{"id":"5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626","title":"AC-JEPA combines cross-modal prediction with supervised training","argument":"Researchers Chaofei Hong, Ziyang Chen, Baochen Wang, Christina Yi Jin and Qing Zhang introduce AC-JEPA, a method designed to translate between visual and symbolic data when paired examples may be imperfect. Their approach combines supervised fine-tuning (SFT), which trains against target tokens, with a joint-embedding objective that predicts intermediate representations across modalities. The authors test it on optical chemical structure recognition, where visual inputs are translated into symbolic chemical structures.\n\nOn the Simple set of the WildMol benchmark, the authors report 76.07% accuracy for AC-JEPA trained on clean pairs, compared with 67.40% for SFT alone. In self-training with the same pseudo-labeled data, AC-JEPA reaches 80.23% accuracy; direct SFT falls from 67.40% to 63.47%. The authors argue that learning from continuous latent targets can better tolerate local errors in pseudo-labeled pairs than treating pseudo-tokens as hard categorical supervision.\n\nThese results are limited to the reported benchmark and setup; the abstract does not establish how the method performs in other scientific domains or real-world use. How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports results on optical chemical structure recognition and the Simple set of the WildMol benchmark; it does not establish performance in other domains or real-world use. The abstract does not provide details about the benchmark setup or evidence of transfer beyond the reported testbed.","speaker":{"name":"Hong, Chaofei, Chen, Ziyang, Wang, Baochen, Jin, Christina Yi, Zhang, Qing","roleAtSource":"Article author"},"source":{"title":"AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic Translation","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.36493","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors introduce AC-JEPA as a method for cross-modal translation that combines a joint-embedding objective with direct token-level supervision.","excerpt":"The joint-embedding objective targets cross-modal structure predictable from context, while the causal SFT path retains direct token-level supervision for the target sequence.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"},{"text":"On the Simple set of the WildMol benchmark, AC-JEPA trained on clean pairs achieved 76.07% accuracy, compared with 67.40% for SFT alone.","excerpt":"AC-JEPA trained on clean pairs achieves 76.07% accuracy on the Simple set of the WildMol benchmark, compared with 67.40% for SFT alone.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"},{"text":"With the same pseudo-labeled data, AC-JEPA-based self-training achieved 80.23% accuracy, while direct SFT accuracy fell from 67.40% to 63.47%.","excerpt":"Using the same pseudo-labeled data for both approaches, AC-JEPA-based self-training further raises accuracy to 80.23%, whereas accuracy under direct SFT drops from 67.40% to 63.47%.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"},{"text":"The authors attribute tolerance to local errors to learning from continuous latent targets rather than treating pseudo-tokens as hard categorical supervision.","excerpt":"AC-JEPA learns from continuous latent targets derived from imperfect pseudo-labeled pairs, offering intrinsic tolerance to local data errors without promoting pseudo-tokens to hard categorical supervision.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.36493","checkedAt":"2026-11-06T14:24:06.509Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"AC-JEPA combines cross-modal prediction with supervised training","summary":"Researchers Chaofei Hong, Ziyang Chen, Baochen Wang, Christina Yi Jin and Qing Zhang introduce AC-JEPA, a method designed to translate between visual and symbolic data when paired examples may be imperfect. Their approach combines supervised fine-tuning (SFT), which trains against target tokens, with a joint-embedding objective that predicts intermediate representations across modalities. The authors test it on optical chemical structure recognition, where visual inputs are translated into symbolic chemical structures.\n\nOn the Simple set of the WildMol benchmark, the authors report 76.07% accuracy for AC-JEPA trained on clean pairs, compared with 67.40% for SFT alone. In self-training with the same pseudo-labeled data, AC-JEPA reaches 80.23% accuracy; direct SFT falls from 67.40% to 63.47%. The authors argue that learning from continuous latent targets can better tolerate local errors in pseudo-labeled pairs than treating pseudo-tokens as hard categorical supervision.