{"id":"20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e","title":"arXiv cs.AI: Zero-Shot Quantization via Weight-Space Arithmetic","argument":"Authors Daniele Solombrino, Antonio Andrea Gargiulo, Alessandro Zirilli, Luca Zhou, Adrian Robert Minut and Emanuele Rodolà argue that a model’s robustness to post-training quantization (PTQ) can be transferred through a “quantization vector.” They extract this vector from a donor task using weight-space arithmetic, then use it to patch a receiver model. The authors report improved PTQ Top-1 accuracy of up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT). They describe the method as a zero-shot, low-cost alternative to QAT when no receiver validation data is used. Across multiple vision and language models and more than 30 tasks, they say donor vectors often yield substantial gains even when the tasks differ markedly. The central idea is that some robustness to quantization may be isolated in model weights and reused, rather than learned anew for each receiver. The abstract reports a maximum improvement, but does not give the underlying task-level results or specify how often that maximum occurred.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper’s abstract and does not provide task-level results, baselines, or details behind the reported maximum improvement. The abstract describes gains as occurring “often” and “substantial,” but does not quantify how frequently gains occur or their distribution across tasks.","speaker":{"name":"Solombrino, Daniele, Gargiulo, Antonio Andrea, Zirilli, Alessandro, Zhou, Luca, Minut, Adrian Robert, Rodolà, Emanuele","roleAtSource":"Article author"},"source":{"title":"Zero-Shot Quantization via Weight-Space Arithmetic","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2604.03420","publishedAt":"2026-04-03","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors argue that robustness to post-training quantization can be transferred as a direction in weight space, which they call the quantization vector.","excerpt":"We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The authors say a donor-task quantization vector can patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting without receiver-side QAT.","excerpt":"it can be used to patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT).","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The authors describe the approach as a zero-shot, low-cost alternative to QAT when no receiver validation data is used.","excerpt":"When using no receiver validation data, the method provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment.","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The authors report gains across multiple vision and language models and more than 30 tasks, including cases where donor and receiver tasks differ markedly.","excerpt":"Across multiple vision and language models and more than 30 tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly.","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The abstract presents quantization robustness as something that can be partially isolated, reused and transferred through weight-space arithmetic.","excerpt":"Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2604.03420","checkedAt":"2026-11-08T02:48:07.468Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which tasks and models produced the reported 60% improvement, and how often did donor quantization vectors improve accuracy across the reported evaluations?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Zero-Shot Quantization via Weight-Space Arithmetic","summary":"Authors Daniele Solombrino, Antonio Andrea Gargiulo, Alessandro Zirilli, Luca Zhou, Adrian Robert Minut and Emanuele Rodolà argue that a model’s robustness to post-training quantization (PTQ) can be transferred through a “quantization vector.” They extract this vector from a donor task using weight-space arithmetic, then use it to patch a receiver model. The authors report improved PTQ Top-1 accuracy of up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT). They describe the method as a zero-shot, low-cost alternative to QAT when no receiver validation data is used. Across multiple vision and language models and more than 30 tasks, they say donor vectors often yield substantial gains even when the tasks differ markedly. The central idea is that some robustness to quantization may be isolated in model weights and reused, rather than learned anew for each receiver. The abstract reports a maximum improvement, but does not give the underlying task-level results or specify how often that maximum occurred.","question":"Which tasks and models produced the reported 60% improvement, and how often did donor quantization vectors improve accuracy across the reported evaluations?","claims":[{"text":"The authors argue that robustness to post-training quantization can be transferred as a direction in weight space, which they call the quantization vector.","excerpt":"We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector","support":"source-stated"},{"text":"The authors say a donor-task quantization vector can patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting without receiver-side QAT.","excerpt":"it can be used to patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT).","support":"source-stated"},{"text":"The authors describe the approach as a zero-shot, low-cost alternative to QAT when no receiver validation data is used.","excerpt":"When using no receiver validation data, the method provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment.","support":"source-stated"},{"text":"The authors report gains across multiple vision and language models and more than 30 tasks, including cases where donor and receiver tasks differ markedly.","excerpt":"Across multiple vision and language models and more than 30 tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly.","support":"source-stated"},{"text":"The abstract presents quantization robustness as something that can be partially isolated, reused and