{"id":"f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a","title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","argument":"The authors propose Vector Policy Optimization (VPO), an RL algorithm that trains language models to produce diverse solutions for downstream search. They argue that standard post-training optimizes a pre-specified scalar reward, which can lead to low-entropy response distributions and leave models ill-suited to search procedures that need varied candidate solutions. VPO uses multi-answer prompting and a set-level reward to train sets of solutions that specialize to different trade-offs in a vector reward space, where performance can be measured across multiple criteria. The authors report that across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search, including pass@k and best@k, with the gap widening as the search budget grows. They also report that evolutionary search with VPO models solves problems that GRPO models cannot solve at all. Their broader claim is that as test-time search becomes standard, diversity may need to become the default post-training objective. The supplied abstract summarizes these results but does not give task-by-task results or experimental details, so the scope of the reported gains is unclear from this evidence alone.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the arXiv abstract, not the paper's full experimental account; it does not provide task-by-task results or experimental details. The abstract reports comparative results but does not specify the search budgets or the magnitude of the reported performance gaps.","speaker":{"name":"Bahlous-Boldi, Ryan, Puri, Isha, Shenfeld, Idan, Kumar, Akarsh, Damani, Mehul, Risi, Sebastian, Khattab, Omar, Hong, Zhang-Wei, Agrawal, Pulkit","roleAtSource":"Article author"},"source":{"title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2605.22817","publishedAt":"2026-05-21","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors propose VPO, an RL algorithm intended to train models for diverse downstream reward functions and solutions.","excerpt":"We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"The authors say VPO trains sets of solutions to specialize to different trade-offs in a vector reward space.","excerpt":"VPO combines multi-answer prompting with a set-level reward to train the LLM to produce sets of solutions that specialize to different trade-offs in the vector reward space.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"Across four tasks, the authors report VPO matches or beats the strongest scalar RL baselines on test-time search, with a widening gap as the search budget grows.","excerpt":"Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"The authors report that VPO models enable evolutionary-search solutions to problems that GRPO models cannot solve.","excerpt":"For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"The authors argue that diversity may need to become the default post-training objective as test-time search becomes standard.","excerpt":"As test-time search becomes standard, diversity may need to become the default post-training objective.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2605.22817","checkedAt":"2026-11-09T03:36:08.088Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What were the task-by-task results, and how did the comparison between VPO and scalar RL baselines change as the search budget grew?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","summary":"The authors propose Vector Policy Optimization (VPO), an RL algorithm that trains language models to produce diverse solutions for downstream search. They argue that standard post-training optimizes a pre-specified scalar reward, which can lead to low-entropy response distributions and leave models ill-suited to search procedures that need varied candidate solutions. VPO uses multi-answer prompting and a set-level reward to train sets of solutions that specialize to different trade-offs in a vector reward space, where performance can be measured across multiple criteria. The authors report that across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search, including pass@k and best@k, with the gap widening as the search budget grows. They also report that evolutionary search with VPO models solves problems that GRPO models cannot solve at all. Their broader claim is that as test-time search becomes standard, diversity may need to become the default post-training objective. The supplied abstract summarizes these results but does not give task-by-task results or experimental details, so the scope of the reported gains is unclear from this evidence alone.","question":"What were the task-by-task results, and how did the comparison between VPO and scalar RL baselines change as the search budget grew?","claims":[{"text":"The authors propose VPO, an RL algorithm intended to train models for diverse downstream reward functions and solutions.","excerpt":"We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions.","support":"source-stated"},{"text":"The authors say VPO trains sets of solutions to specialize to different trade-offs in a vector reward space.","excerpt":"VPO combines multi-answer prompting with a set-level reward to train the LLM to produce sets of solutions that specialize to different trade-offs in the vector reward space.","support":"source-stated"},{"text":"Across four tasks, the authors report VPO matches or beats the strongest scalar RL baselines on test-time search, with a widening gap as the search budget grows.","excerpt":"Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows.","support":"source-stated"},{"text":"The authors report that VPO models enable evolutionary-search solutions to problems that GRPO models cannot solve.","excerpt":"For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all.","support":"source-stated"},{"text":"The authors argue that diversity may need to become the default post-training objective as test-time search becomes standard.","excerpt":"As test-time search becomes standard, diversity may need to become the default post-training objective.","support":"source-stated"}],"limitations":["The supplied evidence