\n\nThese results are limited to the reported benchmark and setup; the abstract does not establish how the method performs in other scientific domains or real-world use. How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?","question":"How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?","claims":[{"text":"The authors introduce AC-JEPA as a method for cross-modal translation that combines a joint-embedding objective with direct token-level supervision.","excerpt":"The joint-embedding objective targets cross-modal structure predictable from context, while the causal SFT path retains direct token-level supervision for the target sequence.","support":"source-stated"},{"text":"On the Simple set of the WildMol benchmark, AC-JEPA trained on clean pairs achieved 76.07% accuracy, compared with 67.40% for SFT alone.","excerpt":"AC-JEPA trained on clean pairs achieves 76.07% accuracy on the Simple set of the WildMol benchmark, compared with 67.40% for SFT alone.","support":"source-stated"},{"text":"With the same pseudo-labeled data, AC-JEPA-based self-training achieved 80.23% accuracy, while direct SFT accuracy fell from 67.40% to 63.47%.","excerpt":"Using the same pseudo-labeled data for both approaches, AC-JEPA-based self-training further raises accuracy to 80.23%, whereas accuracy under direct SFT drops from 67.40% to 63.47%.","support":"source-stated"},{"text":"The authors attribute tolerance to local errors to learning from continuous latent targets rather than treating pseudo-tokens as hard categorical supervision.","excerpt":"AC-JEPA learns from continuous latent targets derived from imperfect pseudo-labeled pairs, offering intrinsic tolerance to local data errors without promoting pseudo-tokens to hard categorical supervision.","support":"source-stated"}],"limitations":["The abstract reports results on optical chemical structure recognition and the Simple set of the WildMol benchmark; it does not establish performance in other domains or real-world use.","The abstract does not provide details about the benchmark setup or evidence of transfer beyond the reported testbed."],"mediaId":"56099918e6c85749d3279236edfd29f40e2911bf545748facdef2066ed1362cd","correctionReason":null,"id":"5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626","publicationId":"844b7639-b42a-4019-8bb2-fb41e72961d5","version":1,"sourceUrl":"https://arxiv.org/abs/2610.36493","sourceTitle":"AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic Translation","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T08:52:32.166Z","createdAt":"2026-11-06T12:16:28.698Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"56099918e6c85749d3279236edfd29f40e2911bf545748facdef2066ed1362cd","kind":"article","url":"https://arxiv.org/abs/2610.36493","sourceUrl":"https://arxiv.org/abs/2610.36493","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T00:47:28.525Z","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/2610.36493","checkedAt":"2026-11-06T14:24:06.509Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"56099918e6c85749d3279236edfd29f40e2911bf545748facdef2066ed1362cd","confidence":0.91,"provider":"typesafe","model":"jev-1.13.0","inputHash":"245d0a458a9926f5fd511dfdfb57a7971782df7c75fae9b1ea7929194752adb2","decidedAt":"2026-10-30T00:47:29.076Z"},"provenance":{"textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observationId":"2b7b408619e5ad040df2f7c086988d479c7463b202d65b06210be5d86819489e","retrievedAt":"2026-10-30T00:47:28.445Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"315c96c0-0caf-4be9-8982-995820236e49"},"reviewer":{"model":"gpt-6-luna","runId":"cbdd190e-d123-40fb-84e6-18c4d60ef522"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-06T12:16:28.698Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Hong, Chaofei","url":null,"kind":"unknown"},{"name":"Chen, Ziyang","url":null,"kind":"unknown"},{"name":"Wang, Baochen","url":null,"kind":"unknown"},{"name":"Jin, Christina Yi","url":null,"kind":"unknown"},{"name":"Zhang, Qing","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Hong, Chaofei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Chaofei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Chen, Ziyang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chen, Ziyang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Wang, Baochen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Baochen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Jin, Christina Yi","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Jin, Christina Yi","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Zhang, Qing","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhang, Qing","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.36493","snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626/source"}},"canonicalUrl":"https://escapemeridian.com/readings/ac-jepa-combines-cross-modal-prediction-with-supervised-training-5c13d48d","formats":{"html":"/readings/ac-jepa-combines-cross-modal-prediction-with-supervised-training-5c13d48d","markdown":"/readings/ac-jepa-combines-cross-modal-prediction-with-supervised-training-5c13d48d.md","json":"/readings/ac-jepa-combines-cross-modal-prediction-with-supervised-training-5c13d48d.