{"id":"0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6","title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","argument":"The authors argue that a fixed rollout cap in RL with verifiable rewards (RLVR) can shape how a model uses inference compute after training, not just reduce training cost. They call this effect budget internalization: models trained with a cap use few additional tokens beyond that budget, perform better than base models near it, but stop improving and fall below the base models at larger budgets.\n\nThe reported effect appears across three models and two reasoning training tasks, and transfers from math training to science and coding tasks. The authors’ ablations suggest it is linked to length-censored post-training rather than an RL-specific objective: SFT on artificially truncated traces reproduces the effect, while SFT on naturally short traces does not. Some changes preserve more test-time scaling, but stronger KL regularization reduces in-budget accuracy; forcing answers on truncated rollouts largely removes budget internalization but also most of the in-budget gain.\n\nThe paper describes a tradeoff: a training cap can help when inference stays near that budget, while limiting gains from additional inference compute. The abstract’s evidence comes from the models, tasks, and training comparisons it reports; how broadly the effect applies beyond them remains open.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports results for three named models, two reasoning training tasks, and specified training-objective ablations; it does not establish how broadly the effect applies beyond those settings.","speaker":{"name":"Zavidnyi, Ilia, Kovrigin, Alexander, Puchkova, Arina, Gelvan, Kirill, Titov, Sergey","roleAtSource":"Article author"},"source":{"title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.26172","publishedAt":"2026-11-19","dateBasis":"Original submission date in the publisher’s citation record."},"evidence":[{"text":"The authors report that models trained with a fixed rollout cap use few tokens beyond the training budget, gain accuracy near it, and perform worse than base models at larger budgets.","excerpt":"After training, models use few additional tokens beyond the training budget. They outperform the base models near that budget, but their accuracy stops improving and falls below the base model at larger budgets.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"},{"text":"The authors report the effect across three models and two reasoning training tasks, with transfer from math training to science and coding tasks.","excerpt":"We observe it across Qwen3.5-4B, Gemma4-E2B, and Nemotron-3-Nano-4B and two reasoning training tasks: mathematical reasoning and diverse, procedurally generated problems. This effect also transfers from math training to science and coding tasks.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"},{"text":"The authors link budget internalization to length-censored post-training rather than an RL-specific objective term, based in part on results from SFT on truncated traces.","excerpt":"SFT on artificially truncated traces reproduces budget internalization, whereas SFT on naturally short traces does not, which shows that the effect is not specific to RL. This result links budget internalization to length-censored post-training rather than to an RL-specific objective term.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"},{"text":"The authors report that interventions preserving more test-time scaling can reduce in-budget accuracy or remove much of the in-budget gain.","excerpt":"Stronger KL regularization and lower training-data truncation preserve more scaling, but stronger KL reduces in-budget accuracy. Forcing an answer on truncated training rollouts largely removes budget internalization, but also removes most of the in-budget gain.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.26172","checkedAt":"2026-11-21T18:38:10.591Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How broadly does budget internalization hold beyond the models and tasks reported in the paper?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","summary":"The authors argue that a fixed rollout cap in RL with verifiable rewards (RLVR) can shape how a model uses inference compute after training, not just reduce training cost. They call this effect budget internalization: models trained with a cap use few additional tokens beyond that budget, perform better than base models near it, but stop improving and fall below the base models at larger budgets.\n\nThe reported effect appears across three models and two reasoning training tasks, and transfers from math training to science and coding tasks. The authors’ ablations suggest it is linked to length-censored post-training rather than an RL-specific objective: SFT on artificially truncated traces reproduces the effect, while SFT on naturally short traces does not. Some changes preserve more test-time scaling, but stronger KL regularization reduces in-budget accuracy; forcing answers on truncated rollouts largely removes budget internalization but also most of the in-budget gain.