{"id":"164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9","title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","argument":"The paper’s authors, Eunkyeong Lee, Junhyung Moon, Seunggwan Hong and Dongsuk Oh, argue that probabilistic safety judges can leave a blind spot: harm they confidently pass is not sent to a stronger LLM judge by a confidence gate. They propose validating the judge before and during training, measuring both the share and amount of confidently passed harm, and re-validating after a judge change. In tests using two versions of Jev and an open guard classifier, harm passed under Jev’s probability reward shrank in all 11 runs under the primary reference, but mostly more slowly than caught harm, and concentrated in the blind spot. The paper also reports that binarizing Jev met the escape criterion on Llama, and that the open classifier showed an escape signal in one seed under a second reference. Frontier LLM judges caught at most one of 10 human-labeled confident misses. Without a trained reward model, adding a helpfulness term to the probability reward lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint. These results illustrate both a potential training benefit and a limitation: optimizing against a judge can leave harm that the judge confidently misses. The abstract reports experiments on specific judges, models and references; it does not establish how well the protocol generalizes beyond them.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The available evidence is the paper’s abstract, so it does not provide the experiment details needed to assess the reported results fully. The reported tests cover two versions of Jev, an open guard classifier, and specified models and references; the abstract does not establish how the findings generalize beyond those tests.","speaker":{"name":"Lee, Eunkyeong, Moon, Junhyung, Hong, Seunggwan, Oh, Dongsuk","roleAtSource":"Article author"},"source":{"title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.20955","publishedAt":"2026-11-16","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors say a confidence gate routes only uncertain verdicts to a stronger LLM judge, leaving confidently passed harm in a blind spot that policies can exploit.","excerpt":"A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"The authors propose validating a judge before and during training, measuring the share and amount of confidently passed harm separately, and re-validating after a judge change.","excerpt":"validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"Under Jev’s probability reward, passed harm shrank in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrated in the blind spot.","excerpt":"Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"The abstract reports that binarizing Jev met the escape criterion on Llama, while the open guard classifier showed an escape signal in one seed under a second reference.","excerpt":"Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"The authors report that frontier LLM judges caught at most one of 10 human-labeled confident misses.","excerpt":"even frontier LLM judges catch at most one of 10 human-labeled confident misses.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"Without a trained reward model, the probability reward with a helpfulness term lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","excerpt":"Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.20955","checkedAt":"2026-11-17T11:46:08.980Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does the proposed validation protocol identify confidently passed harm across other judges, models and references?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","summary":"The paper’s authors, Eunkyeong Lee, Junhyung Moon, Seunggwan Hong and Dongsuk Oh, argue that probabilistic safety judges can leave a blind spot: harm they confidently pass is not sent to a stronger LLM judge by a confidence gate. They propose validating the judge before and during training, measuring both the share and amount of confidently passed harm, and re-validating after a judge change. In tests using two versions of Jev and an open guard classifier, harm passed under Jev’s probability reward shrank in all 11 runs under the primary reference, but mostly more slowly than caught harm, and concentrated in the blind spot. The paper also reports that binarizing Jev met the escape criterion on Llama, and that the open classifier showed an escape signal in one seed under a second reference. Frontier LLM judges caught at most one of 10 human-labeled confident misses. Without a trained reward model, adding a helpfulness term to the probability reward lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint. These results illustrate both a potential training benefit and a limitation: optimizing against a judge can leave harm that the judge confidently misses. The abstract reports experiments on specific judges, models and references; it does not establish how well the protocol generalizes beyond them.","question":"How well does the proposed validation protocol identify confidently passed harm across other judges, models and references?","claims":[{"text":"The authors say a confidence gate routes only uncertain verdicts to a stronger LLM judge, leaving confidently passed harm in a blind spot that policies can exploit.","excerpt":"A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot","support":"source-stated"},{"text":"The authors propose validating a judge before and during training, measuring the share and amount of confidently passed harm separately, and re-validating after a judge change.","excerpt":"validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change.","support":"source-stated"},{"text":"Under Jev’s probability reward, passed harm shrank in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrated in the blind spot.","excerpt":"Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot.","support":"source-stated"},{"text":"The abstract reports that binarizing Jev met the escape criterion on Llama, while the open guard classifier showed an escape signal in one seed under a second