{"id":"49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb","title":"How a paper studies over-refusal in safety-aligned large language models","argument":"The paper’s authors argue that safety-aligned large language models can reject harmless queries because harmful and over-refusal samples become less distinct in deeper layers, while harmful-sample gradients align more with refusal targets than with correct helpful targets. They propose GOAL, a safety-alignment framework that orthogonalizes over-refusal gradients against a low-rank refusal subspace and uses contrastive learning to separate the representations. The authors report that experiments across multiple models and benchmarks mitigate over-refusal while preserving safety performance and general capabilities. If that result holds in the tested settings, the approach could help reduce mistaken refusals without weakening the safeguards the authors measure. The abstract does not identify those models, benchmarks, or the measures used, so it does not show how broadly the result applies. Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper’s abstract; it does not name the models or benchmarks, give performance measures or results, or provide enough detail to assess how broadly the reported findings apply.","speaker":{"name":"Chen, Chao, Shao, Pengyang, Qin, Wei, Hao, Shijie, Shen, Fei, Hong, Richang","roleAtSource":"Article author"},"source":{"title":"Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.39802","publishedAt":"2026-10-29","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors describe over-refusal as rejecting harmless queries that appear superficially risky, and say why it persists under correct safety supervision is not sufficiently understood.","excerpt":"causing harmless but superficially risky queries to be mistakenly rejected","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors report that harmful and over-refusal sample representations are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers.","excerpt":"visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors report that harmful-sample gradients are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts.","excerpt":"more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors propose GOAL, which orthogonalizes over-refusal gradients against a low-rank refusal subspace extracted from harmful gradients and adds a contrastive learning objective.","excerpt":"we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors report that experiments across multiple models and benchmarks show GOAL mitigating over-refusal while preserving safety performance and general capabilities.","excerpt":"mitigates over-refusal while preserving safety performance and general capabilities","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.39802","checkedAt":"2026-11-11T12:56:08.475Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"How a paper studies over-refusal in safety-aligned large language models","summary":"The paper’s authors argue that safety-aligned large language models can reject harmless queries because harmful and over-refusal samples become less distinct in deeper layers, while harmful-sample gradients align more with refusal targets than with correct helpful targets. They propose GOAL, a safety-alignment framework that orthogonalizes over-refusal gradients against a low-rank refusal subspace and uses contrastive learning to separate the representations. The authors report that experiments across multiple models and benchmarks mitigate over-refusal while preserving safety performance and general capabilities. If that result holds in the tested settings, the approach could help reduce mistaken refusals without weakening the safeguards the authors measure. The abstract does not identify those models, benchmarks, or the measures used, so it does not show how broadly the result applies. Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?","question":"Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?","claims":[{"text":"The authors describe over-refusal as rejecting harmless queries that appear superficially risky, and say why it persists under correct safety supervision is not sufficiently understood.","excerpt":"causing harmless but superficially risky queries to be mistakenly rejected","support":"source-stated"},{"text":"The authors report that harmful and over-refusal sample representations are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers.","excerpt":"visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers","support":"source-stated"},{"text":"The authors report that harmful-sample gradients are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts.","excerpt":"more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts","support":"source-stated"},{"text":"The authors propose GOAL, which orthogonalizes over-refusal gradients against a low-rank refusal subspace extracted from harmful gradients and adds a contrastive learning objective.","excerpt":"we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients","support":"source-stated"},{"text":"The authors report that experiments across multiple models and benchmarks show GOAL mitigating over-refusal while preserving safety performance and general capabilities.","excerpt":"mitigates