\n\nThe paper describes a tradeoff: a training cap can help when inference stays near that budget, while limiting gains from additional inference compute. The abstract’s evidence comes from the models, tasks, and training comparisons it reports; how broadly the effect applies beyond them remains open.","question":"How broadly does budget internalization hold beyond the models and tasks reported in the paper?","claims":[{"text":"The authors report that models trained with a fixed rollout cap use few tokens beyond the training budget, gain accuracy near it, and perform worse than base models at larger budgets.","excerpt":"After training, models use few additional tokens beyond the training budget. They outperform the base models near that budget, but their accuracy stops improving and falls below the base model at larger budgets.","support":"source-stated"},{"text":"The authors report the effect across three models and two reasoning training tasks, with transfer from math training to science and coding tasks.","excerpt":"We observe it across Qwen3.5-4B, Gemma4-E2B, and Nemotron-3-Nano-4B and two reasoning training tasks: mathematical reasoning and diverse, procedurally generated problems. This effect also transfers from math training to science and coding tasks.","support":"source-stated"},{"text":"The authors link budget internalization to length-censored post-training rather than an RL-specific objective term, based in part on results from SFT on truncated traces.","excerpt":"SFT on artificially truncated traces reproduces budget internalization, whereas SFT on naturally short traces does not, which shows that the effect is not specific to RL. This result links budget internalization to length-censored post-training rather than to an RL-specific objective term.","support":"source-stated"},{"text":"The authors report that interventions preserving more test-time scaling can reduce in-budget accuracy or remove much of the in-budget gain.","excerpt":"Stronger KL regularization and lower training-data truncation preserve more scaling, but stronger KL reduces in-budget accuracy. Forcing an answer on truncated training rollouts largely removes budget internalization, but also removes most of the in-budget gain.","support":"source-stated"}],"limitations":["The abstract reports results for three named models, two reasoning training tasks, and specified training-objective ablations; it does not establish how broadly the effect applies beyond those settings."],"mediaId":"b4816995b506283a456b8351cd90ab87ea065b46aa4fcf4fca1114180b772f04","correctionReason":null,"id":"0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6","publicationId":"f12b9d11-4614-44a1-8412-28f298fd530f","version":1,"sourceUrl":"https://arxiv.org/abs/2611.26172","sourceTitle":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-19","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-20T05:07:28.417Z","createdAt":"2026-11-20T08:31:07.411Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b4816995b506283a456b8351cd90ab87ea065b46aa4fcf4fca1114180b772f04","kind":"article","url":"https://arxiv.org/abs/2611.26172","sourceUrl":"https://arxiv.org/abs/2611.26172","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-20T08:18:48.529Z","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/2611.26172","checkedAt":"2026-11-21T18:38:10.591Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"b4816995b506283a456b8351cd90ab87ea065b46aa4fcf4fca1114180b772f04","confidence":0.61,"provider":"typesafe","model":"jev-1.13.0","inputHash":"0ce35ad4ba1f0d7d62c7cdb99f5aa161dffe91ba7530fbee82edce916b8e0b5a","decidedAt":"2026-11-20T08:18:49.709Z"},"provenance":{"textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observationId":"27538b910eb3fd878f8030ad952ceaded709fc0a3b428e42f937ac67b5e0290a","retrievedAt":"2026-11-20T08:18:48.325Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"5d381e0c-75c3-4a6a-86bb-5c0cc10066c4"},"reviewer":{"model":"gpt-6-luna","runId":"e5745dfb-2f48-4626-8f59-9f13a07d7b84"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-20T08:31:07.411Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Zavidnyi, Ilia","url":null,"kind":"unknown"},{"name":"Kovrigin, Alexander","url":null,"kind":"unknown"},{"name":"Puchkova, Arina","url":null,"kind":"unknown"},{"name":"Gelvan, Kirill","url":null,"kind":"unknown"},{"name":"Titov, Sergey","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Zavidnyi, Ilia","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zavidnyi, Ilia","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Kovrigin, Alexander","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kovrigin, Alexander","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Puchkova, Arina","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