{"id":"795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087","title":"arXiv: GUI-JEPA: Candidate-Conditioned JEPA Latents for Mobile GUI Agents","argument":"In a paper by Bingchao Wang and coauthors, the authors introduce GUI-JEPA, a mobile GUI agent policy that selects actions using latent representations rather than generating action text token by token. A latent representation is an internal encoding used by a model; JEPA is the architecture named in the paper’s title and abstract.\n\nThe authors say the policy predicts a representation for each candidate action and scores the candidates in one batched pass, without decoding tokens. They report that, on AndroidControl and on the same GPU, GUI-JEPA was 2.8–4.0x faster than a Qwen2.5-VL-7B agent they fine-tuned with the same inputs. They also report 86–92% of that agent’s in-distribution accuracy using 6% of its parameters.\n\nThese results suggest a possible speed and memory advantage for this approach in the tested setting, but the abstract does not establish how it performs across other apps, devices or conditions. Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The reported speed and accuracy comparison in the abstract is on AndroidControl and against one fine-tuned Qwen2.5-VL-7B agent using the same inputs. The abstract reports in-distribution accuracy; it does not establish performance across other apps, devices or conditions.","speaker":{"name":"Wang, Bingchao, Jing, Yongcheng, Liu, Shunyu, Wang, Chengyu, Huang, Jun, Yang, Jie, Liu, Wei, Tao, Dacheng","roleAtSource":"Article author"},"source":{"title":"GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.36877","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors present GUI-JEPA as a decoder-free policy for selecting actions for mobile GUI agents using JEPA latents.","excerpt":"we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA)","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"},{"text":"The policy predicts a latent for each candidate action and scores the candidates together without token decoding.","excerpt":"letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token.","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"},{"text":"On AndroidControl, the authors report GUI-JEPA ran 2.8–4.0x faster on the same GPU than their fine-tuned Qwen2.5-VL-7B comparison agent.","excerpt":"On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"},{"text":"The authors report GUI-JEPA reached 86–92% of the comparison agent’s in-distribution accuracy with 6% of its parameters.","excerpt":"while reaching 86-92% of its in-distribution accuracy with 6% of its parameters.","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.36877","checkedAt":"2026-11-06T01:04:07.737Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv: GUI-JEPA: Candidate-Conditioned JEPA Latents for Mobile GUI Agents","summary":"In a paper by Bingchao Wang and coauthors, the authors introduce GUI-JEPA, a mobile GUI agent policy that selects actions using latent representations rather than generating action text token by token. A latent representation is an internal encoding used by a model; JEPA is the architecture named in the paper’s title and abstract.\n\nThe authors say the policy predicts a representation for each candidate action and scores the candidates in one batched pass, without decoding tokens. They report that, on AndroidControl and on the same GPU, GUI-JEPA was 2.8–4.0x faster than a Qwen2.5-VL-7B agent they fine-tuned with the same inputs. They also report 86–92% of that agent’s in-distribution accuracy using 6% of its parameters.\n\nThese results suggest a possible speed and memory advantage for this approach in the tested setting, but the abstract does not establish how it performs across other apps, devices or conditions. Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?","question":"Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?","claims":[{"text":"The authors present GUI-JEPA as a decoder-free policy for selecting actions for mobile GUI agents using JEPA latents.","excerpt":"we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA)","support":"source-stated"},{"text":"The policy predicts a latent for each candidate action and scores the candidates together without token decoding.","excerpt":"letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token.","support":"source-stated"},{"text":"On AndroidControl, the authors report GUI-JEPA ran 2.8–4.0x faster on the same GPU than their fine-tuned Qwen2.5-VL-7B comparison agent.","excerpt":"On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs","support":"source-stated"},{"text":"The authors report GUI-JEPA reached 86–92% of the comparison agent’s in-distribution accuracy with 6% of its parameters.","excerpt":"while reaching 86-92% of its in-distribution accuracy with 6% of its parameters.","support":"source-stated"}],"limitations":["The reported speed and accuracy comparison