transferred through weight-space arithmetic.","excerpt":"Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.","support":"source-stated"}],"limitations":["The supplied evidence is the paper’s abstract and does not provide task-level results, baselines, or details behind the reported maximum improvement.","The abstract describes gains as occurring “often” and “substantial,” but does not quantify how frequently gains occur or their distribution across tasks."],"mediaId":"f368b1bc8e001d5e476699253c30773f3f4c897da8748a8ecb9b853e36b93e0e","correctionReason":null,"id":"20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e","publicationId":"05b89360-8385-43aa-87d6-0ae580fa4cb8","version":1,"sourceUrl":"https://arxiv.org/abs/2604.03420","sourceTitle":"Zero-Shot Quantization via Weight-Space Arithmetic","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-04-03","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:03:27.710Z","createdAt":"2026-11-08T02:32:24.411Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"f368b1bc8e001d5e476699253c30773f3f4c897da8748a8ecb9b853e36b93e0e","kind":"article","url":"https://arxiv.org/abs/2604.03420","sourceUrl":"https://arxiv.org/abs/2604.03420","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-08T02:18:32.363Z","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/2604.03420","checkedAt":"2026-11-08T02:48:07.468Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"f368b1bc8e001d5e476699253c30773f3f4c897da8748a8ecb9b853e36b93e0e","confidence":0.8200000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"4bb0657023679637a55ba56e3d8725aba3011577e14233f4c96d835c60033c14","decidedAt":"2026-11-08T02:18:32.922Z"},"provenance":{"textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observationId":"30bc7201220ca153a68b37572bb74eb087b97068cfbc9527e8a2c380523befd0","retrievedAt":"2026-11-08T02:18:32.257Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"91a91ae3-591d-4db7-88d7-449845bc1b66"},"reviewer":{"model":"gpt-6-luna","runId":"66dc4265-7250-4786-804c-1b16d72b4c21"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T02:32:24.411Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Solombrino, Daniele","url":null,"kind":"unknown"},{"name":"Gargiulo, Antonio Andrea","url":null,"kind":"unknown"},{"name":"Zirilli, Alessandro","url":null,"kind":"unknown"},{"name":"Zhou, Luca","url":null,"kind":"unknown"},{"name":"Minut, Adrian Robert","url":null,"kind":"unknown"},{"name":"Rodolà, Emanuele","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Solombrino, Daniele","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Solombrino, Daniele","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Gargiulo, Antonio Andrea","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gargiulo, Antonio Andrea","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Zirilli, Alessandro","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zirilli, Alessandro","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Zhou, Luca","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhou, Luca","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Minut, Adrian Robert","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Minut, Adrian Robert","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Rodolà, Emanuele","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Rodolà, Emanuele","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2604.03420","snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z","basis":"publisher-page-metadata","sourceLinks":50,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-cs-ai-zero-shot-quantization-via-weight-space-arithmetic-20282c0c","formats":{"html":"/readings/arxiv-cs-ai-zero-shot-quantization-via-weight-space-arithmetic-20282c0c","markdown":"/readings/arxiv-cs-ai-zero-shot-quantization-via-weight-space-arithmetic-20282c0c.md","json":"/readings/arxiv-cs-ai-zero-shot-quantization-via-weight-space-arithmetic-20282c0c.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-zero-shot","name":"Zero-shot","explanation":"Doing a task without worked examples (zero-shot) or with only a few examples given in the prompt (few-shot).","referenceUrl":"https://developers.google.com/machine-learning/glossary#zero-shot-prompting"},{"id":"term-quantization","name":"Quantization","explanation":"Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.","referenceUrl":"https://developers.google.com/machine-learning/glossary#quantization"},{"id":"term-model-weights","name":"Model weights","explanation":"The learned numbers that make up a trained model; publishing them lets others run it.","referenceUrl":"https://developers.google.com/machine-learning/glossary#weight"}],"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":"20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e","revision":"05b89360-8385-43aa-87d6-0ae580fa4cb8","revisionKind":"publication-id","title":"arXiv cs.AI: Zero-Shot Quantization via Weight-Space Arithmetic","url":"https://escapemeridian.com/readings/arxiv-cs-ai-zero-shot-quantization-via-weight-space-arithmetic-20282c0c"},"source":{"type":"source","id":"https://arxiv.org/abs/2604.03420","revision":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","revisionKind":"sha256-source-text","title":"Zero-Shot Quantization via Weight-Space Arithmetic","url":"https://arxiv.org/abs/2604.03420"},"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":"20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e","title":"arXiv cs.AI: Zero-Shot Quantization via Weight-Space Arithmetic","argument":"Authors Daniele Solombrino, Antonio Andrea Gargiulo, Alessandro Zirilli, Luca Zhou, Adrian Robert Minut and Emanuele Rodolà argue that a model’s robustness to post-training quantization (PTQ) can be transferred through a “quantization vector.” They extract this vector from a donor task using weight-space arithmetic, then use it to patch a receiver model. The authors report improved PTQ Top-1 accuracy of up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT). They describe the method as a zero-shot, low-cost alternative to QAT when no receiver validation data is used. Across multiple vision and language models and more than 30 tasks, they say donor vectors often yield substantial gains even when the tasks differ markedly. The central idea is that some robustness to quantization may be isolated in model weights and reused, rather than learned anew for each receiver. The abstract reports a maximum improvement, but does not give the underlying task-level results or specify how often that maximum occurred.