{"id":"5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b","title":"Peng and coauthors report lower knowledge-centric abstention after biomedical fine-tuning","argument":"The authors report that defined appropriate abstention on MedAbstain was already low in base models and fell further after biomedical fine-tuning: the median went from 8.3% to 4.2%, with decreases in 21 of 25 pairs. They audited 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench. The authors distinguish knowledge-centric abstention from value-centric refusal and say fine-tuning can produce unintended behavioral changes. Their results point to a tension: fine-tuning may improve domain-specific capability, but preserving appropriate caution remains a concern. The direction and size of changes varied across pairs, so the reported overall decrease does not describe every model. Escape Meridian interpretation: these benchmark findings do not establish how models behave in clinical use. The authors say the results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the paper's abstract, which reports benchmark results but does not establish how the models behave in clinical use. The abstract says changes vary across pairs but does not provide pair-specific results or explain the causes of that variation.","speaker":{"name":"Peng, Stephen, Nair, Vineet Jagadeesan, Broestl, Noah, Kleppe, Anne Friederike, Mills, Steven, Zolotov, Adi","roleAtSource":"Article author"},"source":{"title":"Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.08780","publishedAt":"2026-11-05","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors audited 25 biomedical domain fine-tuned models against 11 paired base checkpoints using MedAbstain and ClinDet-Bench.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"The authors report that median defined appropriate abstention on MedAbstain fell from 8.3% in base models to 4.2% after fine-tuning.","excerpt":"already low in base models (median 8.3% on MedAbstain), falls further after fine-tuning (median 4.2%","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"On MedAbstain, the authors report a decrease in 21 of the 25 model pairs, while the magnitude and direction of changes varied across pairs.","excerpt":"decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs.","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"The authors say the findings motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","excerpt":"These results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"The reported benchmark results do not establish model behavior in clinical use.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","url":"https://arxiv.org/abs/2611.08780","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.08780","checkedAt":"2026-11-13T15:06:07.958Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which fine-tuning methods, if any, can improve domain-specific capability while preserving appropriate caution across the models and benchmarks studied?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Peng and coauthors report lower knowledge-centric abstention after biomedical fine-tuning","summary":"The authors report that defined appropriate abstention on MedAbstain was already low in base models and fell further after biomedical fine-tuning: the median went from 8.3% to 4.2%, with decreases in 21 of 25 pairs. They audited 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench. The authors distinguish knowledge-centric abstention from value-centric refusal and say fine-tuning can produce unintended behavioral changes. Their results point to a tension: fine-tuning may improve domain-specific capability, but preserving appropriate caution remains a concern. The direction and size of changes varied across pairs, so the reported overall decrease does not describe every model. Escape Meridian interpretation: these benchmark findings do not establish how models behave in clinical use. The authors say the results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","question":"Which fine-tuning methods, if any, can improve domain-specific capability while preserving appropriate caution across the models and benchmarks studied?","claims":[{"text":"The authors audited 25 biomedical domain fine-tuned models against 11 paired base checkpoints using MedAbstain and ClinDet-Bench.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","support":"source-stated"},{"text":"The authors report that median defined appropriate abstention on MedAbstain fell from 8.3% in base models to 4.2% after fine-tuning.","excerpt":"already low in base models (median 8.3% on MedAbstain), falls further after fine-tuning (median 4.2%","support":"source-stated"},{"text":"On MedAbstain, the authors report a decrease in 21 of the 25 model pairs, while the magnitude and direction of changes varied across pairs.","excerpt":"decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs.","support":"source-stated"},{"text":"The authors say the findings motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","excerpt":"These results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","support":"source-stated"},{"text":"The reported benchmark results do not establish model behavior in clinical use.