over-refusal while preserving safety performance and general capabilities","support":"source-stated"}],"limitations":["The supplied evidence is the paper’s abstract; it does not name the models or benchmarks, give performance measures or results, or provide enough detail to assess how broadly the reported findings apply."],"mediaId":"27d53a73930f47c0b3ec5500dd24f82b9f20f9e1dd8f296ccbdd3b678318ce5b","correctionReason":null,"id":"49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb","publicationId":"ed3f32f5-fb75-4dcb-878f-5b3e740b7592","version":1,"sourceUrl":"https://arxiv.org/abs/2610.39802","sourceTitle":"Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-29","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-30T06:42:58.003Z","createdAt":"2026-11-11T12:25:02.471Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"27d53a73930f47c0b3ec5500dd24f82b9f20f9e1dd8f296ccbdd3b678318ce5b","kind":"article","url":"https://arxiv.org/abs/2610.39802","sourceUrl":"https://arxiv.org/abs/2610.39802","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-31T06:01:54.672Z","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/2610.39802","checkedAt":"2026-11-11T12:56:08.475Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"27d53a73930f47c0b3ec5500dd24f82b9f20f9e1dd8f296ccbdd3b678318ce5b","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"78d94d04e0b05d1d67712369b560137ae3909b3ee0bc47b29d4de6932ebe968c","decidedAt":"2026-10-31T06:01:55.365Z"},"provenance":{"textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observationId":"61fb0a45f29eecfcc4291cfa23e2a9021800e2bc75fdf6795de98aae5de2bb3c","retrievedAt":"2026-10-31T06:01:54.401Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"b5c562af-ae57-4633-8656-8261ace9d9b4"},"reviewer":{"model":"gpt-6-luna","runId":"f3ca4042-3b4c-4f42-8aa2-689e7efb2f1b"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T12:25:02.471Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Chen, Chao","url":null,"kind":"unknown"},{"name":"Shao, Pengyang","url":null,"kind":"unknown"},{"name":"Qin, Wei","url":null,"kind":"unknown"},{"name":"Hao, Shijie","url":null,"kind":"unknown"},{"name":"Shen, Fei","url":null,"kind":"unknown"},{"name":"Hong, Richang","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Chen, Chao","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chen, Chao","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Shao, Pengyang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shao, Pengyang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Qin, Wei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Qin, Wei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Hao, Shijie","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hao, Shijie","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Shen, Fei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shen, Fei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Hong, Richang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Richang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.39802","snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb/source"}},"canonicalUrl":"https://escapemeridian.com/readings/how-a-paper-studies-over-refusal-in-safety-aligned-large-49448b19","formats":{"html":"/readings/how-a-paper-studies-over-refusal-in-safety-aligned-large-49448b19","markdown":"/readings/how-a-paper-studies-over-refusal-in-safety-aligned-large-49448b19.md","json":"/readings/how-a-paper-studies-over-refusal-in-safety-aligned-large-49448b19.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-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"}],"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":"49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb","revision":"ed3f32f5-fb75-4dcb-878f-5b3e740b7592","revisionKind":"publication-id","title":"How a paper studies over-refusal in safety-aligned large language models","url":"https://escapemeridian.com/readings/how-a-paper-studies-over-refusal-in-safety-aligned-large-49448b19"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.39802","revision":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","revisionKind":"sha256-source-text","title":"Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective","url":"https://arxiv.org/abs/2610.39802"},"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":"49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb","title":"How a paper studies over-refusal in safety-aligned large language models","argument":"The paper’s authors argue that safety-aligned large language models can reject harmless queries because harmful and over-refusal samples become less distinct in deeper layers, while harmful-sample gradients align more with refusal targets than with correct helpful targets. They propose GOAL, a safety-alignment framework that orthogonalizes over-refusal gradients against a low-rank refusal subspace and uses contrastive learning to separate the representations. The authors report that experiments across multiple models and benchmarks mitigate over-refusal while preserving safety performance and general capabilities. If that result holds in the tested settings, the approach could help reduce mistaken refusals without weakening the safeguards the authors measure. The abstract does not identify those models, benchmarks, or the measures used, so it does not show how broadly the result applies. Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper’s abstract; it does not name the models or benchmarks, give performance measures or results, or provide enough detail to assess how broadly the reported findings apply.","speaker":{"name":"Chen, Chao, Shao, Pengyang, Qin, Wei, Hao, Shijie, Shen, Fei, Hong, Richang","roleAtSource":"Article author"},"source":{"title":"Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.39802","publishedAt":"2026-10-29","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors describe over-refusal as rejecting harmless queries that appear superficially risky, and say why it persists under correct safety supervision is not sufficiently understood.","excerpt":"causing harmless but superficially risky queries to be mistakenly rejected","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors report that harmful and over-refusal sample representations are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers.","excerpt":"visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors report that harmful-sample gradients are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts.","excerpt":"more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors propose GOAL, which orthogonalizes over-refusal gradients against a low-rank refusal subspace extracted from harmful gradients and adds a contrastive learning objective.","excerpt":"we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"},{"text":"The authors report that experiments across multiple models and benchmarks show GOAL mitigating over-refusal while preserving safety performance and general capabilities.","excerpt":"mitigates over-refusal while preserving safety performance and general capabilities","url":"https://arxiv.org/abs/2610.39802","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.39802","checkedAt":"2026-11-11T12:56:08.475Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"How a paper studies over-refusal in safety-aligned large language models","summary":"The paper’s authors argue that safety-aligned large language models can reject harmless queries because harmful and over-refusal samples become less distinct in deeper layers, while harmful-sample gradients align more with refusal targets than with correct helpful targets. They propose GOAL, a safety-alignment framework that orthogonalizes over-refusal gradients against a low-rank refusal subspace and uses contrastive learning to separate the representations. The authors report that experiments across multiple models and benchmarks mitigate over-refusal while preserving safety performance and general capabilities. If that result holds in the tested settings, the approach could help reduce mistaken refusals without weakening the safeguards the authors measure. The abstract does not identify those models, benchmarks, or the measures used, so it does not show how broadly the result applies. Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?","question":"Which models, benchmarks, and measures support the reported gains, and how do the results vary across them?","claims":[{"text":"The authors describe over-refusal as rejecting harmless queries that appear superficially risky, and say why it persists under correct safety supervision is not sufficiently understood.","excerpt":"causing harmless but superficially risky queries to be mistakenly rejected","support":"source-stated"},{"text":"The authors report that harmful and over-refusal sample representations are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers.","excerpt":"visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers","support":"source-stated"},{"text":"The authors report that harmful-sample gradients are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts.","excerpt":"more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts","support":"source-stated"},{"text":"The authors propose GOAL, which orthogonalizes over-refusal gradients against a low-rank refusal subspace extracted from harmful gradients and adds a contrastive learning objective.","excerpt":"we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients","support":"source-stated"},{"text":"The authors report that experiments across multiple models and benchmarks show GOAL mitigating over-refusal while preserving safety performance and general capabilities.","excerpt":"mitigates over-refusal while preserving safety performance and general capabilities","support":"source-stated"}],"limitations":["The supplied evidence is the paper’s abstract; it does not name the models or benchmarks, give performance measures or results, or provide enough detail to assess how broadly the reported findings