reference.","excerpt":"Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference.","support":"source-stated"},{"text":"The authors report that frontier LLM judges caught at most one of 10 human-labeled confident misses.","excerpt":"even frontier LLM judges catch at most one of 10 human-labeled confident misses.","support":"source-stated"},{"text":"Without a trained reward model, the probability reward with a helpfulness term lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","excerpt":"Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","support":"source-stated"}],"limitations":["The available evidence is the paper’s abstract, so it does not provide the experiment details needed to assess the reported results fully.","The reported tests cover two versions of Jev, an open guard classifier, and specified models and references; the abstract does not establish how the findings generalize beyond those tests."],"mediaId":"a36eb7da82f12fee5cf5887a6decea2079358740b0a81b066f8be3775905ac69","correctionReason":null,"id":"164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9","publicationId":"217a9ef7-3a60-43f0-8ac0-b346f0b1d02c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.20955","sourceTitle":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-16","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-17T08:30:54.112Z","createdAt":"2026-11-17T11:31:57.280Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"a36eb7da82f12fee5cf5887a6decea2079358740b0a81b066f8be3775905ac69","kind":"article","url":"https://arxiv.org/abs/2611.20955","sourceUrl":"https://arxiv.org/abs/2611.20955","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-17T11:20:04.296Z","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/2611.20955","checkedAt":"2026-11-17T11:46:08.980Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"a36eb7da82f12fee5cf5887a6decea2079358740b0a81b066f8be3775905ac69","confidence":0.64,"provider":"typesafe","model":"jev-1.13.0","inputHash":"cd4b051cf8f448fef5297fd7a5effb3f18294126709f0db8154ca50b6386f4fe","decidedAt":"2026-11-17T11:20:06.121Z"},"provenance":{"textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observationId":"8d0a174ab38f1ad5de60bfb0971dd9edb6e7c28984ced102f1353f6453ed581c","retrievedAt":"2026-11-17T11:20:04.026Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"00708694-00d3-4068-8aff-db197425d623"},"reviewer":{"model":"gpt-6-luna","runId":"ddba0ee8-a64f-424c-8b13-2f8258d605c0"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-17T11:31:57.280Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Lee, Eunkyeong","url":null,"kind":"unknown"},{"name":"Moon, Junhyung","url":null,"kind":"unknown"},{"name":"Hong, Seunggwan","url":null,"kind":"unknown"},{"name":"Oh, Dongsuk","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Lee, Eunkyeong","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Lee, Eunkyeong","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}},{"actor":{"name":"Moon, Junhyung","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Moon, Junhyung","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}},{"actor":{"name":"Hong, Seunggwan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Seunggwan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}},{"actor":{"name":"Oh, Dongsuk","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Oh, Dongsuk","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}}],"participationCoverage":{"total":4,"returned":4,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.20955","snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z","basis":"publisher-page-metadata","sourceLinks":43,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9/source"}},"canonicalUrl":"https://escapemeridian.com/readings/jev-as-a-reward-validating-probabilistic-safety-judges-for-164f09f7","formats":{"html":"/readings/jev-as-a-reward-validating-probabilistic-safety-judges-for-164f09f7","markdown":"/readings/jev-as-a-reward-validating-probabilistic-safety-judges-for-164f09f7.md","json":"/readings/jev-as-a-reward-validating-probabilistic-safety-judges-for-164f09f7.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-reinforcement-learning","name":"Reinforcement learning","explanation":"Training by trial and error, in which a system is rewarded for the outcomes its designers want.","referenceUrl":"https://developers.google.com/machine-learning/glossary#reinforcement-learning-rl"},{"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"}],"entities":[{"id":"jev","name":"Jev","kind":"model","context":"Jev is TypeSafe AI’s model for fast evaluation and routing decisions inside AI systems.","referenceUrls":["https://typesafe.ai/blog/introducing-system-one-models-and-jev","https://typesafe.ai/"],"path":"/entities/jev","json":"/entities/jev.json","markdown":"/entities/jev.md"},{"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":"164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9","revision":"217a9ef7-3a60-43f0-8ac0-b346f0b1d02c","revisionKind":"publication-id","title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","url":"https://escapemeridian.com/readings/jev-as-a-reward-validating-probabilistic-safety-judges-for-164f09f7"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.20955","revision":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","revisionKind":"sha256-source-text","title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","url":"https://arxiv.org/abs/2611.20955"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9","title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","argument":"The paper’s authors, Eunkyeong Lee, Junhyung Moon, Seunggwan Hong and Dongsuk Oh, argue that probabilistic safety judges can leave a blind spot: harm they confidently pass is not sent to a stronger LLM judge by a confidence gate. They propose validating the judge before and during training, measuring both the share and amount of confidently passed harm, and re-validating after a judge change. In tests using two versions of Jev and an open guard classifier, harm passed under Jev’s probability reward shrank in all 11 runs under the primary reference, but mostly more slowly than caught harm, and concentrated in the blind spot. The paper also reports that binarizing Jev met the escape criterion on Llama, and that the open classifier showed an escape signal in one seed under a second reference. Frontier LLM judges caught at most one of 10 human-labeled confident misses. Without a trained reward model, adding a helpfulness term to the probability reward lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint. These results illustrate both a potential training benefit and a limitation: optimizing against a judge can leave harm that the judge confidently misses. The abstract reports experiments on specific judges, models and references; it does not establish how well the protocol generalizes beyond them.