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","support":"interpretation"}],"limitations":["The evidence is the paper's abstract, which reports benchmark results but does not establish how the models behave in clinical use.","The abstract says changes vary across pairs but does not provide pair-specific results or explain the causes of that variation."],"mediaId":"833124248916cfb2704b549e644f31c7c8e7ef627cff4c8f10fe368f2c8a725a","correctionReason":null,"id":"5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b","publicationId":"112accaf-ee9e-443f-8cef-49a798b12328","version":1,"sourceUrl":"https://arxiv.org/abs/2611.08780","sourceTitle":"Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-05","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:48.187Z","createdAt":"2026-11-12T22:03:04.636Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"833124248916cfb2704b549e644f31c7c8e7ef627cff4c8f10fe368f2c8a725a","kind":"article","url":"https://arxiv.org/abs/2611.08780","sourceUrl":"https://arxiv.org/abs/2611.08780","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-12T22:01:58.307Z","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/2611.08780","checkedAt":"2026-11-13T15:06:07.958Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v6","mediaId":"833124248916cfb2704b549e644f31c7c8e7ef627cff4c8f10fe368f2c8a725a","confidence":0.74,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b22dad8edf4c9f3b7af5e739b3838adec5a3830fff9d056ac44aac5715b70458","decidedAt":"2026-11-12T22:02:01.301Z"},"provenance":{"textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observationId":"5adb7d3c52cd0ce094a96ece023617e4c131b69ee09075f5962827a7854f97b8","retrievedAt":"2026-11-12T22:01:57.891Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"2cc16758-cfcd-4abb-8302-936c50a8d05d"},"reviewer":{"model":"gpt-6-luna","runId":"6a9dc437-2058-4bda-8928-551f28c4888f"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-12T22:03:04.636Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Peng, Stephen","url":null,"kind":"unknown"},{"name":"Nair, Vineet Jagadeesan","url":null,"kind":"unknown"},{"name":"Broestl, Noah","url":null,"kind":"unknown"},{"name":"Kleppe, Anne Friederike","url":null,"kind":"unknown"},{"name":"Mills, Steven","url":null,"kind":"unknown"},{"name":"Zolotov, Adi","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Peng, Stephen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Peng, Stephen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Nair, Vineet Jagadeesan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Nair, Vineet Jagadeesan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Broestl, Noah","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Broestl, Noah","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Kleppe, Anne Friederike","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kleppe, Anne Friederike","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Mills, Steven","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Mills, Steven","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Zolotov, Adi","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zolotov, Adi","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.08780","snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b/source"}},"canonicalUrl":"https://escapemeridian.com/readings/peng-and-coauthors-report-lower-knowledge-centric-abstention-5b9ddba3","formats":{"html":"/readings/peng-and-coauthors-report-lower-knowledge-centric-abstention-5b9ddba3","markdown":"/readings/peng-and-coauthors-report-lower-knowledge-centric-abstention-5b9ddba3.md","json":"/readings/peng-and-coauthors-report-lower-knowledge-centric-abstention-5b9ddba3.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":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-fine-tuning","name":"Fine-tuning","explanation":"Further training of an existing model to adapt it to a particular task or use.","referenceUrl":"https://developers.google.com/machine-learning/glossary#fine-tuning"},{"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":"5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b","revision":"112accaf-ee9e-443f-8cef-49a798b12328","revisionKind":"publication-id","title":"Peng and coauthors report lower knowledge-centric abstention after biomedical fine-tuning","url":"https://escapemeridian.com/readings/peng-and-coauthors-report-lower-knowledge-centric-abstention-5b9ddba3"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.08780","revision":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","revisionKind":"sha256-source-text","title":"Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs","url":"https://arxiv.org/abs/2611.08780"},"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":"5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b","title":"Peng and coauthors report lower knowledge-centric abstention after biomedical fine-tuning","argument":"The authors report that defined appropriate abstention on MedAbstain was already low in base models and fell further after biomedical fine-tuning: the median went from 8.3% to 4.2%, with decreases in 21 of 25 pairs. They audited 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench. The authors distinguish knowledge-centric abstention from value-centric refusal and say fine-tuning can produce unintended behavioral changes. Their results point to a tension: fine-tuning may improve domain-specific capability, but preserving appropriate caution remains a concern. The direction and size of changes varied across pairs, so the reported overall decrease does not describe every model. Escape Meridian interpretation: these benchmark findings do not establish how models behave in clinical use. The authors say the results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the paper's abstract, which reports benchmark results but does not establish how the models behave in clinical use. The abstract says changes vary across pairs but does not provide pair-specific results or explain the causes of that variation.","speaker":{"name":"Peng, Stephen, Nair, Vineet Jagadeesan, Broestl, Noah, Kleppe, Anne Friederike, Mills, Steven, Zolotov, Adi","roleAtSource":"Article author"},"source":{"title":"Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.08780","publishedAt":"2026-11-05","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors audited 25 biomedical domain fine-tuned models against 11 paired base checkpoints using MedAbstain and ClinDet-Bench.