{"id":"82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da","title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","argument":"Boris Ruf, Adrien de Mailly Nesle and Marcin Detyniecki compare estimated energy and carbon costs for two ways agentic language model systems can improve task performance: refining their own answers and using tools while coding. In their MATH benchmark study, three rounds of self-refinement brought GPT-4.1-mini’s accuracy to within one percentage point of GPT-4.1’s single-pass result, but nearly tripled carbon emissions against the cheaper single-pass baseline. Under the study’s token-based carbon scenario, however, the refined system remained below GPT-4.1’s single-pass emissions. In a separate SWE-bench Lite study, GPT-5.3-codex resolved the largest share of tasks in the tested sample, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task. The comparison therefore depends on what is being optimized: task success or estimated resource use per successful task. The authors caution that their carbon estimates count output-token inference only. They exclude prompt tokens and tool execution, making the reported figures lower bounds rather than full estimates of agentic AI’s footprint.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The carbon estimates exclude prompt tokens and tool execution, so they do not represent the full agentic footprint. The reported task-resolution and efficiency comparisons are from the study’s specified MATH and SWE-bench Lite evaluations; the evidence does not establish how the results generalize beyond those settings.","speaker":{"name":"Ruf, Boris, Nesle, Adrien de Mailly, Detyniecki, Marcin","roleAtSource":"Article author"},"source":{"title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.12128","publishedAt":"2026-11-09","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors present two empirical studies of energy and carbon costs: one on iterative verbal self-refinement and one on tool-augmented coding.","excerpt":"We present two complementary empirical studies that estimate the energy and carbon footprint of agentic AI across settings of increasing complexity.","url":"https://arxiv.org/abs/2611.12128","relationship":"Source-stated"},{"text":"In the SWE-bench Lite sample, GPT-5.3-codex had the highest resolution rate, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task.","excerpt":"GPT-5.3-codex achieves the highest resolution rate in this sample (81.2\\%), ahead of Claude Opus~4.6 (68.9\\%) and GPT-4.1-mini (25.0\\%), but resolved-task-normalized efficiency remains mixed: GPT-4.1-mini is lowest on estimated carbon and cost per resolved task","url":"https://arxiv.org/abs/2611.12128","relationship":"Source-stated"},{"text":"The carbon estimates cover output-token inference only; prompt tokens and tool execution are excluded, so the estimates are lower bounds on the full footprint.","excerpt":"All carbon estimates cover output-token inference only; prompt tokens, which exceed 99\\% of tokens in Study~2, and tool execution are not charged, so the figures are lower bounds on the full agentic footprint.","url":"https://arxiv.org/abs/2611.12128","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.12128","checkedAt":"2026-11-10T19:32:11.668Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How would including prompt tokens and tool execution change the studies’ carbon estimates and comparisons between systems?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","summary":"Boris Ruf, Adrien de Mailly Nesle and Marcin Detyniecki compare estimated energy and carbon costs for two ways agentic language model systems can improve task performance: refining their own answers and using tools while coding. In their MATH benchmark study, three rounds of self-refinement brought GPT-4.1-mini’s accuracy to within one percentage point of GPT-4.1’s single-pass result, but nearly tripled carbon emissions against the cheaper single-pass baseline. Under the study’s token-based carbon scenario, however, the refined system remained below GPT-4.1’s single-pass emissions. In a separate SWE-bench Lite study, GPT-5.3-codex resolved the largest share of tasks in the tested sample, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task. The comparison therefore depends on what is being optimized: task success or estimated resource use per successful task. The authors caution that their carbon estimates count output-token inference only. They exclude prompt tokens and tool execution, making the reported figures lower bounds rather than full estimates of agentic AI’s footprint.","question":"How would including prompt tokens and tool execution change the