{"id":"4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db","title":"Paper authors propose tracing safety degradation to training data","argument":"The authors of “Follow the Gradient: Tracing Safety Degradation Back to Training Data” present SimpleLoRIF, a method for finding fine-tuning examples whose gradients align with unsafe model behavior. They argue that some examples can degrade safety despite appearing benign, making them difficult for standard content filters to detect.\n\nThe authors use SimpleLoRIF to create SafeTulu, a version of Tulu 3 that removes the 20% of examples most aligned with unsafe responses. In their tests, SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability. These are reported experimental results, not evidence of safety in every deployment.\n\nThe authors also identify tool-execution failures and narrative arithmetic as possible, under-studied sources of safety degradation. They release SafeTulu and a Hugging Face Space for researchers to make customized versions. An open question is whether this filtering approach will preserve its reported safety and capability results on other datasets and in deployed systems.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the paper abstract; it does not provide full experimental methods or results. The reported results do not establish performance on other datasets or in deployed systems. The authors describe tool-execution failures and narrative arithmetic as possibly under-studied sources, not established causes in all settings.","speaker":{"name":"Reddy, Saket, Abbas, Amro, Suri, Anshuman, Morcos, Ari, Schwab, David, Maini, Pratyush","roleAtSource":"Article author"},"source":{"title":"Follow the Gradient: Tracing Safety Degradation Back to Training Data","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.37481","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors introduce SimpleLoRIF to identify fine-tuning examples whose gradients align with unsafe behavior.","excerpt":"a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior.","url":"https://arxiv.org/abs/2610.37481","relationship":"Source-stated"},{"text":"The authors report that SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability.","excerpt":"Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability.","url":"https://arxiv.org/abs/2610.37481","relationship":"Source-stated"},{"text":"The authors identify tool-execution failures and narrative arithmetic as possible under-studied sources of safety degradation.","excerpt":"identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic.","url":"https://arxiv.org/abs/2610.37481","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.37481","checkedAt":"2026-11-05T10:56:08.112Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Will SimpleLoRIF-based filtering preserve its reported safety and capability results on other datasets and in deployed systems?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Paper authors propose tracing safety degradation to training data","summary":"The authors of “Follow the Gradient: Tracing Safety Degradation Back to Training Data” present SimpleLoRIF, a method for finding fine-tuning examples whose gradients align with unsafe model behavior. They argue that some examples can degrade safety despite appearing benign, making them difficult for standard content filters to detect.\n\nThe authors use SimpleLoRIF to create SafeTulu, a version of Tulu 3 that removes the 20% of examples most aligned with unsafe responses. In their tests, SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability. These are reported experimental results, not evidence of safety in every deployment.\n\nThe authors also identify tool-execution failures and narrative arithmetic as possible, under-studied sources of safety degradation. They release SafeTulu and a Hugging Face Space for researchers to make customized versions. An open question is whether this filtering approach will preserve its reported safety and capability results on other datasets and in deployed systems.","question":"Will SimpleLoRIF-based filtering preserve its reported safety and capability results on other datasets and in deployed systems?","claims":[{"text":"The authors introduce SimpleLoRIF to identify fine-tuning examples whose gradients align with unsafe behavior.","excerpt":"a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior.","support":"source-stated"},{"text":"The authors report that SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability.","excerpt":"Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability.","support":"source-stated"},{"text":"The authors identify tool-execution failures and narrative arithmetic as possible under-studied sources of safety degradation.","excerpt":"identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic.","support":"source-stated"}],"limitations":["The evidence is the paper abstract; it does not provide full experimental methods or results.","The reported results do not establish performance on other datasets or in deployed systems.","The authors describe tool-execution failures and narrative arithmetic as possibly under-studied sources, not established causes in all