is the arXiv abstract, not the paper's full experimental account; it does not provide task-by-task results or experimental details.","The abstract reports comparative results but does not specify the search budgets or the magnitude of the reported performance gaps."],"mediaId":"d64fb055e7ebddc62a7aab31f702f1fb8c62eb22fee036852d37e6495fa11c78","correctionReason":null,"id":"f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a","publicationId":"91eb60b1-d628-492b-80c1-e946b6abd350","version":1,"sourceUrl":"https://arxiv.org/abs/2605.22817","sourceTitle":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-05-21","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:03:28.673Z","createdAt":"2026-11-09T02:32:56.196Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"d64fb055e7ebddc62a7aab31f702f1fb8c62eb22fee036852d37e6495fa11c78","kind":"article","url":"https://arxiv.org/abs/2605.22817","sourceUrl":"https://arxiv.org/abs/2605.22817","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T02:32:08.281Z","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/2605.22817","checkedAt":"2026-11-09T03:36:08.088Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"d64fb055e7ebddc62a7aab31f702f1fb8c62eb22fee036852d37e6495fa11c78","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"4231663377a93e014f08104e4d9916f3ad1910939bd4e078a653aaa7c4cbcacb","decidedAt":"2026-11-09T02:32:09.931Z"},"provenance":{"textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observationId":"3193bc7670da739c0d23aeca728a7162d18a7354d1cdc381d655c5e8f7036f0f","retrievedAt":"2026-11-09T02:32:08.025Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"d7ec4dd5-7ef9-4df6-8b1e-edb07b0cbe46"},"reviewer":{"model":"gpt-6-luna","runId":"e3fc7841-79fc-49ad-8ea7-c37caea45c20"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T02:32:56.196Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Bahlous-Boldi, Ryan","url":null,"kind":"unknown"},{"name":"Puri, Isha","url":null,"kind":"unknown"},{"name":"Shenfeld, Idan","url":null,"kind":"unknown"},{"name":"Kumar, Akarsh","url":null,"kind":"unknown"},{"name":"Damani, Mehul","url":null,"kind":"unknown"},{"name":"Risi, Sebastian","url":null,"kind":"unknown"},{"name":"Khattab, Omar","url":null,"kind":"unknown"},{"name":"Hong, Zhang-Wei","url":null,"kind":"unknown"},{"name":"Agrawal, Pulkit","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Bahlous-Boldi, Ryan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Bahlous-Boldi, Ryan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Puri, Isha","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Puri, Isha","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Shenfeld, Idan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shenfeld, Idan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Kumar, Akarsh","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kumar, Akarsh","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Damani, Mehul","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Damani, Mehul","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Risi, Sebastian","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Risi, Sebastian","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Khattab, Omar","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Khattab, Omar","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Hong, Zhang-Wei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Zhang-Wei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:7","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Agrawal, Pulkit","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Agrawal, Pulkit","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:8","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}}],"participationCoverage":{"total":9,"returned":9,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2605.22817","snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z","basis":"publisher-page-metadata","sourceLinks":53,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a/source"}},"canonicalUrl":"https://escapemeridian.com/readings/vector-policy-optimization-training-for-diversity-improves-test-f5844aa9","formats":{"html":"/readings/vector-policy-optimization-training-for-diversity-improves-test-f5844aa9","markdown":"/readings/vector-policy-optimization-training-for-diversity-improves-test-f5844aa9.md","json":"/readings/vector-policy-optimization-training-for-diversity-improves-test-f5844aa9.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-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"}],"entities":[{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.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":"f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a","revision":"91eb60b1-d628-492b-80c1-e946b6abd350","revisionKind":"publication-id","title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","url":"https://escapemeridian.com/readings/vector-policy-optimization-training-for-diversity-improves-test-f5844aa9"},"source":{"type":"source","id":"https://arxiv.org/abs/2605.22817","revision":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","revisionKind":"sha256-source-text","title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","url":"https://arxiv.org/abs/2605.22817"},"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":"f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a","title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","argument":"The authors propose Vector Policy Optimization (VPO), an RL algorithm that trains language models to produce diverse solutions for downstream search. They argue that standard post-training optimizes a pre-specified scalar reward, which can lead to low-entropy response distributions and leave models ill-suited to search procedures that need varied candidate solutions. VPO uses multi-answer prompting and a set-level reward to train sets of solutions that specialize to different trade-offs in a vector reward space, where performance can be measured across multiple criteria. The authors report that across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search, including pass@k and best@k, with the gap widening as the search budget grows. They also report that evolutionary search with VPO models solves problems that GRPO models cannot solve at all. Their broader claim is that as test-time search becomes standard, diversity may need to become the default post-training objective. The supplied abstract summarizes these results but does not give task-by-task results or experimental details, so the scope of the reported gains is unclear from this evidence alone.