studies’ carbon estimates and comparisons between systems?","claims":[{"text":"The authors present two empirical studies of energy and carbon costs: one on iterative verbal self-refinement and one on tool-augmented coding.","excerpt":"We present two complementary empirical studies that estimate the energy and carbon footprint of agentic AI across settings of increasing complexity.","support":"source-stated"},{"text":"In the SWE-bench Lite sample, GPT-5.3-codex had the highest resolution rate, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task.","excerpt":"GPT-5.3-codex achieves the highest resolution rate in this sample (81.2\\%), ahead of Claude Opus~4.6 (68.9\\%) and GPT-4.1-mini (25.0\\%), but resolved-task-normalized efficiency remains mixed: GPT-4.1-mini is lowest on estimated carbon and cost per resolved task","support":"source-stated"},{"text":"The carbon estimates cover output-token inference only; prompt tokens and tool execution are excluded, so the estimates are lower bounds on the full footprint.","excerpt":"All carbon estimates cover output-token inference only; prompt tokens, which exceed 99\\% of tokens in Study~2, and tool execution are not charged, so the figures are lower bounds on the full agentic footprint.","support":"source-stated"}],"limitations":["The carbon estimates exclude prompt tokens and tool execution, so they do not represent the full agentic footprint.","The reported task-resolution and efficiency comparisons are from the study’s specified MATH and SWE-bench Lite evaluations; the evidence does not establish how the results generalize beyond those settings."],"mediaId":"b4cb1e10aa72b8a36104026a66b0cdea534f4f535afee7c1bce162365d71b978","correctionReason":null,"id":"82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da","publicationId":"f7cd8571-3443-40a2-8fd8-f2eadd0c68d1","version":1,"sourceUrl":"https://arxiv.org/abs/2611.12128","sourceTitle":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-09","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-10T08:34:15.032Z","createdAt":"2026-11-10T19:01:12.034Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b4cb1e10aa72b8a36104026a66b0cdea534f4f535afee7c1bce162365d71b978","kind":"article","url":"https://arxiv.org/abs/2611.12128","sourceUrl":"https://arxiv.org/abs/2611.12128","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-10T14:19:13.578Z","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/2611.12128","checkedAt":"2026-11-10T19:32:11.668Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"b4cb1e10aa72b8a36104026a66b0cdea534f4f535afee7c1bce162365d71b978","confidence":0.9,"provider":"typesafe","model":"jev-1.13.0","inputHash":"c3a30675a9d251826b3b2930606a45e2c0110d8a372394dcde8bfc0de32ac5f5","decidedAt":"2026-11-10T14:19:15.596Z"},"provenance":{"textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observationId":"126c593ae7d981deb1e83dd92407eca1f7a7841f1ed4946f8d673527051afbc6","retrievedAt":"2026-11-10T14:19:13.303Z","kind":"publisher-page","preparer":{"model":"server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna)","runId":"4e24819b-cc86-431e-80a3-2ba54609d78e"},"reviewer":{"model":"gpt-6-luna","runId":"60cae81d-9dab-4ee4-8140-7060558906e3"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-10T19:01:12.034Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Ruf, Boris","url":null,"kind":"unknown"},{"name":"Nesle, Adrien de Mailly","url":null,"kind":"unknown"},{"name":"Detyniecki, Marcin","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Ruf, Boris","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Ruf, Boris","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.12128","locator":"head/meta[name=author|citation_author]"},"assertionId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T19:32:11.668Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","sourceUrl":"https://arxiv.org/abs/2611.12128","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z"}},{"actor":{"name":"Nesle, Adrien de Mailly","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Nesle, Adrien de