{"id":"3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e","title":"arXiv cs.AI: Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","argument":"Hamid Dashtbani and Alexandra Dmitrienko propose Addition by Subtraction, a way to make safety alignment less dependent on a small set of model components. They describe a weakness in which removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior. Their method gradually removes identified safety-critical neurons during finetuning, then uses preference optimization to re-establish safety-aligned behavior through other parts of the model. The authors report that models trained this way resist pruning-based attacks and also improve against refusal-direction ablation and a broad range of black-box jailbreak attacks. They further report improved truthfulness while largely preserving general capabilities. In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks. Escape Meridian interpretation: the central idea is to rebuild safety behavior after removing components it might otherwise rely on, rather than leave that dependence untouched. The abstract presents this as an initial step and reports experimental results; it does not establish how well the approach transfers beyond the reported evaluations or how its tradeoffs vary across models and attacks.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied arXiv page provides the abstract, not detailed attack setups, evaluation measures, or results by model; the reported robustness and attack-success reductions cannot be assessed in greater detail from this evidence alone. The abstract describes the work as taking “the first steps” toward strengthening safety alignment, and does not establish performance beyond the evaluations it reports.","speaker":{"name":"Dashtbani, Hamid, Dmitrienko, Alexandra","roleAtSource":"Article author"},"source":{"title":"Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.06553","publishedAt":"2026-11-04","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors identify sparse safety alignment as a weakness: removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior.","excerpt":"pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"Addition by Subtraction removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model.","excerpt":"gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"The authors report improved robustness against pruning-based attacks, refusal-direction ablation, and a broad range of black-box jailbreak attacks.","excerpt":"improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"The authors report improved truthfulness while largely preserving general capabilities.","excerpt":"also exhibit improved truthfulness while largely preserving general capabilities","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks.","excerpt":"across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"The proposed method’s central idea is to rebuild safety behavior after removing components it might otherwise rely on.","excerpt":"uses preference optimization to re-establish safety-aligned behavior through other parts of the model","url":"https://arxiv.org/abs/2611.06553","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.06553","checkedAt":"2026-11-09T07:48:06.640Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What attack setups and evaluation details underlie the reported reductions in attack success rates?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","summary":"Hamid Dashtbani and Alexandra Dmitrienko propose Addition by Subtraction, a way to make safety alignment less dependent on a small set of model components. They describe a weakness in which removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior. Their method gradually removes identified safety-critical neurons during finetuning, then uses preference optimization to re-establish safety-aligned behavior through other parts of the model. The authors report that models trained this way resist pruning-based attacks and also improve against refusal-direction ablation and a broad range of black-box jailbreak attacks. They further report improved truthfulness while largely preserving general capabilities. In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks. Escape Meridian interpretation: the central idea is to rebuild safety behavior after removing components it might otherwise rely on, rather than leave that dependence untouched. The abstract presents this as an initial step and reports experimental results; it does not establish how well the approach transfers beyond the reported evaluations or how its tradeoffs vary across models and attacks.","question":"What attack setups and evaluation details underlie the reported reductions in attack success rates?","claims":[{"text":"The authors identify sparse safety alignment as a weakness: removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior.","excerpt":"pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior","support":"source-stated"},{"text":"Addition by Subtraction removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model.","excerpt":"gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model","support":"source-stated"},{"text":"The authors report improved robustness against pruning-based attacks, refusal-direction ablation, and a broad range of black-box jailbreak attacks.","excerpt":"improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks","support":"source-stated"},{"text":"The authors report improved truthfulness while largely preserving general capabilities.","excerpt":"also exhibit improved truthfulness while largely preserving general capabilities","support":"source-stated"},{"text":"In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks.","excerpt":"across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.","support":"source-stated"},{"text":"The proposed method’s central idea is to rebuild safety behavior after removing components it might