{"id":"dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613","title":"arXiv cs.AI: Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","argument":"Yubo Cui, Xianchao Guan and Zheng Zhang propose Textual Response-Aware Guidance (TRAG), a method for adversarial fine-tuning of vision-language models (VLMs). They argue that two existing ways of guiding fine-tuning have complementary limits: matching visual features treats all directions equally, while matching prediction distributions depends on the available candidate texts and prediction probabilities. Those limits, they say, can weaken pre-trained image--text relationships and degrade zero-shot performance.\n\nTRAG uses how a pre-trained text encoder responds to small changes in input embeddings to constrain visual representation drift. It builds constraints around semantic anchors, then combines them according to pre-trained image--anchor relevance to make the constraint image-adaptive. The authors report that experiments across architectures and tasks achieve state-of-the-art clean and robust performance, and that further analysis finds TRAG helps preserve pre-trained image--text relationships. The abstract does not identify the specific architectures, tasks, comparison methods or performance measures, so it does not show how broad or large those reported gains are.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract does not identify the architectures, tasks, comparison methods or performance measures behind the reported results; the breadth and size of the reported gains cannot be assessed from this evidence.","speaker":{"name":"Cui, Yubo, Guan, Xianchao, Zhang, Zheng","roleAtSource":"Article author"},"source":{"title":"Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.40309","publishedAt":"2026-10-29","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors propose TRAG to constrain visual representation drift using responses from a pre-trained text encoder to local variations in input embeddings.","excerpt":"propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"},{"text":"The authors say matching visual features treats all directions equally, while matching prediction distributions is limited by candidate texts and prediction probabilities.","excerpt":"Matching visual features constrains all directions equally, providing broad coverage without semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"},{"text":"TRAG builds constraints around semantic anchors and combines them using pre-trained image--anchor relevance to create an image-adaptive constraint.","excerpt":"TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"},{"text":"The authors report state-of-the-art clean and robust performance across architectures and tasks, and say further analysis shows TRAG helps preserve pre-trained image--text relationships.","excerpt":"Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.40309","checkedAt":"2026-11-12T16:20:10.237Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which architectures, tasks and performance measures support the authors’ reported clean and robust performance, and how large are the gains?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","summary":"Yubo Cui, Xianchao Guan and Zheng Zhang propose Textual Response-Aware Guidance (TRAG), a method for adversarial fine-tuning of vision-language models (VLMs). They argue that two existing ways of guiding fine-tuning have complementary limits: matching visual features treats all directions equally, while matching prediction distributions depends on the available candidate texts and prediction probabilities. Those limits, they say, can weaken pre-trained image--text relationships and degrade zero-shot performance.\n\nTRAG uses how a pre-trained text encoder responds to small changes in input embeddings to constrain visual representation drift. It builds constraints around semantic anchors, then combines them according to pre-trained image--anchor relevance to make the constraint image-adaptive. The authors report that experiments across architectures and tasks achieve state-of-the-art clean and robust performance, and that further analysis finds TRAG helps preserve pre-trained image--text relationships. The abstract does not identify the specific architectures, tasks, comparison methods or performance measures, so it does not show how broad or large those reported gains are.","question":"Which architectures, tasks and performance measures support the authors’ reported clean and robust performance, and how large are the gains?","claims":[{"text":"The authors propose TRAG to constrain visual representation drift using responses from a pre-trained text encoder to local variations in input embeddings.","excerpt":"propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings.","support":"source-stated"},{"text":"The authors say matching visual features treats all directions equally, while matching prediction distributions is limited by candidate texts and prediction probabilities.","excerpt":"Matching visual features constrains all directions equally, providing broad coverage without semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities.","support":"source-stated"},{"text":"TRAG builds constraints around semantic anchors and combines them using pre-trained image--anchor relevance to create an image-adaptive constraint.","excerpt":"TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint.","support":"source-stated"},{"text":"The authors report state-of-the-art clean and robust performance across architectures and tasks, and say further analysis shows TRAG helps preserve pre-trained image--text relationships.","excerpt":"Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.","support":"source-stated"}],"limitations":["The abstract does not identify the architectures, tasks, comparison methods or performance measures behind the reported results; the breadth and size of the reported gains cannot be assessed from