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-fine-tuning","name":"Fine-tuning","explanation":"Further training of an existing model to adapt it to a particular task or use.","referenceUrl":"https://developers.google.com/machine-learning/glossary#fine-tuning"},{"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":"5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626","revision":"844b7639-b42a-4019-8bb2-fb41e72961d5","revisionKind":"publication-id","title":"AC-JEPA combines cross-modal prediction with supervised training","url":"https://escapemeridian.com/readings/ac-jepa-combines-cross-modal-prediction-with-supervised-training-5c13d48d"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.36493","revision":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","revisionKind":"sha256-source-text","title":"AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic Translation","url":"https://arxiv.org/abs/2610.36493"},"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":"5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626","title":"AC-JEPA combines cross-modal prediction with supervised training","argument":"Researchers Chaofei Hong, Ziyang Chen, Baochen Wang, Christina Yi Jin and Qing Zhang introduce AC-JEPA, a method designed to translate between visual and symbolic data when paired examples may be imperfect. Their approach combines supervised fine-tuning (SFT), which trains against target tokens, with a joint-embedding objective that predicts intermediate representations across modalities. The authors test it on optical chemical structure recognition, where visual inputs are translated into symbolic chemical structures.\n\nOn the Simple set of the WildMol benchmark, the authors report 76.07% accuracy for AC-JEPA trained on clean pairs, compared with 67.40% for SFT alone. In self-training with the same pseudo-labeled data, AC-JEPA reaches 80.23% accuracy; direct SFT falls from 67.40% to 63.47%. The authors argue that learning from continuous latent targets can better tolerate local errors in pseudo-labeled pairs than treating pseudo-tokens as hard categorical supervision.\n\nThese results are limited to the reported benchmark and setup; the abstract does not establish how the method performs in other scientific domains or real-world use. How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The abstract reports results on optical chemical structure recognition and the Simple set of the WildMol benchmark; it does not establish performance in other domains or real-world use. The abstract does not provide details about the benchmark setup or evidence of transfer beyond the reported testbed.","speaker":{"name":"Hong, Chaofei, Chen, Ziyang, Wang, Baochen, Jin, Christina Yi, Zhang, Qing","roleAtSource":"Article author"},"source":{"title":"AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic Translation","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.36493","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors introduce AC-JEPA as a method for cross-modal translation that combines a joint-embedding objective with direct token-level supervision.","excerpt":"The joint-embedding objective targets cross-modal structure predictable from context, while the causal SFT path retains direct token-level supervision for the target sequence.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"},{"text":"On the Simple set of the WildMol benchmark, AC-JEPA trained on clean pairs achieved 76.07% accuracy, compared with 67.40% for SFT alone.","excerpt":"AC-JEPA trained on clean pairs achieves 76.07% accuracy on the Simple set of the WildMol benchmark, compared with 67.40% for SFT alone.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"},{"text":"With the same pseudo-labeled data, AC-JEPA-based self-training achieved 80.23% accuracy, while direct SFT accuracy fell from 67.40% to 63.47%.","excerpt":"Using the same pseudo-labeled data for both approaches, AC-JEPA-based self-training further raises accuracy to 80.23%, whereas accuracy under direct SFT drops from 67.40% to 63.47%.