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the arXiv abstract, not the paper's full experimental account; it does not provide task-by-task results or experimental details. The abstract reports comparative results but does not specify the search budgets or the magnitude of the reported performance gaps.","speaker":{"name":"Bahlous-Boldi, Ryan, Puri, Isha, Shenfeld, Idan, Kumar, Akarsh, Damani, Mehul, Risi, Sebastian, Khattab, Omar, Hong, Zhang-Wei, Agrawal, Pulkit","roleAtSource":"Article author"},"source":{"title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2605.22817","publishedAt":"2026-05-21","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors propose VPO, an RL algorithm intended to train models for diverse downstream reward functions and solutions.","excerpt":"We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"The authors say VPO trains sets of solutions to specialize to different trade-offs in a vector reward space.","excerpt":"VPO combines multi-answer prompting with a set-level reward to train the LLM to produce sets of solutions that specialize to different trade-offs in the vector reward space.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"Across four tasks, the authors report VPO matches or beats the strongest scalar RL baselines on test-time search, with a widening gap as the search budget grows.","excerpt":"Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"The authors report that VPO models enable evolutionary-search solutions to problems that GRPO models cannot solve.","excerpt":"For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"},{"text":"The authors argue that diversity may need to become the default post-training objective as test-time search becomes standard.","excerpt":"As test-time search becomes standard, diversity may need to become the default post-training objective.","url":"https://arxiv.org/abs/2605.22817","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2605.22817","checkedAt":"2026-11-09T03:36:08.088Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What were the task-by-task results, and how did the comparison between VPO and scalar RL baselines change as the search budget grew?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","summary":"The authors propose Vector Policy Optimization (VPO), an RL algorithm that trains language models to produce diverse solutions for downstream search. They argue that standard post-training optimizes a pre-specified scalar reward, which can lead to low-entropy response distributions and leave models ill-suited to search procedures that need varied candidate solutions. VPO uses multi-answer prompting and a set-level reward to train sets of solutions that specialize to different trade-offs in a vector reward space, where performance can be measured across multiple criteria. The authors report that across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search, including pass@k and best@k, with the gap widening as the search budget grows. They also report that evolutionary search with VPO models solves problems that GRPO models cannot solve at all. Their broader claim is that as test-time search becomes standard, diversity may need to become the default post-training objective. The supplied abstract summarizes these results but does not give task-by-task results or experimental details, so the scope of the reported gains is unclear from this evidence alone.","question":"What were the task-by-task results, and how did the comparison between VPO and scalar RL baselines change as the search budget grew?","claims":[{"text":"The authors propose VPO, an RL algorithm intended to train models for diverse downstream reward functions and solutions.","excerpt":"We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions.","support":"source-stated"},{"text":"The authors say VPO trains sets of solutions to specialize to different trade-offs in a vector reward space.","excerpt":"VPO combines multi-answer prompting with a set-level reward to train the LLM to produce sets of solutions that specialize to different trade-offs in the vector reward space.","support":"source-stated"},{"text":"Across four tasks, the authors report VPO matches or beats the strongest scalar RL baselines on test-time search, with a widening gap as the search budget grows.","excerpt":"Across four tasks, VPO matches or beats the strongest scalar RL baselines on test-time search (e.g. pass@k and best@k), with the gap widening as the search budget grows.","support":"source-stated"},{"text":"The authors report that VPO models enable evolutionary-search solutions to problems that GRPO models cannot solve.","excerpt":"For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all.","support":"source-stated"},{"text":"The authors argue that diversity may need to become the default post-training objective as test-time search becomes standard.","excerpt":"As test-time search becomes standard, diversity may need to become the default post-training objective.","support":"source-stated"}],"limitations":["The supplied evidence is the arXiv abstract, not the paper's full experimental account; it does not provide task-by-task results or experimental details.","The abstract reports comparative results but does not specify the search budgets or the magnitude of the reported performance gaps."],"mediaId":"d64fb055e7ebddc62a7aab31f702f1fb8c62eb22fee036852d37e6495fa11c78","correctionReason":null,"id":"f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a","publicationId":"91eb60b1-d628-492b-80c1-e946b6abd350","version":1,"sourceUrl":"https://arxiv.org/abs/2605.22817","sourceTitle":"Vector Policy Optimization: Training for Diversity Improves Test-Time