{"id":"e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1","title":"arXiv cs.AI: Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","argument":"Karol Struniawski reports that concentrated bursts of attention head masking disrupted autoregressive reasoning more than budget-matched diffuse dropout in tests on GSM8K. The study evaluated 250 paired derivations across Qwen3-8B and LLaMA-3.1-8B-Instruct. At the stated onset parameter, Qwen3-8B accuracy was 78.8% under burst dropout versus 91.2% under diffuse noise; LLaMA-3.1-8B-Instruct scored 76.8% versus 84.0%. When the same cumulative budget was dispersed uniformly across generation steps, the paper reports substantially milder degradation, with both models retaining their accuracy under diffuse noise. The contrast suggests that where disruption falls in a reasoning sequence, not only the total dropout budget, may matter in these tests. An exploratory analysis associated single-step dropout density with an acute drop in one schedule draw, but found the multi-token continuity compounding trend inconclusive. Error patterns also differed: Qwen failures reflected derivation errors within formatted responses, while some LLaMA losses lacked a final answer format line. Struniawski presents the findings as empirical benchmarks for fault tolerance in transformer architectures. The evaluation covers one benchmark and two models, so whether the results extend beyond these tested conditions remains unresolved.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The reported evaluation covers 250 paired derivations on GSM8K and two 8B models; the abstract does not report results on other benchmarks or models. The paper describes the multi-token continuity result as an inconclusive compounding trend. The abstract reports an exploratory factorial dissociation with N=100; it does not establish how the reported effects generalize beyond the tested conditions.","speaker":{"name":"Struniawski, Karol","roleAtSource":"Article author"},"source":{"title":"Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.14209","publishedAt":"2026-11-10","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper compares clustered bursts of attention head masking with budget-matched diffuse dropout in 250 paired derivations on GSM8K using two 8B models.","excerpt":"Evaluating 250 paired derivations on the GSM8K benchmark across two 8B models","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"At the stated onset parameter, burst dropout lowered Qwen3-8B accuracy more than diffuse noise.","excerpt":"Qwen3-8B accuracy drops from $91.2\\%$ under diffuse noise to $78.8\\%$ under burst dropout","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"At the stated onset parameter, burst dropout lowered LLaMA-3.1-8B-Instruct accuracy more than diffuse noise.","excerpt":"LLaMA-3.1-8B-Instruct drops from $84.0\\%$ to $76.8\\%$","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"Dispersing the same cumulative dropout budget uniformly across generation steps was associated with substantially milder degradation.","excerpt":"dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"In an exploratory analysis, the association with an acute drop for single-step dropout density was reported, while the compounding trend for multi-token continuity was inconclusive.","excerpt":"single-step dropout density is associated with an acute drop in a single schedule draw","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"The paper reports different error patterns for the two models: Qwen failures involved derivation errors, while some LLaMA losses lacked a final answer format line.","excerpt":"Qwen failures reflect derivation errors within formatted responses","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.14209","checkedAt":"2026-11-11T21:50:13.637Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Do concentrated bursts produce similar accuracy losses on other benchmarks and models, beyond the two models tested on GSM8K?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","summary":"Karol Struniawski reports that concentrated bursts of attention head masking disrupted autoregressive reasoning more than budget-matched diffuse dropout in tests on GSM8K. The study evaluated 250 paired derivations across Qwen3-8B and LLaMA-3.1-8B-Instruct. At the stated onset parameter, Qwen3-8B accuracy was 78.8% under burst dropout versus 91.2% under diffuse noise; LLaMA-3.1-8B-Instruct scored 76.8% versus 84.0%. When the same cumulative budget was dispersed uniformly across generation steps, the paper reports substantially milder degradation, with both models retaining their accuracy under diffuse noise. The contrast suggests that where disruption falls in a reasoning sequence, not only the total dropout budget, may matter in these tests. An exploratory analysis associated single-step dropout density with an acute drop in one schedule draw, but found the multi-token continuity compounding