Puchkova, Arina","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Gelvan, Kirill","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gelvan, Kirill","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Titov, Sergey","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Titov, Sergey","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.26172","snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6/source"}},"canonicalUrl":"https://escapemeridian.com/readings/budget-internalization-fixed-budget-rlvr-trades-test-time-0f0218c0","formats":{"html":"/readings/budget-internalization-fixed-budget-rlvr-trades-test-time-0f0218c0","markdown":"/readings/budget-internalization-fixed-budget-rlvr-trades-test-time-0f0218c0.md","json":"/readings/budget-internalization-fixed-budget-rlvr-trades-test-time-0f0218c0.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-inference","name":"Model inference","explanation":"Using a trained model to produce a prediction or response to new input.","referenceUrl":"https://developers.google.com/machine-learning/glossary#inference"}],"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":"0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6","revision":"f12b9d11-4614-44a1-8412-28f298fd530f","revisionKind":"publication-id","title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","url":"https://escapemeridian.com/readings/budget-internalization-fixed-budget-rlvr-trades-test-time-0f0218c0"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.26172","revision":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","revisionKind":"sha256-source-text","title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","url":"https://arxiv.org/abs/2611.26172"},"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":"0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6","title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","argument":"The authors argue that a fixed rollout cap in RL with verifiable rewards (RLVR) can shape how a model uses inference compute after training, not just reduce training cost. They call this effect budget internalization: models trained with a cap use few additional tokens beyond that budget, perform better than base models near it, but stop improving and fall below the base models at larger budgets.\n\nThe reported effect appears across three models and two reasoning training tasks, and transfers from math training to science and coding tasks. The authors’ ablations suggest it is linked to length-censored post-training rather than an RL-specific objective: SFT on artificially truncated traces reproduces the effect, while SFT on naturally short traces does not. Some changes preserve more test-time scaling, but stronger KL regularization reduces in-budget accuracy; forcing answers on truncated rollouts largely removes budget internalization but also most of the in-budget gain.\n\nThe paper describes a tradeoff: a training cap can help when inference stays near that budget, while limiting gains from additional inference compute. The abstract’s evidence comes from the models, tasks, and training comparisons it reports; how broadly the effect applies beyond them remains open.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports results for three named models, two reasoning training tasks, and specified training-objective ablations; it does not establish how broadly the effect applies beyond those settings.","speaker":{"name":"Zavidnyi, Ilia, Kovrigin, Alexander, Puchkova, Arina, Gelvan, Kirill, Titov, Sergey","roleAtSource":"Article author"},"source":{"title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.26172","publishedAt":"2026-11-19","dateBasis":"Original submission date in the publisher’s citation record."},"evidence":[{"text":"The authors report that models trained with a fixed rollout cap use few tokens beyond the training budget, gain accuracy near it, and perform worse than base models at larger budgets.","excerpt":"After training, models use few additional tokens beyond the training budget. They outperform the base models near that budget, but their accuracy stops improving and falls below the base model at larger budgets.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"},{"text":"The authors report the effect across three models and two reasoning training tasks, with transfer from math training to science and coding tasks.","excerpt":"We observe it across Qwen3.5-4B, Gemma4-E2B, and Nemotron-3-Nano-4B and two reasoning training tasks: mathematical reasoning and diverse, procedurally generated problems. This effect also transfers from math training to science and coding tasks.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"},{"text":"The authors link budget internalization to length-censored post-training rather than an RL-specific objective term, based in part on results from SFT on truncated traces.","excerpt":"SFT on artificially truncated traces reproduces budget internalization, whereas SFT on naturally short traces does not, which shows that the effect is not specific to RL. This result links budget internalization to length-censored post-training rather than to an RL-specific objective term.