apply."],"mediaId":"27d53a73930f47c0b3ec5500dd24f82b9f20f9e1dd8f296ccbdd3b678318ce5b","correctionReason":null,"id":"49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb","publicationId":"ed3f32f5-fb75-4dcb-878f-5b3e740b7592","version":1,"sourceUrl":"https://arxiv.org/abs/2610.39802","sourceTitle":"Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-29","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-30T06:42:58.003Z","createdAt":"2026-11-11T12:25:02.471Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"27d53a73930f47c0b3ec5500dd24f82b9f20f9e1dd8f296ccbdd3b678318ce5b","kind":"article","url":"https://arxiv.org/abs/2610.39802","sourceUrl":"https://arxiv.org/abs/2610.39802","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-31T06:01:54.672Z","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/2610.39802","checkedAt":"2026-11-11T12:56:08.475Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"27d53a73930f47c0b3ec5500dd24f82b9f20f9e1dd8f296ccbdd3b678318ce5b","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"78d94d04e0b05d1d67712369b560137ae3909b3ee0bc47b29d4de6932ebe968c","decidedAt":"2026-10-31T06:01:55.365Z"},"provenance":{"textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observationId":"61fb0a45f29eecfcc4291cfa23e2a9021800e2bc75fdf6795de98aae5de2bb3c","retrievedAt":"2026-10-31T06:01:54.401Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"b5c562af-ae57-4633-8656-8261ace9d9b4"},"reviewer":{"model":"gpt-6-luna","runId":"f3ca4042-3b4c-4f42-8aa2-689e7efb2f1b"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T12:25:02.471Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Chen, Chao","url":null,"kind":"unknown"},{"name":"Shao, Pengyang","url":null,"kind":"unknown"},{"name":"Qin, Wei","url":null,"kind":"unknown"},{"name":"Hao, Shijie","url":null,"kind":"unknown"},{"name":"Shen, Fei","url":null,"kind":"unknown"},{"name":"Hong, Richang","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Chen, Chao","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chen, Chao","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Shao, Pengyang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shao, Pengyang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Qin, Wei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Qin, Wei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Hao, Shijie","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hao, Shijie","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Shen, Fei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shen, Fei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}},{"actor":{"name":"Hong, Richang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Richang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.39802","locator":"head/meta[name=author|citation_author]"},"assertionId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T12:56:08.475Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","sourceUrl":"https://arxiv.org/abs/2610.39802","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.39802","snapshotId":"cc4670a9db89e21d8e2fe480d9f6d5be31b7344c1542bc36c5ea1c1deb289e7f","contentSha256":"4d3ea4e6b7657d4834380faa88ed1df74c8025f38309be3efb979ab75b9cc199","textHash":"78eb406d318db028bfb62f9b624c5025c5765cfbb838e785b94d8bc539fe2015","observedAt":"2026-11-11T12:56:08.475Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/49448b19484f0e4b646533f324e3838f2c58e36befa3780bf8e5073fdb7d1cbb/source"}}}],"audits":[{"publicationId":"ed3f32f5-fb75-4dcb-878f-5b3e740b7592","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the analysis, the GOAL method and the experimental claims to the paper's authors. Its claims stay framed as author reports rather than established fact, consistent with the arXiv abstract page.","excerpt":"Authors:Chao Chen, Pengyang Shao, Wei Qin, Shijie Hao, Fei Shen, Richang Hong"},{"dimension":"summary","status":"uncertain","explanation":"Central technical terms are not explained in plain language, which blocks understanding of the core mechanism for a reader without specialist background. These include gradients, representations in deeper layers, orthogonalize, low-rank refusal subspace and contrastive learning. Smallest repair: add tooltips or brief plain glosses for these terms. Separately, the first sentence drops the scope condition on over-refusal prompts and states the cause with 'because', whereas the source hedges with 'can bias' and 'may make it more difficult'. Smallest repair: restore that hedging.","excerpt":"This imbalanced gradient coupling can bias over-refusal samples toward refusal, while the reduced apparent representational separation in deeper layers may make it more difficult"},{"dimension":"date","status":"pass","explanation":"The original publication date of 2026-10-29 matches the v1 submission date. The feed date of Oct 31 is a later announcement and is correctly not used.","excerpt":"[Submitted on 29 Oct 2026]"},{"dimension":"support","status":"uncertain","explanation":"The individual claims are each grounded in verbatim excerpts, and the stated limitations accurately note that the abstract gives no models, benchmarks or measures. However, the summary's causal 'because' framing is firmer than the source, which presents its mechanism as one that can bias outputs and may impair the boundary. The summary also says harmful-sample gradients align with 'refusal targets' and omits that this comparison applies to over-refusal prompts. The conditional benefit sentence is appropriately hedged.","excerpt":"more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the arXiv abstract page, which is the original source.","excerpt":null}],"createdAt":"2026-11-11T12:25:02.471Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}