trend inconclusive. Error patterns also differed: Qwen failures reflected derivation errors within formatted responses, while some LLaMA losses lacked a final answer format line. Struniawski presents the findings as empirical benchmarks for fault tolerance in transformer architectures. The evaluation covers one benchmark and two models, so whether the results extend beyond these tested conditions remains unresolved.","question":"Do concentrated bursts produce similar accuracy losses on other benchmarks and models, beyond the two models tested on GSM8K?","claims":[{"text":"The paper compares clustered bursts of attention head masking with budget-matched diffuse dropout in 250 paired derivations on GSM8K using two 8B models.","excerpt":"Evaluating 250 paired derivations on the GSM8K benchmark across two 8B models","support":"source-stated"},{"text":"At the stated onset parameter, burst dropout lowered Qwen3-8B accuracy more than diffuse noise.","excerpt":"Qwen3-8B accuracy drops from $91.2\\%$ under diffuse noise to $78.8\\%$ under burst dropout","support":"source-stated"},{"text":"At the stated onset parameter, burst dropout lowered LLaMA-3.1-8B-Instruct accuracy more than diffuse noise.","excerpt":"LLaMA-3.1-8B-Instruct drops from $84.0\\%$ to $76.8\\%$","support":"source-stated"},{"text":"Dispersing the same cumulative dropout budget uniformly across generation steps was associated with substantially milder degradation.","excerpt":"dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation","support":"source-stated"},{"text":"In an exploratory analysis, the association with an acute drop for single-step dropout density was reported, while the compounding trend for multi-token continuity was inconclusive.","excerpt":"single-step dropout density is associated with an acute drop in a single schedule draw","support":"source-stated"},{"text":"The paper reports different error patterns for the two models: Qwen failures involved derivation errors, while some LLaMA losses lacked a final answer format line.","excerpt":"Qwen failures reflect derivation errors within formatted responses","support":"source-stated"}],"limitations":["The reported evaluation covers 250 paired derivations on GSM8K and two 8B models; the abstract does not report results on other benchmarks or models.","The paper describes the multi-token continuity result as an inconclusive compounding trend.","The abstract reports an exploratory factorial dissociation with N=100; it does not establish how the reported effects generalize beyond the tested conditions."],"mediaId":"549eeb81d58ba343e51795fd5b62d8164456ac67bd92ec81675b818af989cd48","correctionReason":null,"id":"e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1","publicationId":"0df16650-c366-4d29-844c-8be202b3adb3","version":1,"sourceUrl":"https://arxiv.org/abs/2611.14209","sourceTitle":"Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-10","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-11T05:09:26.088Z","createdAt":"2026-11-11T20:47:24.994Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"549eeb81d58ba343e51795fd5b62d8164456ac67bd92ec81675b818af989cd48","kind":"article","url":"https://arxiv.org/abs/2611.14209","sourceUrl":"https://arxiv.org/abs/2611.14209","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-11T20:46:29.873Z","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/2611.14209","checkedAt":"2026-11-11T21:50:13.637Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"549eeb81d58ba343e51795fd5b62d8164456ac67bd92ec81675b818af989cd48","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"277ba4550cb24efa2a413f9b13847c8c5442e47bd589d250e1176ef99770908e","decidedAt":"2026-11-11T20:46:30.742Z"},"provenance":{"textHash":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","observationId":"caa1e825953deb395e98e8d0f39f317721290eddca62c666a41bd2b99f24d484","retrievedAt":"2026-11-11T20:46:29.788Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"0cf5c905-ba3e-4fe9-8b71-cb06a90fbfb0"},"reviewer":{"model":"gpt-6-luna","runId":"e7101a0b-0d18-4960-8314-e50ec4e0a60e"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T20:47:24.994Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Struniawski, Karol","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Struniawski, Karol","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Struniawski, Karol","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.14209","locator":"head/meta[name=author|citation_author]"},"assertionId":"649337fe31aa6edbe35d22aee8a734ca444dae6764aa2e98df8832e9c9fa4136:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T21:50:13.637Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"649337fe31aa6edbe35d22aee8a734ca444dae6764aa2e98df8832e9c9fa4136","sourceUrl":"https://arxiv.org/abs/2611.14209","contentSha256":"732a83b866799f5241fad3aa06352c16672357e379ff7c05ac40525d7805aa36","textHash":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","observedAt":"2026-11-11T21:50:13