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"The authors report that median defined appropriate abstention on MedAbstain fell from 8.3% in base models to 4.2% after fine-tuning.","excerpt":"already low in base models (median 8.3% on MedAbstain), falls further after fine-tuning (median 4.2%","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"On MedAbstain, the authors report a decrease in 21 of the 25 model pairs, while the magnitude and direction of changes varied across pairs.","excerpt":"decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs.","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"The authors say the findings motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","excerpt":"These results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","url":"https://arxiv.org/abs/2611.08780","relationship":"Source-stated"},{"text":"The reported benchmark results do not establish model behavior in clinical use.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","url":"https://arxiv.org/abs/2611.08780","relationship":"Escape Meridian interpretation"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.08780","checkedAt":"2026-11-13T15:06:07.958Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which fine-tuning methods, if any, can improve domain-specific capability while preserving appropriate caution across the models and benchmarks studied?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Peng and coauthors report lower knowledge-centric abstention after biomedical fine-tuning","summary":"The authors report that defined appropriate abstention on MedAbstain was already low in base models and fell further after biomedical fine-tuning: the median went from 8.3% to 4.2%, with decreases in 21 of 25 pairs. They audited 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench. The authors distinguish knowledge-centric abstention from value-centric refusal and say fine-tuning can produce unintended behavioral changes. Their results point to a tension: fine-tuning may improve domain-specific capability, but preserving appropriate caution remains a concern. The direction and size of changes varied across pairs, so the reported overall decrease does not describe every model. Escape Meridian interpretation: these benchmark findings do not establish how models behave in clinical use. The authors say the results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","question":"Which fine-tuning methods, if any, can improve domain-specific capability while preserving appropriate caution across the models and benchmarks studied?","claims":[{"text":"The authors audited 25 biomedical domain fine-tuned models against 11 paired base checkpoints using MedAbstain and ClinDet-Bench.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","support":"source-stated"},{"text":"The authors report that median defined appropriate abstention on MedAbstain fell from 8.3% in base models to 4.2% after fine-tuning.","excerpt":"already low in base models (median 8.3% on MedAbstain), falls further after fine-tuning (median 4.2%","support":"source-stated"},{"text":"On MedAbstain, the authors report a decrease in 21 of the 25 model pairs, while the magnitude and direction of changes varied across pairs.","excerpt":"decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs.","support":"source-stated"},{"text":"The authors say the findings motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","excerpt":"These results motivate fine-tuning methods that improve domain-specific capability while preserving appropriate caution.","support":"source-stated"},{"text":"The reported benchmark results do not establish model behavior in clinical use.","excerpt":"We audit 25 biomedical domain fine-tuned large language models against 11 paired base checkpoints on MedAbstain and ClinDet-Bench","support":"interpretation"}],"limitations":["The evidence is the paper's abstract, which reports benchmark results but does not establish how the models behave in clinical use.","The abstract says changes vary across pairs but does not provide pair-specific results or explain the causes of that