otherwise rely on.","excerpt":"uses preference optimization to re-establish safety-aligned behavior through other parts of the model","support":"interpretation"}],"limitations":["The supplied arXiv page provides the abstract, not detailed attack setups, evaluation measures, or results by model; the reported robustness and attack-success reductions cannot be assessed in greater detail from this evidence alone.","The abstract describes the work as taking “the first steps” toward strengthening safety alignment, and does not establish performance beyond the evaluations it reports."],"mediaId":"4ec820d09c23a54f84d8987016cb76673862ad073b8dcb65d6b7018da1a2e60d","correctionReason":null,"id":"3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e","publicationId":"c43f9020-934a-4618-828e-a563db8a7f9c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.06553","sourceTitle":"Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-04","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-05T05:00:27.341Z","createdAt":"2026-11-09T07:17:41.956Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"4ec820d09c23a54f84d8987016cb76673862ad073b8dcb65d6b7018da1a2e60d","kind":"article","url":"https://arxiv.org/abs/2611.06553","sourceUrl":"https://arxiv.org/abs/2611.06553","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T07:09:15.215Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.06553","checkedAt":"2026-11-09T07:48:06.640Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"4ec820d09c23a54f84d8987016cb76673862ad073b8dcb65d6b7018da1a2e60d","confidence":0.92,"provider":"typesafe","model":"jev-1.13.0","inputHash":"eff3e913fa3995e2e7f2a9a95bebe49f54ad5f1d4e5b8dcb4b0870a341b4f5bd","decidedAt":"2026-11-09T07:09:15.833Z"},"provenance":{"textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observationId":"14ea36792cad8cf9b1c339fa1ba6a68b4f5069b68ffeea13cb05b860c37919dc","retrievedAt":"2026-11-09T07:09:15.126Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"a85ba6cf-82d5-4279-8929-99b67f391799"},"reviewer":{"model":"gpt-6-luna","runId":"7a48dc75-c747-40be-8b85-ce8ffcb352df"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T07:17:41.956Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Dashtbani, Hamid","url":null,"kind":"unknown"},{"name":"Dmitrienko, Alexandra","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Dashtbani, Hamid","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Dashtbani, Hamid","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06553","locator":"head/meta[name=author|citation_author]"},"assertionId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T07:48:06.640Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607","sourceUrl":"https://arxiv.org/abs/2611.06553","contentSha256":"335fddbbe7a70c953a0defc8544d4bbd79c80459eb24861efb0f6a4bada9613c","textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observedAt":"2026-11-09T07:48:06.640Z"}},{"actor":{"name":"Dmitrienko, Alexandra","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Dmitrienko, Alexandra","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06553","locator":"head/meta[name=author|citation_author]"},"assertionId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T07:48:06.640Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607","sourceUrl":"https://arxiv.org/abs/2611.06553","contentSha256":"335fddbbe7a70c953a0defc8544d4bbd79c80459eb24861efb0f6a4bada9613c","textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observedAt":"2026-11-09T07:48:06.640Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.06553","snapshotId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607","contentSha256":"335fddbbe7a70c953a0defc8544d4bbd79c80459eb24861efb0f6a4bada9613c","textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observedAt":"2026-11-09T07:48:06.640Z","basis":"publisher-page-metadata","sourceLinks":41,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-cs-ai-addition-by-subtraction-toward-mitigating-fragile-3d7af2f4","formats":{"html":"/readings/arxiv-cs-ai-addition-by-subtraction-toward-mitigating-fragile-3d7af2f4","markdown":"/readings/arxiv-cs-ai-addition-by-subtraction-toward-mitigating-fragile-3d7af2f4.md","json":"/readings/arxiv-cs-ai-addition-by-subtraction-toward-mitigating-fragile-3d7af2f4.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-llm","name":"Large language model","explanation":"A language model with many learned numerical settings, used to process and generate language.","referenceUrl":"https://developers.google.com/machine-learning/glossary#large-language-model"},{"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"},{"id":"term-jailbreak","name":"Jailbreak","explanation":"A prompt written to trick an AI system into ignoring its safety rules.","referenceUrl":"https://arxiv.org/abs/2307.02483"}],"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":"3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e","revision":"c43f9020-934a-4618-828e-a563db8a7f9c","revisionKind":"publication-id","title":"arXiv cs.AI: Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","url":"https://escapemeridian.com/readings/arxiv-cs-ai-addition-by-subtraction-toward-mitigating-fragile-3d7af2f4"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.06553","revision":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","revisionKind":"sha256-source-text","title":"Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","url":"https://arxiv.org/abs/2611.06553"},"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":"3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e","title":"arXiv cs.AI: Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","argument":"Hamid Dashtbani and Alexandra Dmitrienko propose Addition by Subtraction, a way to make safety alignment less dependent on a small set of model components. They describe a weakness in which removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior. Their method gradually removes identified safety-critical neurons during finetuning, then uses preference optimization to re-establish safety-aligned behavior through other parts of the model. The authors report that models trained this way resist pruning-based attacks and also improve against refusal-direction ablation and a broad range of black-box jailbreak attacks. They further report improved truthfulness while largely preserving general capabilities. In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks. Escape Meridian interpretation: the central idea is to rebuild safety behavior after removing components it might otherwise rely on, rather than leave that dependence untouched. The abstract presents this as an initial step and reports experimental results; it does not establish how well the approach transfers beyond the reported evaluations or how its tradeoffs vary across models and attacks.