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"},{"text":"The authors report that interventions preserving more test-time scaling can reduce in-budget accuracy or remove much of the in-budget gain.","excerpt":"Stronger KL regularization and lower training-data truncation preserve more scaling, but stronger KL reduces in-budget accuracy. Forcing an answer on truncated training rollouts largely removes budget internalization, but also removes most of the in-budget gain.","url":"https://arxiv.org/abs/2611.26172","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.26172","checkedAt":"2026-11-21T18:38:10.591Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How broadly does budget internalization hold beyond the models and tasks reported in the paper?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","summary":"The authors argue that a fixed rollout cap in RL with verifiable rewards (RLVR) can shape how a model uses inference compute after training, not just reduce training cost. They call this effect budget internalization: models trained with a cap use few additional tokens beyond that budget, perform better than base models near it, but stop improving and fall below the base models at larger budgets.\n\nThe reported effect appears across three models and two reasoning training tasks, and transfers from math training to science and coding tasks. The authors’ ablations suggest it is linked to length-censored post-training rather than an RL-specific objective: SFT on artificially truncated traces reproduces the effect, while SFT on naturally short traces does not. Some changes preserve more test-time scaling, but stronger KL regularization reduces in-budget accuracy; forcing answers on truncated rollouts largely removes budget internalization but also most of the in-budget gain.\n\nThe paper describes a tradeoff: a training cap can help when inference stays near that budget, while limiting gains from additional inference compute. The abstract’s evidence comes from the models, tasks, and training comparisons it reports; how broadly the effect applies beyond them remains open.","question":"How broadly does budget internalization hold beyond the models and tasks reported in the paper?","claims":[{"text":"The authors report that models trained with a fixed rollout cap use few tokens beyond the training budget, gain accuracy near it, and perform worse than base models at larger budgets.","excerpt":"After training, models use few additional tokens beyond the training budget. They outperform the base models near that budget, but their accuracy stops improving and falls below the base model at larger budgets.","support":"source-stated"},{"text":"The authors report the effect across three models and two reasoning training tasks, with transfer from math training to science and coding tasks.","excerpt":"We observe it across Qwen3.5-4B, Gemma4-E2B, and Nemotron-3-Nano-4B and two reasoning training tasks: mathematical reasoning and diverse, procedurally generated problems. This effect also transfers from math training to science and coding tasks.","support":"source-stated"},{"text":"The authors link budget internalization to length-censored post-training rather than an RL-specific objective term, based in part on results from SFT on truncated traces.","excerpt":"SFT on artificially truncated traces reproduces budget internalization, whereas SFT on naturally short traces does not, which shows that the effect is not specific to RL. This result links budget internalization to length-censored post-training rather than to an RL-specific objective term.","support":"source-stated"},{"text":"The authors report that interventions preserving more test-time scaling can reduce in-budget accuracy or remove much of the in-budget gain.","excerpt":"Stronger KL regularization and lower training-data truncation preserve more scaling, but stronger KL reduces in-budget accuracy. Forcing an answer on truncated training rollouts largely removes budget internalization, but also removes most of the in-budget gain.","support":"source-stated"}],"limitations":["The abstract reports results for three named models, two reasoning training tasks, and specified training-objective ablations; it does not establish how broadly the effect applies beyond those settings."],"mediaId":"b4816995b506283a456b8351cd90ab87ea065b46aa4fcf4fca1114180b772f04","correctionReason":null,"id":"0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6","publicationId":"f12b9d11-4614-44a1-8412-28f298fd530f","version":1,"sourceUrl":"https://arxiv.org/abs/2611.26172","sourceTitle":"Budget Internalization: Fixed-Budget RLVR Trades Test-Time Scaling for In-Budget