variation."],"mediaId":"833124248916cfb2704b549e644f31c7c8e7ef627cff4c8f10fe368f2c8a725a","correctionReason":null,"id":"5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b","publicationId":"112accaf-ee9e-443f-8cef-49a798b12328","version":1,"sourceUrl":"https://arxiv.org/abs/2611.08780","sourceTitle":"Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-05","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:48.187Z","createdAt":"2026-11-12T22:03:04.636Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"833124248916cfb2704b549e644f31c7c8e7ef627cff4c8f10fe368f2c8a725a","kind":"article","url":"https://arxiv.org/abs/2611.08780","sourceUrl":"https://arxiv.org/abs/2611.08780","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-12T22:01:58.307Z","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/2611.08780","checkedAt":"2026-11-13T15:06:07.958Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v6","mediaId":"833124248916cfb2704b549e644f31c7c8e7ef627cff4c8f10fe368f2c8a725a","confidence":0.74,"provider":"typesafe","model":"jev-1.13.0","inputHash":"b22dad8edf4c9f3b7af5e739b3838adec5a3830fff9d056ac44aac5715b70458","decidedAt":"2026-11-12T22:02:01.301Z"},"provenance":{"textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observationId":"5adb7d3c52cd0ce094a96ece023617e4c131b69ee09075f5962827a7854f97b8","retrievedAt":"2026-11-12T22:01:57.891Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"2cc16758-cfcd-4abb-8302-936c50a8d05d"},"reviewer":{"model":"gpt-6-luna","runId":"6a9dc437-2058-4bda-8928-551f28c4888f"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-12T22:03:04.636Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Peng, Stephen","url":null,"kind":"unknown"},{"name":"Nair, Vineet Jagadeesan","url":null,"kind":"unknown"},{"name":"Broestl, Noah","url":null,"kind":"unknown"},{"name":"Kleppe, Anne Friederike","url":null,"kind":"unknown"},{"name":"Mills, Steven","url":null,"kind":"unknown"},{"name":"Zolotov, Adi","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Peng, Stephen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Peng, Stephen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Nair, Vineet Jagadeesan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Nair, Vineet Jagadeesan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Broestl, Noah","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Broestl, Noah","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Kleppe, Anne Friederike","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kleppe, Anne Friederike","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Mills, Steven","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Mills, Steven","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}},{"actor":{"name":"Zolotov, Adi","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zolotov, Adi","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.08780","locator":"head/meta[name=author|citation_author]"},"assertionId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-13T15:06:07.958Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","sourceUrl":"https://arxiv.org/abs/2611.08780","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.08780","snapshotId":"ca1314083d2751d5a23bc2ecd8625c91fac04355d118457db055b802a916f09a","contentSha256":"72c1bca4e6d2e1f12f7123583b3bc34a5fccb0545f6c0975a706a8330fcd1088","textHash":"38967d37c1b67aff3fbbb1d76c1380f4ea4eaf08ef4ba200b878b580b5098e44","observedAt":"2026-11-13T15:06:07.958Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/5b9ddba30f96dc3c2c036ebaebfeb0c1ae19435422ad11deba2b079e788e296b/source"}}}],"audits":[{"publicationId":"112accaf-ee9e-443f-8cef-49a798b12328","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly credits the findings to the paper's authors, Stephen Peng and coauthors, as reported in an arXiv abstract. Source claims stay attributed with phrases such as 'the authors report,' and the Escape Meridian interpretation is labeled separately.","excerpt":"Authors:Stephen Peng, Vineet Jagadeesan Nair, Noah Broestl, Anne Friederike Kleppe, Steven Mills, Adi Zolotov"},{"dimension":"summary","status":"uncertain","explanation":"The figures and conditions are faithful, but the central idea is never explained in plain language. 'Knowledge-centric abstention' means a model declining to answer when it lacks the knowledge to answer reliably, and the reading does not say so. It also leaves 'value-centric refusal,' 'base checkpoints' and 'fine-tuning' unexplained, and copies the awkward phrase 'defined appropriate abstention.' A non-specialist may not grasp what fell or why it matters. Smallest repair: one plain-language clause saying the models less often declined to answer biomedical questions when declining was appropriate, plus term-library tooltips for these terms.","excerpt":"Where prior work in this space has studied value-centric refusal, this work focuses on knowledge-centric abstention."},{"dimension":"date","status":"pass","explanation":"The originalPublished date of 2026-11-05 matches the arXiv v1 submission date. It correctly differs from the later feed date of 2026-11-09.","excerpt":"[v1]\nThu, 5 Nov 2026 23:03:11 UTC (619 KB)"},{"dimension":"support","status":"pass","explanation":"The numeric claims (25 models, 11 base checkpoints, 8.3% to 4.2%, decreases in 21 of 25 pairs, variation across pairs) and the authors' motivation statement match the abstract. The claim that fine-tuning may improve capability is hedged and grounded in the abstract's 'promises' framing. The clinical-use caveat is labeled as interpretation and does not overreach. The question follows from the authors' stated motivation.","excerpt":"falls further after fine-tuning (median 4.2%, decreasing in 21 of 25 pairs on MedAbstain), and that the magnitude and direction of change are heterogeneous across pairs."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page, the direct original source.","excerpt":null}],"createdAt":"2026-11-12T22:03:04.636Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}