Mailly","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.12128","locator":"head/meta[name=author|citation_author]"},"assertionId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T19:32:11.668Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","sourceUrl":"https://arxiv.org/abs/2611.12128","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z"}},{"actor":{"name":"Detyniecki, Marcin","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Detyniecki, Marcin","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.12128","locator":"head/meta[name=author|citation_author]"},"assertionId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T19:32:11.668Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","sourceUrl":"https://arxiv.org/abs/2611.12128","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z"}}],"participationCoverage":{"total":3,"returned":3,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.12128","snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z","basis":"publisher-page-metadata","sourceLinks":41,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da/source"}},"canonicalUrl":"https://escapemeridian.com/readings/toward-sustainable-agentic-ai-carbon-tradeoffs-across-iterative-82944e3b","formats":{"html":"/readings/toward-sustainable-agentic-ai-carbon-tradeoffs-across-iterative-82944e3b","markdown":"/readings/toward-sustainable-agentic-ai-carbon-tradeoffs-across-iterative-82944e3b.md","json":"/readings/toward-sustainable-agentic-ai-carbon-tradeoffs-across-iterative-82944e3b.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-agentic","name":"Agentic AI","explanation":"AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.","referenceUrl":"https://developers.google.com/machine-learning/glossary#agentic"},{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da","revision":"f7cd8571-3443-40a2-8fd8-f2eadd0c68d1","revisionKind":"publication-id","title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","url":"https://escapemeridian.com/readings/toward-sustainable-agentic-ai-carbon-tradeoffs-across-iterative-82944e3b"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.12128","revision":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","revisionKind":"sha256-source-text","title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","url":"https://arxiv.org/abs/2611.12128"},"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":"82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da","title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","argument":"Boris Ruf, Adrien de Mailly Nesle and Marcin Detyniecki compare estimated energy and carbon costs for two ways agentic language model systems can improve task performance: refining their own answers and using tools while coding. In their MATH benchmark study, three rounds of self-refinement brought GPT-4.1-mini’s accuracy to within one percentage point of GPT-4.1’s single-pass result, but nearly tripled carbon emissions against the cheaper single-pass baseline. Under the study’s token-based carbon scenario, however, the refined system remained below GPT-4.1’s single-pass emissions. In a separate SWE-bench Lite study, GPT-5.3-codex resolved the largest share of tasks in the tested sample, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task. The comparison therefore depends on what is being optimized: task success or estimated resource use per successful task. The authors caution that their carbon estimates count output-token inference only. They exclude prompt tokens and tool execution, making the reported figures lower bounds rather than full estimates of agentic AI’s footprint.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The carbon estimates exclude prompt tokens and tool execution, so they do not represent the full agentic footprint. The reported task-resolution and efficiency comparisons are from the study’s specified MATH and SWE-bench Lite evaluations; the evidence does not establish how the results generalize beyond those settings.","speaker":{"name":"Ruf, Boris, Nesle, Adrien de Mailly, Detyniecki, Marcin","roleAtSource":"Article author"},"source":{"title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.12128","publishedAt":"2026-11-09","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors present two empirical studies of energy and carbon costs: one on iterative verbal self-refinement and one on tool-augmented coding.","excerpt":"We present two complementary empirical studies that estimate the energy and carbon footprint of agentic AI across settings of increasing complexity.","url":"https://arxiv.org/abs/2611.12128","relationship":"Source-stated"},{"text":"In the SWE-bench Lite sample, GPT-5.3-codex had the highest resolution rate, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task.","excerpt":"GPT-5.3-codex achieves the highest resolution rate in this sample (81.2\\%), ahead of Claude Opus~4.6 (68.9\\%) and GPT-4.1-mini (25.0\\%), but resolved-task-normalized efficiency remains mixed: GPT-4.1-mini is lowest on estimated carbon and cost per resolved task","url":"https://arxiv.org/abs/2611.12128","relationship":"Source-stated"},{"text":"The carbon estimates cover output-token inference only; prompt tokens and tool execution are excluded, so the estimates are lower bounds on the full footprint.","excerpt":"All carbon estimates cover output-token inference only; prompt tokens, which exceed 99\\% of tokens in Study~2, and tool execution are not charged, so the figures are lower bounds on the full agentic footprint.","url":"https://arxiv.org/abs/2611.12128","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.12128","checkedAt":"2026-11-10T19:32:11.668Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How would including prompt tokens and tool execution change the studies’ carbon estimates and comparisons between systems?