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The available evidence is the paper’s abstract, so it does not provide the experiment details needed to assess the reported results fully. The reported tests cover two versions of Jev, an open guard classifier, and specified models and references; the abstract does not establish how the findings generalize beyond those tests.","speaker":{"name":"Lee, Eunkyeong, Moon, Junhyung, Hong, Seunggwan, Oh, Dongsuk","roleAtSource":"Article author"},"source":{"title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.20955","publishedAt":"2026-11-16","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors say a confidence gate routes only uncertain verdicts to a stronger LLM judge, leaving confidently passed harm in a blind spot that policies can exploit.","excerpt":"A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"The authors propose validating a judge before and during training, measuring the share and amount of confidently passed harm separately, and re-validating after a judge change.","excerpt":"validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"Under Jev’s probability reward, passed harm shrank in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrated in the blind spot.","excerpt":"Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"The abstract reports that binarizing Jev met the escape criterion on Llama, while the open guard classifier showed an escape signal in one seed under a second reference.","excerpt":"Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"The authors report that frontier LLM judges caught at most one of 10 human-labeled confident misses.","excerpt":"even frontier LLM judges catch at most one of 10 human-labeled confident misses.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"},{"text":"Without a trained reward model, the probability reward with a helpfulness term lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","excerpt":"Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","url":"https://arxiv.org/abs/2611.20955","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.20955","checkedAt":"2026-11-17T11:46:08.980Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does the proposed validation protocol identify confidently passed harm across other judges, models and references?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","summary":"The paper’s authors, Eunkyeong Lee, Junhyung Moon, Seunggwan Hong and Dongsuk Oh, argue that probabilistic safety judges can leave a blind spot: harm they confidently pass is not sent to a stronger LLM judge by a confidence gate. They propose validating the judge before and during training, measuring both the share and amount of confidently passed harm, and re-validating after a judge change. In tests using two versions of Jev and an open guard classifier, harm passed under Jev’s probability reward shrank in all 11 runs under the primary reference, but mostly more slowly than caught harm, and concentrated in the blind spot. The paper also reports that binarizing Jev met the escape criterion on Llama, and that the open classifier showed an escape signal in one seed under a second reference. Frontier LLM judges caught at most one of 10 human-labeled confident misses. Without a trained reward model, adding a helpfulness term to the probability reward lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint. These results illustrate both a potential training benefit and a limitation: optimizing against a judge can leave harm that the judge confidently misses. The abstract reports experiments on specific judges, models and references; it does not establish how well the protocol generalizes beyond them.","question":"How well does the proposed validation protocol identify confidently passed harm across other judges, models and references?","claims":[{"text":"The authors say a confidence gate routes only uncertain verdicts to a stronger LLM judge, leaving confidently passed harm in a blind spot that policies can exploit.","excerpt":"A confidence gate sends only uncertain verdicts to a stronger LLM judge, which leaves confidently passed harm in a blind spot. Policies can exploit this blind spot","support":"source-stated"},{"text":"The authors propose validating a judge before and during training, measuring the share and amount of confidently passed harm separately, and re-validating after a judge change.","excerpt":"validates such a judge before and during training, separates the share of confidently passed harm from its amount, and re-validates it after each judge change.","support":"source-stated"},{"text":"Under Jev’s probability reward, passed harm shrank in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrated in the blind spot.","excerpt":"Under the probability reward of Jev, passed harm shrinks in all 11 runs under the primary reference, mostly more slowly than caught harm, and concentrates in the blind spot.","support":"source-stated"},{"text":"The abstract reports that binarizing Jev met the escape criterion on Llama, while the open guard classifier showed an escape signal in one seed under a second reference.","excerpt":"Binarizing this judge meets the escape criterion on Llama, and the open guard classifier shows an escape signal in one seed under a second reference.","support":"source-stated"},{"text":"The authors report that frontier LLM judges caught at most one of 10 human-labeled confident misses.","excerpt":"even frontier