.637Z"}}],"participationCoverage":{"total":1,"returned":1,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.14209","snapshotId":"649337fe31aa6edbe35d22aee8a734ca444dae6764aa2e98df8832e9c9fa4136","contentSha256":"732a83b866799f5241fad3aa06352c16672357e379ff7c05ac40525d7805aa36","textHash":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","observedAt":"2026-11-11T21:50:13.637Z","basis":"publisher-page-metadata","sourceLinks":40,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-cs-ai-breakdown-of-autoregressive-multi-step-reasoning-e68d31f4","formats":{"html":"/readings/arxiv-cs-ai-breakdown-of-autoregressive-multi-step-reasoning-e68d31f4","markdown":"/readings/arxiv-cs-ai-breakdown-of-autoregressive-multi-step-reasoning-e68d31f4.md","json":"/readings/arxiv-cs-ai-breakdown-of-autoregressive-multi-step-reasoning-e68d31f4.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-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"},{"id":"term-transformer","name":"Transformer architecture","explanation":"The neural-network design behind most current language models; it weighs how each part of the input relates to the others.","referenceUrl":"https://developers.google.com/machine-learning/glossary#transformer"}],"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":"e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1","revision":"0df16650-c366-4d29-844c-8be202b3adb3","revisionKind":"publication-id","title":"arXiv cs.AI: Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","url":"https://escapemeridian.com/readings/arxiv-cs-ai-breakdown-of-autoregressive-multi-step-reasoning-e68d31f4"},"source":{"type":"source","id":"https://arxiv.org/abs/2611.14209","revision":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","revisionKind":"sha256-source-text","title":"Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","url":"https://arxiv.org/abs/2611.14209"},"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":"e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1","title":"arXiv cs.AI: Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","argument":"Karol Struniawski reports that concentrated bursts of attention head masking disrupted autoregressive reasoning more than budget-matched diffuse dropout in tests on GSM8K. The study evaluated 250 paired derivations across Qwen3-8B and LLaMA-3.1-8B-Instruct. At the stated onset parameter, Qwen3-8B accuracy was 78.8% under burst dropout versus 91.2% under diffuse noise; LLaMA-3.1-8B-Instruct scored 76.8% versus 84.0%. When the same cumulative budget was dispersed uniformly across generation steps, the paper reports substantially milder degradation, with both models retaining their accuracy under diffuse noise. The contrast suggests that where disruption falls in a reasoning sequence, not only the total dropout budget, may matter in these tests. An exploratory analysis associated single-step dropout density with an acute drop in one schedule draw, but found the multi-token continuity compounding trend inconclusive. Error patterns also differed: Qwen failures reflected derivation errors within formatted responses, while some LLaMA losses lacked a final answer format line. Struniawski presents the findings as empirical benchmarks for fault tolerance in transformer architectures. The evaluation covers one benchmark and two models, so whether the results extend beyond these tested conditions remains unresolved.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The reported evaluation covers 250 paired derivations on GSM8K and two 8B models; the abstract does not report results on other benchmarks or models. The paper describes the multi-token continuity result as an inconclusive compounding trend. The abstract reports an exploratory factorial dissociation with N=100; it does not establish how the reported effects generalize beyond the tested conditions.","speaker":{"name":"Struniawski, Karol","roleAtSource":"Article author"},"source":{"title":"Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2611.14209","publishedAt":"2026-11-10","dateBasis":"Date basis not provided."},"evidence":[{"text":"The paper compares clustered bursts of attention head masking with budget-matched diffuse dropout in 250 paired derivations on GSM8K using two 8B models.","excerpt":"Evaluating 250 paired derivations on the GSM8K benchmark across two 8B models","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"At the stated onset parameter, burst dropout lowered Qwen3-8B accuracy more than diffuse noise.","excerpt":"Qwen3-8B accuracy drops from $91.2\\%$ under diffuse noise to $78.8\\%$ under burst dropout","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"At the stated onset parameter, burst dropout lowered LLaMA-3.1-8B-Instruct accuracy more than diffuse noise.","excerpt":"LLaMA-3.1-8B-Instruct drops from $84.0\\%$ to $76.8\\%$","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"Dispersing the same cumulative dropout budget uniformly across