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","summary":"Boris Ruf, Adrien de Mailly Nesle and Marcin Detyniecki compare estimated energy and carbon costs for two ways agentic language model systems can improve task performance: refining their own answers and using tools while coding. In their MATH benchmark study, three rounds of self-refinement brought GPT-4.1-mini’s accuracy to within one percentage point of GPT-4.1’s single-pass result, but nearly tripled carbon emissions against the cheaper single-pass baseline. Under the study’s token-based carbon scenario, however, the refined system remained below GPT-4.1’s single-pass emissions. In a separate SWE-bench Lite study, GPT-5.3-codex resolved the largest share of tasks in the tested sample, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task. The comparison therefore depends on what is being optimized: task success or estimated resource use per successful task. The authors caution that their carbon estimates count output-token inference only. They exclude prompt tokens and tool execution, making the reported figures lower bounds rather than full estimates of agentic AI’s footprint.","question":"How would including prompt tokens and tool execution change the studies’ carbon estimates and comparisons between systems?","claims":[{"text":"The authors present two empirical studies of energy and carbon costs: one on iterative verbal self-refinement and one on tool-augmented coding.","excerpt":"We present two complementary empirical studies that estimate the energy and carbon footprint of agentic AI across settings of increasing complexity.","support":"source-stated"},{"text":"In the SWE-bench Lite sample, GPT-5.3-codex had the highest resolution rate, while GPT-4.1-mini had the lowest estimated carbon and cost per resolved task.","excerpt":"GPT-5.3-codex achieves the highest resolution rate in this sample (81.2\\%), ahead of Claude Opus~4.6 (68.9\\%) and GPT-4.1-mini (25.0\\%), but resolved-task-normalized efficiency remains mixed: GPT-4.1-mini is lowest on estimated carbon and cost per resolved task","support":"source-stated"},{"text":"The carbon estimates cover output-token inference only; prompt tokens and tool execution are excluded, so the estimates are lower bounds on the full footprint.","excerpt":"All carbon estimates cover output-token inference only; prompt tokens, which exceed 99\\% of tokens in Study~2, and tool execution are not charged, so the figures are lower bounds on the full agentic footprint.","support":"source-stated"}],"limitations":["The carbon estimates exclude prompt tokens and tool execution, so they do not represent the full agentic footprint.","The reported task-resolution and efficiency comparisons are from the study’s specified MATH and SWE-bench Lite evaluations; the evidence does not establish how the results generalize beyond those settings."],"mediaId":"b4cb1e10aa72b8a36104026a66b0cdea534f4f535afee7c1bce162365d71b978","correctionReason":null,"id":"82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da","publicationId":"f7cd8571-3443-40a2-8fd8-f2eadd0c68d1","version":1,"sourceUrl":"https://arxiv.org/abs/2611.12128","sourceTitle":"Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-09","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-10T08:34:15.032Z","createdAt":"2026-11-10T19:01:12.034Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"b4cb1e10aa72b8a36104026a66b0cdea534f4f535afee7c1bce162365d71b978","kind":"article","url":"https://arxiv.org/abs/2611.12128","sourceUrl":"https://arxiv.org/abs/2611.12128","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-10T14:19:13.578Z","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/2611.12128","checkedAt":"2026-11-10T19:32:11.668Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"b4cb1e10aa72b8a36104026a66b0cdea534f4f535afee7c1bce162365d71b978","confidence":0.9,"provider":"typesafe","model":"jev-1.13.0","inputHash":"c3a30675a9d251826b3b2930606a45e2c0110d8a372394dcde8bfc0de32ac5f5","decidedAt":"2026-11-10T14:19:15.596Z"},"provenance":{"textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observationId":"126c593ae7d981deb1e83dd92407eca1f7a7841f1ed4946f8d673527051afbc6","retrievedAt":"2026-11-10T14:19:13.303Z","kind":"publisher-page","preparer":{"model":"server-excerpt-repair/stream-excerpt-repair-v1 (original draft: gpt-6-luna)","runId":"4e24819b-cc86-431e-80a3-2ba54609d78e"},"reviewer":{"model":"gpt-6-luna","runId":"60cae81d-9dab-4ee4-8140-7060558906e3"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-10T19:01:12.034Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Ruf, Boris","url":null,"kind":"unknown"},{"name":"Nesle, Adrien de Mailly","url":null,"kind":"unknown"},{"name":"Detyniecki, Marcin","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Ruf, Boris","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Ruf, Boris","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.12128","locator":"head/meta[name=author|citation_author]"},"assertionId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T19:32:11.668Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","sourceUrl":"https://arxiv.org/abs/2611.12128","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z"}},{"actor":{"name":"Nesle, Adrien de Mailly","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Nesle, Adrien de Mailly","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.12128","locator":"head/meta[name=author|citation_author]"},"assertionId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T19:32:11.668Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","sourceUrl":"https://arxiv.org/abs/2611.12128","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z"}},{"actor":{"name":"Detyniecki, Marcin","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Detyniecki, Marcin","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.12128","locator":"head/meta[name=author|citation_author]"},"assertionId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T19:32:11.668Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","sourceUrl":"https://arxiv.org/abs/2611.12128","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z"}}],"participationCoverage":{"total":3,"returned":3,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.12128","snapshotId":"7999cb840903be9dd7004cd985a9c4c1c03aee7c21cd8ff5105f142bb7db6f52","contentSha256":"9550a0e2dddbc3a0669919cb6006311fc0ebed1c577a89f2b846dca3a9185c5b","textHash":"0fcd11aa45df200d2d5bd6edcbca825029ae4ff2a92b63ad77e91eba76fa0441","observedAt":"2026-11-10T19:32:11.668Z","basis":"publisher-page-metadata","sourceLinks":41,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/82944e3ba7bbb246a93a285d9c6ad096ec66cffabc78a5be6d7b6fa6abc6b7da/source"}}}],"audits":[{"publicationId":"f7cd8571-3443-40a2-8fd8-f2eadd0c68d1","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The draft correctly attributes the paper to Boris Ruf, Adrien de Mailly Nesle and Marcin Detyniecki. It presents the findings as the authors' estimates rather than as established fact.","excerpt":"Authors:Boris Ruf, Adrien de Mailly Nesle, Marcin Detyniecki"},{"dimension":"summary","status":"uncertain","explanation":"The draft is faithful in substance but has two gaps. First, it says the estimates are lower bounds without noting that excluded prompt tokens exceed 99% of tokens in Study 2. That figure is what gives the caveat its force. Smallest repair: add it. Second, central terms go unexplained for a general reader: tokens, output-token inference, prompt tokens, agentic, MATH and SWE-bench Lite. Short plain-language glosses or term tooltips would fix this. A reader without that background may not grasp what is excluded or how large the omission could be.","excerpt":"prompt tokens, which exceed 99\\% of tokens in Study~2, and tool execution are not charged"},{"dimension":"date","status":"pass","explanation":"The displayed original publication date, 2026-11-09, matches the v1 submission date. The feed date of 2026-11-10 is later and is correctly not used.","excerpt":"[Submitted on 9 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"The claims match the abstract: the k=3 refinement result, the carbon comparison against both baselines, the SWE-bench Lite results and the lower-bound caveat. The framing that the outcome depends on what is optimized follows from the source calling efficiency mixed. The question follows directly from the stated exclusions. No benefits or costs are invented.","excerpt":"In Study~1, $k=3$ refinements close a 16-percentage-point accuracy gap to within 1~point, but nearly triple the carbon footprint relative to the cheap single-pass baseline"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page for this paper. It is the direct original source.","excerpt":null}],"createdAt":"2026-11-10T19:01:12.034Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}