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The supplied evidence is the paper’s abstract and does not provide task-level results, baselines, or details behind the reported maximum improvement. The abstract describes gains as occurring “often” and “substantial,” but does not quantify how frequently gains occur or their distribution across tasks.","speaker":{"name":"Solombrino, Daniele, Gargiulo, Antonio Andrea, Zirilli, Alessandro, Zhou, Luca, Minut, Adrian Robert, Rodolà, Emanuele","roleAtSource":"Article author"},"source":{"title":"Zero-Shot Quantization via Weight-Space Arithmetic","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2604.03420","publishedAt":"2026-04-03","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors argue that robustness to post-training quantization can be transferred as a direction in weight space, which they call the quantization vector.","excerpt":"We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The authors say a donor-task quantization vector can patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting without receiver-side QAT.","excerpt":"it can be used to patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT).","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The authors describe the approach as a zero-shot, low-cost alternative to QAT when no receiver validation data is used.","excerpt":"When using no receiver validation data, the method provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment.","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The authors report gains across multiple vision and language models and more than 30 tasks, including cases where donor and receiver tasks differ markedly.","excerpt":"Across multiple vision and language models and more than 30 tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly.","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"},{"text":"The abstract presents quantization robustness as something that can be partially isolated, reused and transferred through weight-space arithmetic.","excerpt":"Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.","url":"https://arxiv.org/abs/2604.03420","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2604.03420","checkedAt":"2026-11-08T02:48:07.468Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Which tasks and models produced the reported 60% improvement, and how often did donor quantization vectors improve accuracy across the reported evaluations?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv cs.AI: Zero-Shot Quantization via Weight-Space Arithmetic","summary":"Authors Daniele Solombrino, Antonio Andrea Gargiulo, Alessandro Zirilli, Luca Zhou, Adrian Robert Minut and Emanuele Rodolà argue that a model’s robustness to post-training quantization (PTQ) can be transferred through a “quantization vector.” They extract this vector from a donor task using weight-space arithmetic, then use it to patch a receiver model. The authors report improved PTQ Top-1 accuracy of up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT). They describe the method as a zero-shot, low-cost alternative to QAT when no receiver validation data is used. Across multiple vision and language models and more than 30 tasks, they say donor vectors often yield substantial gains even when the tasks differ markedly. The central idea is that some robustness to quantization may be isolated in model weights and reused, rather than learned anew for each receiver. The abstract reports a maximum improvement, but does not give the underlying task-level results or specify how often that maximum occurred.","question":"Which tasks and models produced the reported 60% improvement, and how often did donor quantization vectors improve accuracy across the reported evaluations?","claims":[{"text":"The authors argue that robustness to post-training quantization can be transferred as a direction in weight space, which they call the quantization vector.","excerpt":"We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector","support":"source-stated"},{"text":"The authors say a donor-task quantization vector can patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting without receiver-side QAT.","excerpt":"it can be used to patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT).","support":"source-stated"},{"text":"The authors describe the approach as a zero-shot, low-cost alternative to QAT when no receiver validation data is used.","excerpt":"When using no receiver validation data, the method provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment.","support":"source-stated"},{"text":"The authors report gains across multiple vision and language models and more than 30 tasks, including cases where donor and receiver tasks differ markedly.","excerpt":"Across multiple vision and language models and more than 30 tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly.","support":"source-stated"},{"text":"The abstract presents quantization robustness as something that can be partially isolated, reused and transferred through weight-space arithmetic.","excerpt":"Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.","support":"source-stated"}],"limitations":["The supplied evidence is the paper’s abstract and does not provide task-level results, baselines, or details behind the reported maximum improvement.","The abstract describes gains as occurring “often” and “substantial,” but does not quantify how frequently gains occur or their distribution across