generation steps was associated with substantially milder degradation.","excerpt":"dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"In an exploratory analysis, the association with an acute drop for single-step dropout density was reported, while the compounding trend for multi-token continuity was inconclusive.","excerpt":"single-step dropout density is associated with an acute drop in a single schedule draw","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"},{"text":"The paper reports different error patterns for the two models: Qwen failures involved derivation errors, while some LLaMA losses lacked a final answer format line.","excerpt":"Qwen failures reflect derivation errors within formatted responses","url":"https://arxiv.org/abs/2611.14209","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2611.14209","checkedAt":"2026-11-11T21:50:13.637Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Do concentrated bursts produce similar accuracy losses on other benchmarks and models, beyond the two models tested on GSM8K?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","summary":"Karol Struniawski reports that concentrated bursts of attention head masking disrupted autoregressive reasoning more than budget-matched diffuse dropout in tests on GSM8K. The study evaluated 250 paired derivations across Qwen3-8B and LLaMA-3.1-8B-Instruct. At the stated onset parameter, Qwen3-8B accuracy was 78.8% under burst dropout versus 91.2% under diffuse noise; LLaMA-3.1-8B-Instruct scored 76.8% versus 84.0%. When the same cumulative budget was dispersed uniformly across generation steps, the paper reports substantially milder degradation, with both models retaining their accuracy under diffuse noise. The contrast suggests that where disruption falls in a reasoning sequence, not only the total dropout budget, may matter in these tests. An exploratory analysis associated single-step dropout density with an acute drop in one schedule draw, but found the multi-token continuity compounding trend inconclusive. Error patterns also differed: Qwen failures reflected derivation errors within formatted responses, while some LLaMA losses lacked a final answer format line. Struniawski presents the findings as empirical benchmarks for fault tolerance in transformer architectures. The evaluation covers one benchmark and two models, so whether the results extend beyond these tested conditions remains unresolved.","question":"Do concentrated bursts produce similar accuracy losses on other benchmarks and models, beyond the two models tested on GSM8K?","claims":[{"text":"The paper compares clustered bursts of attention head masking with budget-matched diffuse dropout in 250 paired derivations on GSM8K using two 8B models.","excerpt":"Evaluating 250 paired derivations on the GSM8K benchmark across two 8B models","support":"source-stated"},{"text":"At the stated onset parameter, burst dropout lowered Qwen3-8B accuracy more than diffuse noise.","excerpt":"Qwen3-8B accuracy drops from $91.2\\%$ under diffuse noise to $78.8\\%$ under burst dropout","support":"source-stated"},{"text":"At the stated onset parameter, burst dropout lowered LLaMA-3.1-8B-Instruct accuracy more than diffuse noise.","excerpt":"LLaMA-3.1-8B-Instruct drops from $84.0\\%$ to $76.8\\%$","support":"source-stated"},{"text":"Dispersing the same cumulative dropout budget uniformly across generation steps was associated with substantially milder degradation.","excerpt":"dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation","support":"source-stated"},{"text":"In an exploratory analysis, the association with an acute drop for single-step dropout density was reported, while the compounding trend for multi-token continuity was inconclusive.","excerpt":"single-step dropout density is associated with an acute drop in a single schedule draw","support":"source-stated"},{"text":"The paper reports different error patterns for the two models: Qwen failures involved derivation errors, while some LLaMA losses lacked a final answer format line.","excerpt":"Qwen failures reflect derivation errors within formatted responses","support":"source-stated"}],"limitations":["The reported evaluation covers 250 paired derivations on GSM8K and two 8B models; the abstract does not report results on other benchmarks or models.","The paper describes the multi-token continuity result as an inconclusive compounding trend.","The abstract reports an exploratory factorial dissociation with N=100; it does not establish how the reported effects generalize beyond the tested conditions."],"mediaId":"549eeb81d58ba343e51795fd5b62d8164456ac67bd92ec81675b818af989cd48","correctionReason":null,"id":"e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1","publicationId":"0df16650-c366-4d29-844c-8be202b3adb3","version":1,"sourceUrl":"https://arxiv.org/abs/2611.14209","sourceTitle":"Breakdown of Autoregressive Multi-Step Reasoning Under Concentrated vs. Diffuse Attention Head