settings."],"mediaId":"aff31c4fee36adb6fa0c176bafc20cd13185b095d99d1f5b52293c037f3521c3","correctionReason":null,"id":"4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db","publicationId":"666c58fd-ee8e-4e20-856b-d45955a50957","version":1,"sourceUrl":"https://arxiv.org/abs/2610.37481","sourceTitle":"Follow the Gradient: Tracing Safety Degradation Back to Training Data","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T06:28:41.180Z","createdAt":"2026-11-05T08:48:21.085Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"aff31c4fee36adb6fa0c176bafc20cd13185b095d99d1f5b52293c037f3521c3","kind":"article","url":"https://arxiv.org/abs/2610.37481","sourceUrl":"https://arxiv.org/abs/2610.37481","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T00:16:18.999Z","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/2610.37481","checkedAt":"2026-11-05T10:56:08.112Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"aff31c4fee36adb6fa0c176bafc20cd13185b095d99d1f5b52293c037f3521c3","confidence":0.9400000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"9e3f8bc578f8086c14385811b4809e89c7ac955ad9ef21b47e3a10610e090ea9","decidedAt":"2026-10-30T00:16:20.244Z"},"provenance":{"textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observationId":"4208e03867ca4ce0c8104edcab3691a8632cb5a1d1c13f363a7fd70275beb159","retrievedAt":"2026-10-30T00:16:18.695Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"ccb240d2-b71e-4f34-8097-04b41460ebad"},"reviewer":{"model":"gpt-6-luna","runId":"55bcdb01-3fab-472e-8ccc-2d412682ea05"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-05T08:48:21.085Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Reddy, Saket","url":null,"kind":"unknown"},{"name":"Abbas, Amro","url":null,"kind":"unknown"},{"name":"Suri, Anshuman","url":null,"kind":"unknown"},{"name":"Morcos, Ari","url":null,"kind":"unknown"},{"name":"Schwab, David","url":null,"kind":"unknown"},{"name":"Maini, Pratyush","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Reddy, Saket","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Reddy, Saket","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Abbas, Amro","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Abbas, Amro","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Suri, Anshuman","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Suri, Anshuman","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Morcos, Ari","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Morcos, Ari","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Schwab, David","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Schwab, David","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Maini, Pratyush","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Maini, Pratyush","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.37481","snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z","basis":"publisher-page-metadata","sourceLinks":47,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db/source"}},"canonicalUrl":"https://escapemeridian.com/readings/paper-authors-propose-tracing-safety-degradation-to-training-4ddd17c1","formats":{"html":"/readings/paper-authors-propose-tracing-safety-degradation-to-training-4ddd17c1","markdown":"/readings/paper-authors-propose-tracing-safety-degradation-to-training-4ddd17c1.md","json":"/readings/paper-authors-propose-tracing-safety-degradation-to-training-4ddd17c1.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-fine-tuning","name":"Fine-tuning","explanation":"Further training of an existing model to adapt it to a particular task or use.","referenceUrl":"https://developers.google.com/machine-learning/glossary#fine-tuning"}],"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":"olmo","name":"Olmo","kind":"model","context":"Olmo is Ai2’s family of language models accompanied by research artifacts intended to make development more inspectable.","referenceUrls":["https://allenai.org/olmo"],"path":"/entities/olmo","json":"/entities/olmo.json","markdown":"/entities/olmo.md"},{"id":"hugging-face","name":"Hugging Face","kind":"organization","context":"Hugging Face hosts models, datasets and software that let people share, adapt and inspect AI systems.","referenceUrls":["https://huggingface.co/blog"],"path":"/entities/hugging-face","json":"/entities/hugging-face.json","markdown":"/entities/hugging-face.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db","revision":"666c58fd-ee8e-4e20-856b-d45955a50957","revisionKind":"publication-id","title":"Paper authors propose tracing safety degradation to training data","url":"https://escapemeridian.com/readings/paper-authors-propose-tracing-safety-degradation-to-training-4ddd17c1"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.37481","revision":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","revisionKind":"sha256-source-text","title":"Follow the Gradient: Tracing Safety Degradation Back to Training Data","url":"https://arxiv.org/abs/2610.37481"},"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":"4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db","title":"Paper authors propose tracing safety degradation to training data","argument":"The authors of “Follow the Gradient: Tracing Safety Degradation Back to Training Data” present SimpleLoRIF, a method for finding fine-tuning examples whose gradients align with unsafe model behavior. They argue that some examples can degrade safety despite appearing benign, making them difficult for standard content filters to detect.