this evidence."],"mediaId":"9ce097f6edce91e0b3c41d49fee5ae428d03bb82aefff751ccb5a259d1b3651b","correctionReason":null,"id":"dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613","publicationId":"1ceb993c-1d82-4095-84a3-8b8290b238a2","version":1,"sourceUrl":"https://arxiv.org/abs/2610.40309","sourceTitle":"Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-29","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-11T05:09:36.458Z","createdAt":"2026-11-11T23:17:57.436Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"9ce097f6edce91e0b3c41d49fee5ae428d03bb82aefff751ccb5a259d1b3651b","kind":"article","url":"https://arxiv.org/abs/2610.40309","sourceUrl":"https://arxiv.org/abs/2610.40309","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-11T23:17:11.862Z","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/2610.40309","checkedAt":"2026-11-12T16:20:10.237Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"9ce097f6edce91e0b3c41d49fee5ae428d03bb82aefff751ccb5a259d1b3651b","confidence":0.81,"provider":"typesafe","model":"jev-1.13.0","inputHash":"80a185ae32fc0465e9cbc9bb3df8d004fddcc825e86977abfc80d90599aa32e4","decidedAt":"2026-11-11T23:17:12.721Z"},"provenance":{"textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observationId":"f2ec5c0fb0453ba0f9f81e7b126802eb6c135deeea384632533b2e0f2412284f","retrievedAt":"2026-11-11T23:17:11.757Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"e04715a1-b2be-4d6c-8a27-971eb7743a1c"},"reviewer":{"model":"gpt-6-luna","runId":"6ee8b82b-e5b9-4f47-86e9-bed450e768ee"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T23:17:57.436Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Cui, Yubo","url":null,"kind":"unknown"},{"name":"Guan, Xianchao","url":null,"kind":"unknown"},{"name":"Zhang, Zheng","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Cui, Yubo","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Cui, Yubo","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.40309","locator":"head/meta[name=author|citation_author]"},"assertionId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-12T16:20:10.237Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","sourceUrl":"https://arxiv.org/abs/2610.40309","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z"}},{"actor":{"name":"Guan, Xianchao","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Guan, Xianchao","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.40309","locator":"head/meta[name=author|citation_author]"},"assertionId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-12T16:20:10.237Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","sourceUrl":"https://arxiv.org/abs/2610.40309","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z"}},{"actor":{"name":"Zhang, Zheng","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhang, Zheng","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.40309","locator":"head/meta[name=author|citation_author]"},"assertionId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-12T16:20:10.237Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","sourceUrl":"https://arxiv.org/abs/2610.40309","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z"}}],"participationCoverage":{"total":3,"returned":3,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.40309","snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z","basis":"publisher-page-metadata","sourceLinks":46,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-cs-ai-textual-response-aware-guidance-for-zero-shot-dbae875e","formats":{"html":"/readings/arxiv-cs-ai-textual-response-aware-guidance-for-zero-shot-dbae875e","markdown":"/readings/arxiv-cs-ai-textual-response-aware-guidance-for-zero-shot-dbae875e.md","json":"/readings/arxiv-cs-ai-textual-response-aware-guidance-for-zero-shot-dbae875e.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-zero-shot","name":"Zero-shot","explanation":"Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).","referenceUrl":"https://developers.google.com/machine-learning/glossary#zero-shot-prompting"},{"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-pre-training","name":"Pre-training","explanation":"A model's initial training on a large dataset, before further adaptation for particular uses.","referenceUrl":"https://developers.google.com/machine-learning/glossary#pre-training"},{"id":"term-embedding-vector","name":"Embedding vector","explanation":"A list of numbers that represents a piece of text, an image or another item so software can compare meanings.","referenceUrl":"https://developers.google.com/machine-learning/glossary#embedding-vector"}],"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":"dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613","revision":"1ceb993c-1d82-4095-84a3-8b8290b238a2","revisionKind":"publication-id","title":"arXiv cs.AI: Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","url":"https://escapemeridian.com/readings/arxiv-cs-ai-textual-response-aware-guidance-for-zero-shot-dbae875e"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.40309","revision":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","revisionKind":"sha256-source-text","title":"Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","url":"https://arxiv.org/abs/2610.40309"},"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":"dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613","title":"arXiv cs.AI: Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","argument":"Yubo Cui, Xianchao Guan and Zheng Zhang propose Textual Response-Aware Guidance (TRAG), a method for adversarial fine-tuning of vision-language models (VLMs). They argue that two existing ways of guiding fine-tuning have complementary limits: matching visual features treats all directions equally, while matching prediction distributions depends on the available candidate texts and prediction probabilities. Those limits, they say, can weaken pre-trained image--text relationships and degrade zero-shot performance.