Dropout","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-11-10","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-11T05:09:26.088Z","createdAt":"2026-11-11T20:47:24.994Z","historical":false,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"549eeb81d58ba343e51795fd5b62d8164456ac67bd92ec81675b818af989cd48","kind":"article","url":"https://arxiv.org/abs/2611.14209","sourceUrl":"https://arxiv.org/abs/2611.14209","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-11T20:46:29.873Z","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/2611.14209","checkedAt":"2026-11-11T21:50:13.637Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"549eeb81d58ba343e51795fd5b62d8164456ac67bd92ec81675b818af989cd48","confidence":0.86,"provider":"typesafe","model":"jev-1.13.0","inputHash":"277ba4550cb24efa2a413f9b13847c8c5442e47bd589d250e1176ef99770908e","decidedAt":"2026-11-11T20:46:30.742Z"},"provenance":{"textHash":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","observationId":"caa1e825953deb395e98e8d0f39f317721290eddca62c666a41bd2b99f24d484","retrievedAt":"2026-11-11T20:46:29.788Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"0cf5c905-ba3e-4fe9-8b71-cb06a90fbfb0"},"reviewer":{"model":"gpt-6-luna","runId":"e7101a0b-0d18-4960-8314-e50ec4e0a60e"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-11T20:47:24.994Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Struniawski, Karol","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Struniawski, Karol","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Struniawski, Karol","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2611.14209","locator":"head/meta[name=author|citation_author]"},"assertionId":"649337fe31aa6edbe35d22aee8a734ca444dae6764aa2e98df8832e9c9fa4136:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-11T21:50:13.637Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"649337fe31aa6edbe35d22aee8a734ca444dae6764aa2e98df8832e9c9fa4136","sourceUrl":"https://arxiv.org/abs/2611.14209","contentSha256":"732a83b866799f5241fad3aa06352c16672357e379ff7c05ac40525d7805aa36","textHash":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","observedAt":"2026-11-11T21:50:13.637Z"}}],"participationCoverage":{"total":1,"returned":1,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2611.14209","snapshotId":"649337fe31aa6edbe35d22aee8a734ca444dae6764aa2e98df8832e9c9fa4136","contentSha256":"732a83b866799f5241fad3aa06352c16672357e379ff7c05ac40525d7805aa36","textHash":"68ef8abbd575d501061960b4b74b628643dc9235676b24d8ca7437bdd207eb7b","observedAt":"2026-11-11T21:50:13.637Z","basis":"publisher-page-metadata","sourceLinks":40,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/e68d31f45b1ce7765124860f3f666c58c1fd9e1e7fd44940d70d19f922b5d2c1/source"}}}],"audits":[{"publicationId":"0df16650-c366-4d29-844c-8be202b3adb3","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly attributes the paper to Karol Struniawski on arXiv cs.AI. It keeps the results as the paper's reported findings rather than established facts.","excerpt":"Authors:Karol Struniawski"},{"dimension":"summary","status":"uncertain","explanation":"The numbers are faithful, but a non-specialist cannot follow the central idea. Key terms are never explained in plain language: attention head masking/dropout, autoregressive, GSM8K, onset parameter, budget-matched, schedule draw, multi-token continuity. Readers cannot tell what was disrupted or what GSM8K measures. The reading also drops the paper's motivating analogy to brief neuronal silences in tired brains (off-states). Smallest repair: one plain-language line saying the test temporarily switches off parts of the model's attention while it generates text, GSM8K is a set of grade-school math word problems, and the paper frames the work as an analogue of brain off-states.","excerpt":"During extended wakefulness, mammalian neocortical networks experience localized periods of transient neuronal silence termed off-states."},{"dimension":"date","status":"pass","explanation":"The originalPublished date of 2026-11-10 matches the v1 submission date on arXiv.","excerpt":"[Submitted on 10 Nov 2026]"},{"dimension":"support","status":"pass","explanation":"The figures for 250 derivations, two 8B models and both accuracy pairs match the source. So do the milder degradation under dispersed budget, the exploratory acute and inconclusive results, and the error patterns. The interpretive sentence that where disruption falls may matter is hedged and limited to these tests. Its sharper source framing is concentrated versus dispersed rather than position, but it does not contradict the source. The limitations and question follow from the stated scope.","excerpt":"dispersing the identical cumulative budget uniformly across generation steps via stratified quintile matching yields substantially milder degradation"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-11T20:47:24.994Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}