\n\nThe authors use SimpleLoRIF to create SafeTulu, a version of Tulu 3 that removes the 20% of examples most aligned with unsafe responses. In their tests, SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability. These are reported experimental results, not evidence of safety in every deployment.\n\nThe authors also identify tool-execution failures and narrative arithmetic as possible, under-studied sources of safety degradation. They release SafeTulu and a Hugging Face Space for researchers to make customized versions. An open question is whether this filtering approach will preserve its reported safety and capability results on other datasets and in deployed systems.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the paper abstract; it does not provide full experimental methods or results. The reported results do not establish performance on other datasets or in deployed systems. The authors describe tool-execution failures and narrative arithmetic as possibly under-studied sources, not established causes in all settings.","speaker":{"name":"Reddy, Saket, Abbas, Amro, Suri, Anshuman, Morcos, Ari, Schwab, David, Maini, Pratyush","roleAtSource":"Article author"},"source":{"title":"Follow the Gradient: Tracing Safety Degradation Back to Training Data","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.37481","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors introduce SimpleLoRIF to identify fine-tuning examples whose gradients align with unsafe behavior.","excerpt":"a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior.","url":"https://arxiv.org/abs/2610.37481","relationship":"Source-stated"},{"text":"The authors report that SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability.","excerpt":"Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability.","url":"https://arxiv.org/abs/2610.37481","relationship":"Source-stated"},{"text":"The authors identify tool-execution failures and narrative arithmetic as possible under-studied sources of safety degradation.","excerpt":"identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic.","url":"https://arxiv.org/abs/2610.37481","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.37481","checkedAt":"2026-11-05T10:56:08.112Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Will SimpleLoRIF-based filtering preserve its reported safety and capability results on other datasets and in deployed systems?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Paper authors propose tracing safety degradation to training data","summary":"The authors of “Follow the Gradient: Tracing Safety Degradation Back to Training Data” present SimpleLoRIF, a method for finding fine-tuning examples whose gradients align with unsafe model behavior. They argue that some examples can degrade safety despite appearing benign, making them difficult for standard content filters to detect.\n\nThe authors use SimpleLoRIF to create SafeTulu, a version of Tulu 3 that removes the 20% of examples most aligned with unsafe responses. In their tests, SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability. These are reported experimental results, not evidence of safety in every deployment.\n\nThe authors also identify tool-execution failures and narrative arithmetic as possible, under-studied sources of safety degradation. They release SafeTulu and a Hugging Face Space for researchers to make customized versions. An open question is whether this filtering approach will preserve its reported safety and capability results on other datasets and in deployed systems.","question":"Will SimpleLoRIF-based filtering preserve its reported safety and capability results on other datasets and in deployed systems?","claims":[{"text":"The authors introduce SimpleLoRIF to identify fine-tuning examples whose gradients align with unsafe behavior.","excerpt":"a scalable gradient embedding method that identifies fine-tuning examples whose gradients align with unsafe behavior.","support":"source-stated"},{"text":"The authors report that SafeTulu reduced GCG attack success rates by up to 22% across Qwen, Llama, and Olmo, while largely preserving or improving capability.","excerpt":"Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo, outperforming both LLM-as-a-judge and prior gradient-based filtering, while largely preserving or even improving capability.","support":"source-stated"},{"text":"The authors identify tool-execution failures and narrative arithmetic as possible under-studied sources of safety degradation.","excerpt":"identify possibly under-studied sources of safety degradation, such as tool-execution failures and narrative arithmetic.","support":"source-stated"}],"limitations":["The evidence is the paper abstract; it does not provide full experimental methods or results.","The reported results do not establish performance on other datasets or in deployed systems.","The authors describe tool-execution failures and narrative arithmetic as possibly under-studied sources, not established causes in all settings."],"mediaId":"aff31c4fee36adb6fa0c176bafc20cd13185b095d99d1f5b52293c037f3521c3","correctionReason":null,"id":"4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db","publicationId":"666c58fd-ee8e-4e20-856b-d45955a50957","version":1,"sourceUrl":"https://arxiv.org/abs/2610.37481","sourceTitle":"Follow