tasks."],"mediaId":"f368b1bc8e001d5e476699253c30773f3f4c897da8748a8ecb9b853e36b93e0e","correctionReason":null,"id":"20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e","publicationId":"05b89360-8385-43aa-87d6-0ae580fa4cb8","version":1,"sourceUrl":"https://arxiv.org/abs/2604.03420","sourceTitle":"Zero-Shot Quantization via Weight-Space Arithmetic","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-04-03","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-06T05:03:27.710Z","createdAt":"2026-11-08T02:32:24.411Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"f368b1bc8e001d5e476699253c30773f3f4c897da8748a8ecb9b853e36b93e0e","kind":"article","url":"https://arxiv.org/abs/2604.03420","sourceUrl":"https://arxiv.org/abs/2604.03420","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-08T02:18:32.363Z","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/2604.03420","checkedAt":"2026-11-08T02:48:07.468Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"f368b1bc8e001d5e476699253c30773f3f4c897da8748a8ecb9b853e36b93e0e","confidence":0.8200000000000001,"provider":"typesafe","model":"jev-1.13.0","inputHash":"4bb0657023679637a55ba56e3d8725aba3011577e14233f4c96d835c60033c14","decidedAt":"2026-11-08T02:18:32.922Z"},"provenance":{"textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observationId":"30bc7201220ca153a68b37572bb74eb087b97068cfbc9527e8a2c380523befd0","retrievedAt":"2026-11-08T02:18:32.257Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"91a91ae3-591d-4db7-88d7-449845bc1b66"},"reviewer":{"model":"gpt-6-luna","runId":"66dc4265-7250-4786-804c-1b16d72b4c21"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-08T02:32:24.411Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Solombrino, Daniele","url":null,"kind":"unknown"},{"name":"Gargiulo, Antonio Andrea","url":null,"kind":"unknown"},{"name":"Zirilli, Alessandro","url":null,"kind":"unknown"},{"name":"Zhou, Luca","url":null,"kind":"unknown"},{"name":"Minut, Adrian Robert","url":null,"kind":"unknown"},{"name":"Rodolà, Emanuele","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Solombrino, Daniele","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Solombrino, Daniele","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Gargiulo, Antonio Andrea","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Gargiulo, Antonio Andrea","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Zirilli, Alessandro","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zirilli, Alessandro","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Zhou, Luca","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhou, Luca","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Minut, Adrian Robert","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Minut, Adrian Robert","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}},{"actor":{"name":"Rodolà, Emanuele","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Rodolà, Emanuele","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2604.03420","locator":"head/meta[name=author|citation_author]"},"assertionId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-08T02:48:07.468Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","sourceUrl":"https://arxiv.org/abs/2604.03420","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z"}}],"participationCoverage":{"total":6,"returned":6,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2604.03420","snapshotId":"8eb9a5bbd85034af66e03834e764e5ae8fcf54b92afef25d6e8f3c98662b2afd","contentSha256":"69f10d80aea1f0ba42e420ee996553eefdfd9d3a9b98ed1de0128e5b55a95715","textHash":"75389c3510107e491362077038e345c3e1a33b8366c7156fb42bb04fb3f75fd3","observedAt":"2026-11-08T02:48:07.468Z","basis":"publisher-page-metadata","sourceLinks":50,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/20282c0c79de95663239677ee6cb30098cd0ee604b8bee1aca23c6e88eda668e/source"}}}],"audits":[{"publicationId":"05b89360-8385-43aa-87d6-0ae580fa4cb8","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"All six authors are listed correctly. Each claim is attributed to them as their assertion, not stated as established fact. The arXiv cs.AI label matches the supplied publisher; the paper's primary listing is cs.CV, but it is cross-listed in cs.AI.","excerpt":"Authors:Daniele Solombrino, Antonio Andrea Gargiulo, Alessandro Zirilli, Luca Zhou, Adrian Robert Minut, Emanuele Rodolà"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful, but the central concepts remain opaque to a non-specialist. It never explains in plain language what quantization is or why 3-bit deployment matters, so the cost and deployment stakes are not stated. Smallest repair: one plain sentence on quantization, plus term-library tooltips for PTQ, QAT, Top-1 accuracy, weight space and zero-shot. Two smaller fidelity points: the reading drops the scope condition 'for extremely low-bit deployment' and omits the claimed proof that quantization vectors are well-defined despite reparameterization symmetries. It also softens the authors' 'We show' to 'argue'.","excerpt":"When using no receiver validation data, the method provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment."},{"dimension":"date","status":"pass","explanation":"The displayed original publication date of 2026-04-03 matches the v1 submission date. The v4 revision of 5 Nov 2026 and the November feed date are correctly not used.","excerpt":"[Submitted on 3 Apr 2026 (v1), last revised 5 Nov 2026 (this version, v4)]"},{"dimension":"support","status":"pass","explanation":"Each claim matches its verbatim abstract excerpt and is framed as the authors' report. The limitations correctly note that the abstract gives no task-level results or frequency for 'often'. The question follows from that gap without implying an answer. No unsupported benefits or costs are asserted. This grounding does not independently verify the paper's results.","excerpt":"donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the original arXiv abstract page, which is the source itself.","excerpt":null}],"createdAt":"2026-11-08T02:32:24.411Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}