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied arXiv page provides the abstract, not detailed attack setups, evaluation measures, or results by model; the reported robustness and attack-success reductions cannot be assessed in greater detail from this evidence alone. The abstract describes the work as taking “the first steps” toward strengthening safety alignment, and does not establish performance beyond the evaluations it reports.","speaker":{"name":"Dashtbani, Hamid, Dmitrienko, Alexandra","roleAtSource":"Article author"},"source":{"title":"Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.06553","publishedAt":"2026-11-04","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors identify sparse safety alignment as a weakness: removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior.","excerpt":"pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"Addition by Subtraction removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model.","excerpt":"gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"The authors report improved robustness against pruning-based attacks, refusal-direction ablation, and a broad range of black-box jailbreak attacks.","excerpt":"improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"The authors report improved truthfulness while largely preserving general capabilities.","excerpt":"also exhibit improved truthfulness while largely preserving general capabilities","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks.","excerpt":"across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.","url":"https://arxiv.org/abs/2611.06553","relationship":"Source-stated"},{"text":"The proposed method’s central idea is to rebuild safety behavior after removing components it might otherwise rely on.","excerpt":"uses preference optimization to re-establish safety-aligned behavior through other parts of the model","url":"https://arxiv.org/abs/2611.06553","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.06553","checkedAt":"2026-11-09T07:48:06.640Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What attack setups and evaluation details underlie the reported reductions in attack success rates?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","summary":"Hamid Dashtbani and Alexandra Dmitrienko propose Addition by Subtraction, a way to make safety alignment less dependent on a small set of model components. They describe a weakness in which removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior. Their method gradually removes identified safety-critical neurons during finetuning, then uses preference optimization to re-establish safety-aligned behavior through other parts of the model. The authors report that models trained this way resist pruning-based attacks and also improve against refusal-direction ablation and a broad range of black-box jailbreak attacks. They further report improved truthfulness while largely preserving general capabilities. In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks. Escape Meridian interpretation: the central idea is to rebuild safety behavior after removing components it might otherwise rely on, rather than leave that dependence untouched. The abstract presents this as an initial step and reports experimental results; it does not establish how well the approach transfers beyond the reported evaluations or how its tradeoffs vary across models and attacks.","question":"What attack setups and evaluation details underlie the reported reductions in attack success rates?","claims":[{"text":"The authors identify sparse safety alignment as a weakness: removing safety-critical neurons or ablating a refusal direction can undermine aligned behavior.","excerpt":"pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior","support":"source-stated"},{"text":"Addition by Subtraction removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model.","excerpt":"gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model","support":"source-stated"},{"text":"The authors report improved robustness against pruning-based attacks, refusal-direction ablation, and a broad range of black-box jailbreak attacks.","excerpt":"improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks","support":"source-stated"},{"text":"The authors report improved truthfulness while largely preserving general capabilities.","excerpt":"also exhibit improved truthfulness while largely preserving general capabilities","support":"source-stated"},{"text":"In experiments across four models from three vendors, attack success rates fell by up to 90% for some attacks.","excerpt":"across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.","support":"source-stated"},{"text":"The proposed method’s central idea is to rebuild safety behavior after removing components it might otherwise rely on.","excerpt":"uses preference optimization to re-establish safety-aligned behavior through other parts of the model","support":"interpretation"}],"limitations":["The supplied arXiv page provides the abstract, not detailed attack setups, evaluation measures, or results