Search","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-05-21","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:03:28.673Z","createdAt":"2026-11-09T02:32:56.196Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"d64fb055e7ebddc62a7aab31f702f1fb8c62eb22fee036852d37e6495fa11c78","kind":"article","url":"https://arxiv.org/abs/2605.22817","sourceUrl":"https://arxiv.org/abs/2605.22817","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-09T02:32:08.281Z","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/2605.22817","checkedAt":"2026-11-09T03:36:08.088Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"d64fb055e7ebddc62a7aab31f702f1fb8c62eb22fee036852d37e6495fa11c78","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"4231663377a93e014f08104e4d9916f3ad1910939bd4e078a653aaa7c4cbcacb","decidedAt":"2026-11-09T02:32:09.931Z"},"provenance":{"textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observationId":"3193bc7670da739c0d23aeca728a7162d18a7354d1cdc381d655c5e8f7036f0f","retrievedAt":"2026-11-09T02:32:08.025Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"d7ec4dd5-7ef9-4df6-8b1e-edb07b0cbe46"},"reviewer":{"model":"gpt-6-luna","runId":"e3fc7841-79fc-49ad-8ea7-c37caea45c20"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-09T02:32:56.196Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Bahlous-Boldi, Ryan","url":null,"kind":"unknown"},{"name":"Puri, Isha","url":null,"kind":"unknown"},{"name":"Shenfeld, Idan","url":null,"kind":"unknown"},{"name":"Kumar, Akarsh","url":null,"kind":"unknown"},{"name":"Damani, Mehul","url":null,"kind":"unknown"},{"name":"Risi, Sebastian","url":null,"kind":"unknown"},{"name":"Khattab, Omar","url":null,"kind":"unknown"},{"name":"Hong, Zhang-Wei","url":null,"kind":"unknown"},{"name":"Agrawal, Pulkit","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Bahlous-Boldi, Ryan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Bahlous-Boldi, Ryan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Puri, Isha","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Puri, Isha","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Shenfeld, Idan","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Shenfeld, Idan","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Kumar, Akarsh","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kumar, Akarsh","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Damani, Mehul","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Damani, Mehul","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Risi, Sebastian","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Risi, Sebastian","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Khattab, Omar","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Khattab, Omar","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Hong, Zhang-Wei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Zhang-Wei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:7","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}},{"actor":{"name":"Agrawal, Pulkit","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Agrawal, Pulkit","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2605.22817","locator":"head/meta[name=author|citation_author]"},"assertionId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a:participation:8","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-09T03:36:08.088Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","sourceUrl":"https://arxiv.org/abs/2605.22817","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z"}}],"participationCoverage":{"total":9,"returned":9,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2605.22817","snapshotId":"b85f84c2040f9cea2b9aabff821c7cd48fd934df49f98a9f736ffe714fc1a12a","contentSha256":"56e60c3bdd139c5f3c06a8cc1f18edbcc3dab3e39f45b32dae2cee05642fe3f4","textHash":"86f1eb6df2605d044a28b21394480a0eeebd4dc2200d7bda27e0ffc40892be2d","observedAt":"2026-11-09T03:36:08.088Z","basis":"publisher-page-metadata","sourceLinks":53,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/f5844aa924f2d181c530e912e9635534af297a2e12d06591630c060d7f24066a/source"}}}],"audits":[{"publicationId":"91eb60b1-d628-492b-80c1-e946b6abd350","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading consistently attributes the proposal, results and broader claim to the paper's authors via arXiv, not to Escape Meridian. The reported results stay framed as author reports.","excerpt":"We propose Vector Policy Optimization (VPO), an RL algorithm that explicitly trains policies to anticipate diverse downstream reward functions and to produce diverse solutions."},{"dimension":"summary","status":"uncertain","explanation":"The summary faithfully tracks the abstract, but a non-specialist would struggle with central terms left unexplained: RL, post-training, scalar reward, low-entropy response distributions, test-time search, pass@k, best@k and GRPO. It also drops grounding context that aids understanding, such as AlphaEvolve as an example search procedure and per-test-case correctness as an example vector reward. Smallest repair: add short plain-language glosses or tooltips for these terms and restore one concrete example. The title and question are acceptable.","excerpt":"VPO exploits that rewards are often vector-valued in practice, like per-test-case correctness in code generation or, say, multiple different user personas or reward models."},{"dimension":"date","status":"pass","explanation":"The original publication date of 2026-05-21 matches the v1 submission date. It does not use the 5 Nov 2026 revision or the later feed date.","excerpt":"[Submitted on 21 May 2026 (v1), last revised 5 Nov 2026 (this version, v2)]"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract verbatim in substance. The hedged 'may need to become' is preserved. The limitations about missing task-by-task results, budgets and gap magnitudes accurately reflect an abstract-only evidence base. No unsupported benefits or costs are implied.","excerpt":"For evolutionary search, VPO models unlock problems that GRPO models cannot solve at all."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page for this exact paper. It is the direct original source.","excerpt":null}],"createdAt":"2026-11-09T02:32:56.196Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}