class. The authors add this signal to a calibrated base score at one fixed global weight, with calibration statistics estimated from ID validation data alone. They report improved near-OOD AUROC across 126 detector–backbone combinations using 18 existing detectors, while far-OOD performance was broadly maintained or improved, with marginal declines where the base score was near-optimal. The proposed approach matters because it aims to add sensitivity to near-OOD inputs without discarding detectors already performing well on far-OOD. The abstract reports these evaluation results but does not identify the datasets, so their scope cannot be assessed from this evidence alone.","question":"How does PLUS perform on near-OOD and far-OOD datasets beyond those used in the reported evaluations?","claims":[{"text":"The authors argue that far-OOD and near-OOD differ in how much they activate class-characteristic feature directions, making a score optimized for one setting misaligned with the other.","excerpt":"far-OOD inputs largely fail to activate those of their predicted class, whereas near-OOD inputs activate them much as in-distribution samples do.","support":"source-stated"},{"text":"The authors propose PLUS as a complement to existing scores, using class-specific subspaces derived from a sparse autoencoder and an input's per-class whitened squared Mahalanobis distance.","excerpt":"We propose PLUS (Per-class Low-dimensional sUbspace Signal). For each class, PLUS selects the $m$ most class-active neurons from a sparse autoencoder","support":"source-stated"},{"text":"The authors report improved near-OOD AUROC in 126 detector–backbone combinations and broadly maintained or improved far-OOD performance, with some marginal declines.","excerpt":"consistently improves near-OOD AUROC in all 126 detector--backbone combinations. Far-OOD performance is broadly maintained or improved","support":"source-stated"}],"limitations":["The supplied evidence is the arXiv abstract; it does not identify the datasets used in the reported evaluations, limiting assessment of their scope.","The reported evaluation results are claims made by the paper's authors, not an independent assessment."],"mediaId":"b9324470af05aec39997243d8659bec00d17005bdee26c5fbd11f63fd1f40a7d","correctionReason":null,"id":"5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1","publicationId":"884cc872-9cfc-49fe-8662-d5bf022061a4","version":1,"sourceUrl":"https://arxiv.org/abs/2611.06688","sourceTitle":"Don't Replace, Complement: A Detector-Agnostic Subspace Signal for Near-OOD Detection","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-04","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-05T05:09:25.183Z","createdAt":"2026-11-09T06:10:31.462Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b9324470af05aec39997243d8659bec00d17005bdee26c5fbd11f63fd1f40a7d","kind":"article","url":"https://arxiv.org/abs/2611.06688","sourceUrl":"https://arxiv.org/abs/2611.06688","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T06:09:45.624Z","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/2611.06688","checkedAt":"2026-11-09T06:42:06.462Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"b9324470af05aec39997243d8659bec00d17005bdee26c5fbd11f63fd1f40a7d","confidence":0.8300000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"7aa9f6f73f837032bbe9034b50b895f1ca6cfd2229c7d5367cdfe35496f0b004","decidedAt":"2026-11-09T06:09:46.574Z"},"provenance":{"textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observationId":"dac8dc6178258746f8b03baa29d74429c83dd51ed523a0036db12005c15c6f2c","retrievedAt":"2026-11-09T06:09:45.474Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"8bf7b311-9b8d-4929-8468-be10055f841a"},"reviewer":{"model":"gpt-6-luna","runId":"73e43a78-6ff3-443c-8910-465c2788a0e7"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T06:10:31.462Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Shin, Wonseok","url":null,"kind":"unknown"},{"name":"Kim, Songkuk","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Shin, Wonseok","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shin, Wonseok","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06688","locator":"head/meta[name=author|citation_author]"},"assertionId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:42:06.462Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf","sourceUrl":"https://arxiv.org/abs/2611.06688","contentSha256":"fba1184cb591d90c06ec29c4fc68bc8726ec371929452c56e8f6db76ae7f8acc","textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observedAt":"2026-11-09T06:42:06.462Z"}},{"actor":{"name":"Kim, Songkuk","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kim, Songkuk","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06688","locator":"head/meta[name=author|citation_author]"},"assertionId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T06:42:06.462Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf","sourceUrl":"https://arxiv.org/abs/2611.06688","contentSha256":"fba1184cb591d90c06ec29c4fc68bc8726ec371929452c56e8f6db76ae7f8acc","textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observedAt":"2026-11-09T06:42:06.462Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.06688","snapshotId":"166368472d9cabc80029c9ae85f5a726d6bbd87eefc21b44505206bcff2705bf","contentSha256":"fba1184cb591d90c06ec29c4fc68bc8726ec371929452c56e8f6db76ae7f8acc","textHash":"a1c33c768e792d392d6dc78ba6c380d8adcd22623152b470f8865945157a449b","observedAt":"2026-11-09T06:42:06.462Z","basis":"publisher-page-metadata","sourceLinks":42,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/5e00286c47dfd928dae4cecf68e537c4c888d5869586899019b177d1344fe4f1/source"}}}],"audits":[{"publicationId":"884cc872-9cfc-49fe-8662-d5bf022061a4","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly attributes the argument, method and reported results to authors Wonseok Shin and Songkuk Kim. It frames the results as reported by the authors, and the limitations note that they are not an independent assessment.","excerpt":"Authors:Wonseok Shin, Songkuk Kim"},{"dimension":"summary","status":"uncertain","explanation":"The summary faithfully tracks the abstract, but central terms are never explained for a general reader. It never expands OOD (out-of-distribution) or says what near- versus far-OOD inputs are, such as inputs unlike a model's training data. AUROC, sparse autoencoder, Mahalanobis distance, backbone and calibrated base score also go unexplained. Without these, a non-specialist cannot grasp what problem PLUS addresses or why it matters. Smallest repair: in the opening sentence, say plainly that OOD detection flags inputs unlike what the model was trained on, and that near-OOD inputs are deceptively similar ones; add tooltips for the remaining terms.","excerpt":"Post-hoc out-of-distribution (OOD) detection is mature on far-OOD yet remains weak on near-OOD."},{"dimension":"date","status":"pass","explanation":"The displayed original publication date of 2026-11-04 matches the arXiv v1 submission date. The feed date of 2026-11-05 was correctly not used.","excerpt":"[Submitted on 4 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"The claims match the abstract: the activation-direction argument, the PLUS construction, the single global fused weight using ID-only calibration, the 126 detector-backbone combinations across 18 detectors, and the far-OOD caveat. The stated significance of adding near-OOD sensitivity without discarding far-OOD detectors follows from the authors' complement framing. The limitations accurately note that the abstract does not identify the datasets.","excerpt":"Designed as a dedicated complement rather than a standalone detector, the signal $z$ is fused additively with any calibrated base score at a single, fixed weight."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-09T06:10:31.462Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}