\n\nTRAG uses how a pre-trained text encoder responds to small changes in input embeddings to constrain visual representation drift. It builds constraints around semantic anchors, then combines them according to pre-trained image--anchor relevance to make the constraint image-adaptive. The authors report that experiments across architectures and tasks achieve state-of-the-art clean and robust performance, and that further analysis finds TRAG helps preserve pre-trained image--text relationships. The abstract does not identify the specific architectures, tasks, comparison methods or performance measures, so it does not show how broad or large those reported gains are.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract does not identify the architectures, tasks, comparison methods or performance measures behind the reported results; the breadth and size of the reported gains cannot be assessed from this evidence.","speaker":{"name":"Cui, Yubo, Guan, Xianchao, Zhang, Zheng","roleAtSource":"Article author"},"source":{"title":"Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.40309","publishedAt":"2026-10-29","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors propose TRAG to constrain visual representation drift using responses from a pre-trained text encoder to local variations in input embeddings.","excerpt":"propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"},{"text":"The authors say matching visual features treats all directions equally, while matching prediction distributions is limited by candidate texts and prediction probabilities.","excerpt":"Matching visual features constrains all directions equally, providing broad coverage without semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"},{"text":"TRAG builds constraints around semantic anchors and combines them using pre-trained image--anchor relevance to create an image-adaptive constraint.","excerpt":"TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"},{"text":"The authors report state-of-the-art clean and robust performance across architectures and tasks, and say further analysis shows TRAG helps preserve pre-trained image--text relationships.","excerpt":"Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.","url":"https://arxiv.org/abs/2610.40309","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.40309","checkedAt":"2026-11-12T16:20:10.237Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which architectures, tasks and performance measures support the authors’ reported clean and robust performance, and how large are the gains?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","summary":"Yubo Cui, Xianchao Guan and Zheng Zhang propose Textual Response-Aware Guidance (TRAG), a method for adversarial fine-tuning of vision-language models (VLMs). They argue that two existing ways of guiding fine-tuning have complementary limits: matching visual features treats all directions equally, while matching prediction distributions depends on the available candidate texts and prediction probabilities. Those limits, they say, can weaken pre-trained image--text relationships and degrade zero-shot performance.\n\nTRAG uses how a pre-trained text encoder responds to small changes in input embeddings to constrain visual representation drift. It builds constraints around semantic anchors, then combines them according to pre-trained image--anchor relevance to make the constraint image-adaptive. The authors report that experiments across architectures and tasks achieve state-of-the-art clean and robust performance, and that further analysis finds TRAG helps preserve pre-trained image--text relationships. The abstract does not identify the specific architectures, tasks, comparison methods or performance measures, so it does not show how broad or large those reported gains are.","question":"Which architectures, tasks and performance measures support the authors’ reported clean and robust performance, and how large are the gains?","claims":[{"text":"The authors propose TRAG to constrain visual representation drift using responses from a pre-trained text encoder to local variations in input embeddings.","excerpt":"propose Textual Response-Aware Guidance (TRAG), which constrains visual representation drift using the pre-trained text encoder's responses to local variations in input embeddings.","support":"source-stated"},{"text":"The authors say matching visual features treats all directions equally, while matching prediction distributions is limited by candidate texts and prediction probabilities.","excerpt":"Matching visual features constrains all directions equally, providing broad coverage without semantic emphasis. Matching prediction distributions prioritizes semantically relevant directions, but remains limited by the candidate texts and prediction probabilities.","support":"source-stated"},{"text":"TRAG builds constraints around semantic anchors and combines them using pre-trained image--anchor relevance to create an image-adaptive constraint.","excerpt":"TRAG constructs anchor-level constraints from the response directions and magnitudes around semantic anchors, then aggregates them using pre-trained image--anchor relevance, yielding an image-adaptive constraint.","support":"source-stated"},{"text":"The authors report state-of-the-art clean and robust performance across architectures and tasks, and say further analysis shows TRAG helps preserve pre-trained image--text relationships.","excerpt":"Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance. Further analysis shows that TRAG helps preserve pre-trained image--text relationships.","support":"source-stated"}],"limitations":["The abstract does not identify the architectures, tasks, comparison methods or performance measures behind the reported results; the breadth and size of the reported gains cannot be assessed from this