Accuracy","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-19","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-20T05:07:28.417Z","createdAt":"2026-11-20T08:31:07.411Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b4816995b506283a456b8351cd90ab87ea065b46aa4fcf4fca1114180b772f04","kind":"article","url":"https://arxiv.org/abs/2611.26172","sourceUrl":"https://arxiv.org/abs/2611.26172","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-20T08:18:48.529Z","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/2611.26172","checkedAt":"2026-11-21T18:38:10.591Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v9","mediaId":"b4816995b506283a456b8351cd90ab87ea065b46aa4fcf4fca1114180b772f04","confidence":0.61,"provider":"typesafe","model":"jev-1.13.0","inputHash":"0ce35ad4ba1f0d7d62c7cdb99f5aa161dffe91ba7530fbee82edce916b8e0b5a","decidedAt":"2026-11-20T08:18:49.709Z"},"provenance":{"textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observationId":"27538b910eb3fd878f8030ad952ceaded709fc0a3b428e42f937ac67b5e0290a","retrievedAt":"2026-11-20T08:18:48.325Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"5d381e0c-75c3-4a6a-86bb-5c0cc10066c4"},"reviewer":{"model":"gpt-6-luna","runId":"e5745dfb-2f48-4626-8f59-9f13a07d7b84"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-20T08:31:07.411Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Zavidnyi, Ilia","url":null,"kind":"unknown"},{"name":"Kovrigin, Alexander","url":null,"kind":"unknown"},{"name":"Puchkova, Arina","url":null,"kind":"unknown"},{"name":"Gelvan, Kirill","url":null,"kind":"unknown"},{"name":"Titov, Sergey","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Zavidnyi, Ilia","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zavidnyi, Ilia","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Kovrigin, Alexander","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kovrigin, Alexander","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Puchkova, Arina","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Puchkova, Arina","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Gelvan, Kirill","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gelvan, Kirill","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}},{"actor":{"name":"Titov, Sergey","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Titov, Sergey","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.26172","locator":"head/meta[name=author|citation_author]"},"assertionId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-21T18:38:10.591Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","sourceUrl":"https://arxiv.org/abs/2611.26172","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.26172","snapshotId":"71bbba16dc4d7c55b4e657047c97eaff90428f9383a60ecf806f51ddbd032a49","contentSha256":"dc654dc0c9dc27976c19551fc21af10be2a90e5720e0060d13bffa4e7d5eab0f","textHash":"3c1640a3ce6bdf819318a4019ed8cf336208971f59b38a2a1d3cca54ee365179","observedAt":"2026-11-21T18:38:10.591Z","basis":"publisher-page-metadata","sourceLinks":45,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/0f0218c029c712329017c165bc5aaaddfe3a4527ad012d8a9bcafba886f7b8c6/source"}}}],"audits":[{"publicationId":"f12b9d11-4614-44a1-8412-28f298fd530f","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft consistently attributes the findings to the paper's authors. It uses 'the authors argue/report' and presents the results as their claims, not as established fact.","excerpt":"Authors:Ilia Zavidnyi, Alexander Kovrigin, Arina Puchkova, Kirill Gelvan, Sergey Titov"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful, but the central terms stay unexplained for a non-specialist: RLVR, rollout cap, test-time scaling, tokens, SFT, KL regularization, ablations and length-censored. The source itself gives a plain definition that the draft omits: reasoning models get more accurate when allowed longer answers. Smallest repair: open with that definition, say a rollout cap limits answer length during training, and gloss SFT and KL briefly or via tooltips. The draft also drops that most objective changes recovered little scaling.","excerpt":"Reasoning models often become more accurate when they can generate longer answers at test time. This behavior is known as test-time scaling."},{"dimension":"date","status":"pass","explanation":"The originalPublished date of 2026-11-19 matches the v1 submission date on the arXiv page. The feed date of Nov 20 is a later announcement.","excerpt":"[Submitted on 19 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract. The tradeoff framing and the open question about generality follow from the stated scope without inventing benefits, costs or forecasts. Saying 'some changes preserve more test-time scaling' is accurate. Omitting lower training-data truncation's in-budget cost is fine, because the source states no such cost.","excerpt":"It can trade low-budget gains for lower performance when additional inference compute is available."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-20T08:31:07.411Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}