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"},{"text":"The authors attribute tolerance to local errors to learning from continuous latent targets rather than treating pseudo-tokens as hard categorical supervision.","excerpt":"AC-JEPA learns from continuous latent targets derived from imperfect pseudo-labeled pairs, offering intrinsic tolerance to local data errors without promoting pseudo-tokens to hard categorical supervision.","url":"https://arxiv.org/abs/2610.36493","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.36493","checkedAt":"2026-11-06T14:24:06.509Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"AC-JEPA combines cross-modal prediction with supervised training","summary":"Researchers Chaofei Hong, Ziyang Chen, Baochen Wang, Christina Yi Jin and Qing Zhang introduce AC-JEPA, a method designed to translate between visual and symbolic data when paired examples may be imperfect. Their approach combines supervised fine-tuning (SFT), which trains against target tokens, with a joint-embedding objective that predicts intermediate representations across modalities. The authors test it on optical chemical structure recognition, where visual inputs are translated into symbolic chemical structures.\n\nOn the Simple set of the WildMol benchmark, the authors report 76.07% accuracy for AC-JEPA trained on clean pairs, compared with 67.40% for SFT alone. In self-training with the same pseudo-labeled data, AC-JEPA reaches 80.23% accuracy; direct SFT falls from 67.40% to 63.47%. The authors argue that learning from continuous latent targets can better tolerate local errors in pseudo-labeled pairs than treating pseudo-tokens as hard categorical supervision.\n\nThese results are limited to the reported benchmark and setup; the abstract does not establish how the method performs in other scientific domains or real-world use. How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?","question":"How well does AC-JEPA transfer beyond optical chemical structure recognition, especially when paired data are imperfect?","claims":[{"text":"The authors introduce AC-JEPA as a method for cross-modal translation that combines a joint-embedding objective with direct token-level supervision.","excerpt":"The joint-embedding objective targets cross-modal structure predictable from context, while the causal SFT path retains direct token-level supervision for the target sequence.","support":"source-stated"},{"text":"On the Simple set of the WildMol benchmark, AC-JEPA trained on clean pairs achieved 76.07% accuracy, compared with 67.40% for SFT alone.","excerpt":"AC-JEPA trained on clean pairs achieves 76.07% accuracy on the Simple set of the WildMol benchmark, compared with 67.40% for SFT alone.","support":"source-stated"},{"text":"With the same pseudo-labeled data, AC-JEPA-based self-training achieved 80.23% accuracy, while direct SFT accuracy fell from 67.40% to 63.47%.","excerpt":"Using the same pseudo-labeled data for both approaches, AC-JEPA-based self-training further raises accuracy to 80.23%, whereas accuracy under direct SFT drops from 67.40% to 63.47%.","support":"source-stated"},{"text":"The authors attribute tolerance to local errors to learning from continuous latent targets rather than treating pseudo-tokens as hard categorical supervision.","excerpt":"AC-JEPA learns from continuous latent targets derived from imperfect pseudo-labeled pairs, offering intrinsic tolerance to local data errors without promoting pseudo-tokens to hard categorical supervision.","support":"source-stated"}],"limitations":["The abstract reports results on optical chemical structure recognition and the Simple set of the WildMol benchmark; it does not establish performance in other domains or real-world use.","The abstract does not provide details about the benchmark setup or evidence of transfer beyond the reported testbed."],"mediaId":"56099918e6c85749d3279236edfd29f40e2911bf545748facdef2066ed1362cd","correctionReason":null,"id":"5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626","publicationId":"844b7639-b42a-4019-8bb2-fb41e72961d5","version":1,"sourceUrl":"https://arxiv.org/abs/2610.36493","sourceTitle":"AC-JEPA: Anchored Cross-Modal Joint Embedding Predictive Architecture for Visual-to-Symbolic