evidence."],"mediaId":"9ce097f6edce91e0b3c41d49fee5ae428d03bb82aefff751ccb5a259d1b3651b","correctionReason":null,"id":"dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613","publicationId":"1ceb993c-1d82-4095-84a3-8b8290b238a2","version":1,"sourceUrl":"https://arxiv.org/abs/2610.40309","sourceTitle":"Textual Response-Aware Guidance for Zero-Shot Adversarial Robustness of Vision-Language Models","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-29","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-11T05:09:36.458Z","createdAt":"2026-11-11T23:17:57.436Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"9ce097f6edce91e0b3c41d49fee5ae428d03bb82aefff751ccb5a259d1b3651b","kind":"article","url":"https://arxiv.org/abs/2610.40309","sourceUrl":"https://arxiv.org/abs/2610.40309","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-11T23:17:11.862Z","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/2610.40309","checkedAt":"2026-11-12T16:20:10.237Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"9ce097f6edce91e0b3c41d49fee5ae428d03bb82aefff751ccb5a259d1b3651b","confidence":0.81,"provider":"typesafe","model":"jev-1.13.0","inputHash":"80a185ae32fc0465e9cbc9bb3df8d004fddcc825e86977abfc80d90599aa32e4","decidedAt":"2026-11-11T23:17:12.721Z"},"provenance":{"textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observationId":"f2ec5c0fb0453ba0f9f81e7b126802eb6c135deeea384632533b2e0f2412284f","retrievedAt":"2026-11-11T23:17:11.757Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"e04715a1-b2be-4d6c-8a27-971eb7743a1c"},"reviewer":{"model":"gpt-6-luna","runId":"6ee8b82b-e5b9-4f47-86e9-bed450e768ee"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T23:17:57.436Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Cui, Yubo","url":null,"kind":"unknown"},{"name":"Guan, Xianchao","url":null,"kind":"unknown"},{"name":"Zhang, Zheng","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Cui, Yubo","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Cui, Yubo","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.40309","locator":"head/meta[name=author|citation_author]"},"assertionId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-12T16:20:10.237Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","sourceUrl":"https://arxiv.org/abs/2610.40309","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z"}},{"actor":{"name":"Guan, Xianchao","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Guan, Xianchao","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.40309","locator":"head/meta[name=author|citation_author]"},"assertionId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-12T16:20:10.237Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","sourceUrl":"https://arxiv.org/abs/2610.40309","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z"}},{"actor":{"name":"Zhang, Zheng","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhang, Zheng","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.40309","locator":"head/meta[name=author|citation_author]"},"assertionId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-12T16:20:10.237Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","sourceUrl":"https://arxiv.org/abs/2610.40309","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z"}}],"participationCoverage":{"total":3,"returned":3,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.40309","snapshotId":"cf6b4d8ad08f7527e291f68106cf5dac285d03b1244d2a8375c489a88a785453","contentSha256":"d1b14e39d992a9a81cc562083b2e1cde445397c0e52813850767b31eb0d1de46","textHash":"3b1a40dc8ca50252ad8cb2939734790231d41e893fb35bd22d9e0688e7e5889f","observedAt":"2026-11-12T16:20:10.237Z","basis":"publisher-page-metadata","sourceLinks":46,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/dbae875eebfa11319a76bdfc6e01eabb9ec0068167ab8bb49d62b3fb12842613/source"}}}],"audits":[{"publicationId":"1ceb993c-1d82-4095-84a3-8b8290b238a2","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly names Yubo Cui, Xianchao Guan and Zheng Zhang as authors. It keeps the method's rationale and results attributed to them through phrases such as 'they argue', 'they say' and 'the authors report'.","excerpt":"Authors:Yubo Cui, Xianchao Guan, Zheng Zhang"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful but hard for a non-specialist to follow. It never states the problem the abstract opens with: VLMs generalize zero-shot but are vulnerable to adversarial attacks. Central terms are left unexplained, including adversarial attacks, adversarial fine-tuning, zero-shot, visual representation drift, semantic anchors, embeddings and clean vs robust performance. It also drops the complementary strengths the source names: broad coverage, and prioritizing semantically relevant directions. Smallest repair: open with the vulnerability-versus-generalization problem, keep both strengths, and add term-library tooltips or plain glosses for these terms.","excerpt":"Vision-language models (VLMs) exhibit strong zero-shot generalization but remain vulnerable to adversarial attacks."},{"dimension":"date","status":"pass","explanation":"The displayed original publication date of 2026-10-29 matches the v1 submission date. It is not the date of the 9 Nov 2026 revision or the 11 Nov feed entry.","excerpt":"Submitted on 29 Oct 2026 (v1), last revised 9 Nov 2026 (this version, v2)"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract and is presented as the authors' assertion. The state-of-the-art results are reported, not endorsed. The limitation and the question correctly note that the abstract omits the specific architectures, tasks, comparison methods and metrics. The draft invents no benefits or costs.","excerpt":"Extensive experiments across architectures and tasks demonstrate state-of-the-art clean and robust performance."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page for this paper, which is the direct original source. No better source media was found.","excerpt":null}],"createdAt":"2026-11-11T23:17:57.436Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}