LLM judges catch at most one of 10 human-labeled confident misses.","support":"source-stated"},{"text":"Without a trained reward model, the probability reward with a helpfulness term lowered HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","excerpt":"Without a trained reward model, the probability reward with a helpfulness term lowers HarmBench attack success on Llama-3.1-8B from 0.287 to at most 0.033 at the final checkpoint.","support":"source-stated"}],"limitations":["The available evidence is the paper’s abstract, so it does not provide the experiment details needed to assess the reported results fully.","The reported tests cover two versions of Jev, an open guard classifier, and specified models and references; the abstract does not establish how the findings generalize beyond those tests."],"mediaId":"a36eb7da82f12fee5cf5887a6decea2079358740b0a81b066f8be3775905ac69","correctionReason":null,"id":"164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9","publicationId":"217a9ef7-3a60-43f0-8ac0-b346f0b1d02c","version":1,"sourceUrl":"https://arxiv.org/abs/2611.20955","sourceTitle":"Jev-as-a-Reward: Validating Probabilistic Safety Judges for Reinforcement Learning","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-16","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-17T08:30:54.112Z","createdAt":"2026-11-17T11:31:57.280Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"a36eb7da82f12fee5cf5887a6decea2079358740b0a81b066f8be3775905ac69","kind":"article","url":"https://arxiv.org/abs/2611.20955","sourceUrl":"https://arxiv.org/abs/2611.20955","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-17T11:20:04.296Z","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/2611.20955","checkedAt":"2026-11-17T11:46:08.980Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"a36eb7da82f12fee5cf5887a6decea2079358740b0a81b066f8be3775905ac69","confidence":0.64,"provider":"typesafe","model":"jev-1.13.0","inputHash":"cd4b051cf8f448fef5297fd7a5effb3f18294126709f0db8154ca50b6386f4fe","decidedAt":"2026-11-17T11:20:06.121Z"},"provenance":{"textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observationId":"8d0a174ab38f1ad5de60bfb0971dd9edb6e7c28984ced102f1353f6453ed581c","retrievedAt":"2026-11-17T11:20:04.026Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"00708694-00d3-4068-8aff-db197425d623"},"reviewer":{"model":"gpt-6-luna","runId":"ddba0ee8-a64f-424c-8b13-2f8258d605c0"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-17T11:31:57.280Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Lee, Eunkyeong","url":null,"kind":"unknown"},{"name":"Moon, Junhyung","url":null,"kind":"unknown"},{"name":"Hong, Seunggwan","url":null,"kind":"unknown"},{"name":"Oh, Dongsuk","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Lee, Eunkyeong","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Lee, Eunkyeong","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}},{"actor":{"name":"Moon, Junhyung","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Moon, Junhyung","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}},{"actor":{"name":"Hong, Seunggwan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Seunggwan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}},{"actor":{"name":"Oh, Dongsuk","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Oh, Dongsuk","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.20955","locator":"head/meta[name=author|citation_author]"},"assertionId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-17T11:46:08.980Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","sourceUrl":"https://arxiv.org/abs/2611.20955","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z"}}],"participationCoverage":{"total":4,"returned":4,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.20955","snapshotId":"466b315181358e171ae750c0b19c5b32abfc7d09430f81d75457922339500ef8","contentSha256":"9524d4cb2d2b1797cbcfe859da62501a3a62fa93508b6a5f58301fca9881c5d7","textHash":"00b78f6870f082e00003dbaacf1474c5e39e0d45717e8733b951a113b4f6d323","observedAt":"2026-11-17T11:46:08.980Z","basis":"publisher-page-metadata","sourceLinks":43,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/164f09f725f10b36ca50276cb3fd22b5051da1cc9660b3a1d6cf8b2cc53e11c9/source"}}}],"audits":[{"publicationId":"217a9ef7-3a60-43f0-8ac0-b346f0b1d02c","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly credits the four listed authors and frames the findings as the authors' reported arguments and results, not as established fact.","excerpt":"Authors:Eunkyeong Lee, Junhyung Moon, Seunggwan Hong, Dongsuk Oh"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful, but central terms are left unexplained for a general reader: escape criterion, escape signal in one seed, primary or second reference, binarizing, rollouts, and HarmBench attack success. It also never says that Jev is a commercial decision-only judge, or that commercial judges release new versions, which is why the paper calls for re-validation. It drops the abstract's point that guard classifiers share this blind spot. Smallest repair: identify Jev as a commercial judge, add the version-change rationale, and explain or tooltip escape criterion, seed, reference and attack success.","excerpt":"We apply it to two versions of Jev, a commercial decision-only judge, and to an open guard classifier."},{"dimension":"date","status":"pass","explanation":"originalPublished is 2026-11-16, which matches the arXiv v1 submission date. The feed date of 2026-11-17 is later and correctly not used.","excerpt":"[Submitted on 16 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Every claim excerpt is verbatim from the abstract. The closing interpretation, that the results show a potential training benefit and a blind-spot limitation, follows from the reported HarmBench reduction and the confident-miss findings. The generalization caveat and the question are appropriately framed and do not invent benefits or costs.","excerpt":"Policies can exploit this blind spot, and commercial judges release new versions."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-17T11:31:57.280Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}