by model; the reported robustness and attack-success reductions cannot be assessed in greater detail from this evidence alone.","The abstract describes the work as taking “the first steps” toward strengthening safety alignment, and does not establish performance beyond the evaluations it reports."],"mediaId":"4ec820d09c23a54f84d8987016cb76673862ad073b8dcb65d6b7018da1a2e60d","correctionReason":null,"id":"3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e","publicationId":"c43f9020-934a-4618-828e-a563db8a7f9c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.06553","sourceTitle":"Addition by Subtraction: Toward Mitigating Fragile Safety Alignment in Large Language Models","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-04","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-05T05:00:27.341Z","createdAt":"2026-11-09T07:17:41.956Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"4ec820d09c23a54f84d8987016cb76673862ad073b8dcb65d6b7018da1a2e60d","kind":"article","url":"https://arxiv.org/abs/2611.06553","sourceUrl":"https://arxiv.org/abs/2611.06553","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T07:09:15.215Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.06553","checkedAt":"2026-11-09T07:48:06.640Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"4ec820d09c23a54f84d8987016cb76673862ad073b8dcb65d6b7018da1a2e60d","confidence":0.92,"provider":"typesafe","model":"jev-1.13.0","inputHash":"eff3e913fa3995e2e7f2a9a95bebe49f54ad5f1d4e5b8dcb4b0870a341b4f5bd","decidedAt":"2026-11-09T07:09:15.833Z"},"provenance":{"textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observationId":"14ea36792cad8cf9b1c339fa1ba6a68b4f5069b68ffeea13cb05b860c37919dc","retrievedAt":"2026-11-09T07:09:15.126Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"a85ba6cf-82d5-4279-8929-99b67f391799"},"reviewer":{"model":"gpt-6-luna","runId":"7a48dc75-c747-40be-8b85-ce8ffcb352df"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T07:17:41.956Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Dashtbani, Hamid","url":null,"kind":"unknown"},{"name":"Dmitrienko, Alexandra","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Dashtbani, Hamid","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Dashtbani, Hamid","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06553","locator":"head/meta[name=author|citation_author]"},"assertionId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T07:48:06.640Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607","sourceUrl":"https://arxiv.org/abs/2611.06553","contentSha256":"335fddbbe7a70c953a0defc8544d4bbd79c80459eb24861efb0f6a4bada9613c","textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observedAt":"2026-11-09T07:48:06.640Z"}},{"actor":{"name":"Dmitrienko, Alexandra","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Dmitrienko, Alexandra","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.06553","locator":"head/meta[name=author|citation_author]"},"assertionId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T07:48:06.640Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607","sourceUrl":"https://arxiv.org/abs/2611.06553","contentSha256":"335fddbbe7a70c953a0defc8544d4bbd79c80459eb24861efb0f6a4bada9613c","textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observedAt":"2026-11-09T07:48:06.640Z"}}],"participationCoverage":{"total":2,"returned":2,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.06553","snapshotId":"9adf04e027b8a538eddd2c78b537829b416461cf0d678dde1865573f2e5c1607","contentSha256":"335fddbbe7a70c953a0defc8544d4bbd79c80459eb24861efb0f6a4bada9613c","textHash":"bdcab4dd0cb00071c469c5c565b2d05a02aa7c3216b1d11896f68786ce1cfcde","observedAt":"2026-11-09T07:48:06.640Z","basis":"publisher-page-metadata","sourceLinks":41,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/3d7af2f47fe4a0fc8382fefff45bef1dcea6467b5705c4b10a01e9727d75023e/source"}}}],"audits":[{"publicationId":"c43f9020-934a-4618-828e-a563db8a7f9c","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly credits Hamid Dashtbani and Alexandra Dmitrienko and keeps the results framed as what the authors report. The Escape Meridian interpretation is clearly labeled and separated from the source's claims.","excerpt":"Authors:Hamid Dashtbani, Alexandra Dmitrienko"},{"dimension":"summary","status":"uncertain","explanation":"The draft is faithful and keeps the sparsity point ('a small set of model components'). A general reader still meets central jargon with no plain explanation in the reading: safety alignment, safety-critical neurons, pruning, ablating a refusal direction, preference optimization, and black-box jailbreak attacks. These terms carry the whole argument, so they need term-library tooltips. Where a term is missing from the library, add it. The draft should keep these terms exactly rather than paraphrase them. The title only repeats the source title, which is acceptable. The question is a fair open issue.","excerpt":"pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior"},{"dimension":"date","status":"pass","explanation":"The originalPublished date, 2026-11-04, matches the arXiv v1 submission date. The feed date of 2026-11-05 is correctly not used.","excerpt":"[Submitted on 4 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract: the method, robustness to pruning, refusal-direction ablation and black-box jailbreaks, improved truthfulness, and the reduction of up to 90% for some attacks across four models from three vendors. The interpretation is labeled and conditional. The limitations correctly note the 'first steps' framing and that the evidence is abstract-only. The draft adds no unsupported benefits or costs.","excerpt":"reducing attack success rates by up to 90% for some attacks"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the arXiv abstract page. This is the original source.","excerpt":null}],"createdAt":"2026-11-09T07:17:41.956Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}