Translation","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T08:52:32.166Z","createdAt":"2026-11-06T12:16:28.698Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"56099918e6c85749d3279236edfd29f40e2911bf545748facdef2066ed1362cd","kind":"article","url":"https://arxiv.org/abs/2610.36493","sourceUrl":"https://arxiv.org/abs/2610.36493","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T00:47:28.525Z","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/2610.36493","checkedAt":"2026-11-06T14:24:06.509Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"56099918e6c85749d3279236edfd29f40e2911bf545748facdef2066ed1362cd","confidence":0.91,"provider":"typesafe","model":"jev-1.13.0","inputHash":"245d0a458a9926f5fd511dfdfb57a7971782df7c75fae9b1ea7929194752adb2","decidedAt":"2026-10-30T00:47:29.076Z"},"provenance":{"textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observationId":"2b7b408619e5ad040df2f7c086988d479c7463b202d65b06210be5d86819489e","retrievedAt":"2026-10-30T00:47:28.445Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"315c96c0-0caf-4be9-8982-995820236e49"},"reviewer":{"model":"gpt-6-luna","runId":"cbdd190e-d123-40fb-84e6-18c4d60ef522"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-06T12:16:28.698Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Hong, Chaofei","url":null,"kind":"unknown"},{"name":"Chen, Ziyang","url":null,"kind":"unknown"},{"name":"Wang, Baochen","url":null,"kind":"unknown"},{"name":"Jin, Christina Yi","url":null,"kind":"unknown"},{"name":"Zhang, Qing","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Hong, Chaofei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Hong, Chaofei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Chen, Ziyang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chen, Ziyang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Wang, Baochen","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Baochen","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Jin, Christina Yi","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Jin, Christina Yi","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}},{"actor":{"name":"Zhang, Qing","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Zhang, Qing","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36493","locator":"head/meta[name=author|citation_author]"},"assertionId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T14:24:06.509Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","sourceUrl":"https://arxiv.org/abs/2610.36493","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z"}}],"participationCoverage":{"total":5,"returned":5,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.36493","snapshotId":"6bf52515d44ded5274961cf201d0e55f262bec94e38d3e019d1fc65f4e779b25","contentSha256":"475328d6e513ce2eae14b43822b791bf18c3bd0fabd27acbdd25c9ea99086e83","textHash":"a0040252db7f6f6726697775d74007b66ed99d89ce8b4a4ea776bcea05fe0b6e","observedAt":"2026-11-06T14:24:06.509Z","basis":"publisher-page-metadata","sourceLinks":44,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/5c13d48db3464ee702c119ef8c4eba43b662307422fb9f4c31d317d13aaa4626/source"}}}],"audits":[{"publicationId":"844b7639-b42a-4019-8bb2-fb41e72961d5","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly names the five listed authors and attributes the method, results and interpretation to them. It keeps benchmark figures and the error-tolerance argument as the authors' reported claims.","excerpt":"Authors:Chaofei Hong, Ziyang Chen, Baochen Wang, Christina Yi Jin, Qing Zhang"},{"dimension":"summary","status":"uncertain","explanation":"The reading is faithful, but the terms behind its central finding are not explained for a non-specialist. Pseudo-labeled data, self-training, continuous latent targets, tokens and joint-embedding (JEPA, never expanded) are unexplained. The headline result is that AC-JEPA gains while SFT drops on imperfect pseudo-labels, so readers cannot see why that matters. Smallest repair: briefly explain pseudo-labels as model-generated, possibly wrong training answers, and self-training as learning from them. Add term-library tooltips for JEPA, SFT and latent targets. The title and question are otherwise sound.","excerpt":"During self-training, AC-JEPA learns from continuous latent targets derived from imperfect pseudo-labeled pairs"},{"dimension":"date","status":"pass","explanation":"The displayed original publication date of 2026-10-27 matches the arXiv v1 submission date. The later feed date is not used.","excerpt":"[Submitted on 27 Oct 2026]"},{"dimension":"support","status":"pass","explanation":"All four claims match the abstract, including 76.07% vs 67.40%, 80.23%, and the SFT drop to 63.47%. The limitations paragraph accurately notes that only one testbed and benchmark subset are reported. It does not invent benefits or stakeholders, and the question follows from the evidence.","excerpt":"Using the same pseudo-labeled data for both approaches, AC-JEPA-based self-training further raises accuracy to 80.23%, whereas accuracy under direct SFT drops from 67.40% to 63.47%."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the original arXiv abstract page. This is the most direct access to the source.","excerpt":null}],"createdAt":"2026-11-06T12:16:28.698Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}