the Gradient: Tracing Safety Degradation Back to Training Data","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T06:28:41.180Z","createdAt":"2026-11-05T08:48:21.085Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"aff31c4fee36adb6fa0c176bafc20cd13185b095d99d1f5b52293c037f3521c3","kind":"article","url":"https://arxiv.org/abs/2610.37481","sourceUrl":"https://arxiv.org/abs/2610.37481","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T00:16:18.999Z","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/2610.37481","checkedAt":"2026-11-05T10:56:08.112Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"aff31c4fee36adb6fa0c176bafc20cd13185b095d99d1f5b52293c037f3521c3","confidence":0.9400000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"9e3f8bc578f8086c14385811b4809e89c7ac955ad9ef21b47e3a10610e090ea9","decidedAt":"2026-10-30T00:16:20.244Z"},"provenance":{"textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observationId":"4208e03867ca4ce0c8104edcab3691a8632cb5a1d1c13f363a7fd70275beb159","retrievedAt":"2026-10-30T00:16:18.695Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"ccb240d2-b71e-4f34-8097-04b41460ebad"},"reviewer":{"model":"gpt-6-luna","runId":"55bcdb01-3fab-472e-8ccc-2d412682ea05"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-05T08:48:21.085Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Reddy, Saket","url":null,"kind":"unknown"},{"name":"Abbas, Amro","url":null,"kind":"unknown"},{"name":"Suri, Anshuman","url":null,"kind":"unknown"},{"name":"Morcos, Ari","url":null,"kind":"unknown"},{"name":"Schwab, David","url":null,"kind":"unknown"},{"name":"Maini, Pratyush","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Reddy, Saket","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Reddy, Saket","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Abbas, Amro","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Abbas, Amro","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Suri, Anshuman","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Suri, Anshuman","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Morcos, Ari","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Morcos, Ari","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Schwab, David","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Schwab, David","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}},{"actor":{"name":"Maini, Pratyush","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Maini, Pratyush","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.37481","locator":"head/meta[name=author|citation_author]"},"assertionId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-10-30T00:16:18.695Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","sourceUrl":"https://arxiv.org/abs/2610.37481","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.37481","snapshotId":"0643f8bc458ba8150a3f79fbbfd2e0696ea44170fd3058ced2eea5ae10cdccf3","contentSha256":"bc33c0f141de2d21c4f40f4dc0b964125d04a553d3b7fe75fcc84a940acaeb08","textHash":"1e95eb9e4426f8352f57965bb0fb22fb0c0b156756af832c5db356136a73a0df","observedAt":"2026-10-30T00:16:18.695Z","basis":"publisher-page-metadata","sourceLinks":47,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/4ddd17c12970257e07e609be266d2f401acfd5ee628bb8b56c4d7bb4c8c089db/source"}}}],"audits":[{"publicationId":"666c58fd-ee8e-4e20-856b-d45955a50957","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft attributes SimpleLoRIF, SafeTulu, the reported results and the possible data sources to the paper's authors. It does not present these claims as established fact or as Escape Meridian's view.","excerpt":"Authors:Saket Reddy, Amro Abbas, Anshuman Suri, Ari Morcos, David Schwab, Pratyush Maini"},{"dimension":"summary","status":"uncertain","explanation":"The reading is faithful but hard to follow without specialist background. Tulu 3 is not identified as a post-training instruction dataset, and the central terms gradients and GCG attack are unexplained; they need tooltips. The draft also omits the paper's main practical claim: scaling to about a million examples turns gradient attribution into a filter for full datasets. It drops the reported advantage over LLM-as-a-judge and prior gradient filtering. Smallest repair: add one sentence on scale and the comparison, and state that Tulu 3 is a training dataset.","excerpt":"boosting gradient attribution from a small-scale analysis tool into a practical filter for full post-training datasets"},{"dimension":"date","status":"pass","explanation":"The originalPublished date of 2026-10-27 matches the arXiv v1 submission date. The later feed date was correctly not used.","excerpt":"[Submitted on 27 Oct 2026]"},{"dimension":"support","status":"pass","explanation":"The claims match the abstract: the method, the 20% removal, the up-to-22% GCG reduction across three model families, the capability preservation, the possible under-studied sources and the release. The caveat against generalizing to deployment and the open question are framed appropriately. No benefits or costs are invented.","excerpt":"Training on SafeTulu reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which gives readers the easiest access to the original source.","excerpt":null}],"createdAt":"2026-11-05T08:48:21.085Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}