in the abstract is on AndroidControl and against one fine-tuned Qwen2.5-VL-7B agent using the same inputs.","The abstract reports in-distribution accuracy; it does not establish performance across other apps, devices or conditions."],"mediaId":"2a957930710eebf4bcf015228466006c9e5b5e1b317b14ccf510e0fc4f784b17","correctionReason":null,"id":"795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087","publicationId":"389b9f8d-2c27-43a2-8d31-432649c1700b","version":1,"sourceUrl":"https://arxiv.org/abs/2610.36877","sourceTitle":"GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T08:46:32.247Z","createdAt":"2026-11-06T00:32:06.935Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"2a957930710eebf4bcf015228466006c9e5b5e1b317b14ccf510e0fc4f784b17","kind":"article","url":"https://arxiv.org/abs/2610.36877","sourceUrl":"https://arxiv.org/abs/2610.36877","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T02:32:59.707Z","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/2610.36877","checkedAt":"2026-11-06T01:04:07.737Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"2a957930710eebf4bcf015228466006c9e5b5e1b317b14ccf510e0fc4f784b17","confidence":0.94,"provider":"typesafe","model":"jev-1.13.0","inputHash":"62b490853595d2e5acf4a4a7690e1cee2157a09f8e6d94d6e60b5aacb2193d7a","decidedAt":"2026-10-30T02:33:00.280Z"},"provenance":{"textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observationId":"7d6ffcab577286b76a5f8e53e569d0800eb48cf6e393d02a813e0e829302fd2a","retrievedAt":"2026-10-30T02:32:59.628Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"e693e725-b32a-4e7f-8e3d-4f7917a7f9a4"},"reviewer":{"model":"gpt-6-luna","runId":"c4fe618f-15fe-4983-8be4-abb84291ae56"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-06T00:32:06.935Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Wang, Bingchao","url":null,"kind":"unknown"},{"name":"Jing, Yongcheng","url":null,"kind":"unknown"},{"name":"Liu, Shunyu","url":null,"kind":"unknown"},{"name":"Wang, Chengyu","url":null,"kind":"unknown"},{"name":"Huang, Jun","url":null,"kind":"unknown"},{"name":"Yang, Jie","url":null,"kind":"unknown"},{"name":"Liu, Wei","url":null,"kind":"unknown"},{"name":"Tao, Dacheng","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Wang, Bingchao","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Bingchao","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Jing, Yongcheng","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Jing, Yongcheng","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Liu, Shunyu","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Liu, Shunyu","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Wang, Chengyu","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Chengyu","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Huang, Jun","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Huang, Jun","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Yang, Jie","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Yang, Jie","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Liu, Wei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Liu, Wei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Tao, Dacheng","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Tao, Dacheng","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:7","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}}],"participationCoverage":{"total":8,"returned":8,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.36877","snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z","basis":"publisher-page-metadata","sourceLinks":49,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087/source"}},"canonicalUrl":"https://escapemeridian.com/readings/arxiv-gui-jepa-candidate-conditioned-jepa-latents-for-mobile-795c7138","formats":{"html":"/readings/arxiv-gui-jepa-candidate-conditioned-jepa-latents-for-mobile-795c7138","markdown":"/readings/arxiv-gui-jepa-candidate-conditioned-jepa-latents-for-mobile-795c7138.md","json":"/readings/arxiv-gui-jepa-candidate-conditioned-jepa-latents-for-mobile-795c7138.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-gpu","name":"Graphics processing unit","explanation":"A chip that performs many calculations at once; the main hardware used to train and run AI models.","referenceUrl":"https://en.wikipedia.org/wiki/Graphics_processing_unit"},{"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"}],"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":"795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087","revision":"389b9f8d-2c27-43a2-8d31-432649c1700b","revisionKind":"publication-id","title":"arXiv: GUI-JEPA: Candidate-Conditioned JEPA Latents for Mobile GUI Agents","url":"https://escapemeridian.com/readings/arxiv-gui-jepa-candidate-conditioned-jepa-latents-for-mobile-795c7138"},"source":{"type":"source","id":"https://arxiv.org/abs/2610.36877","revision":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","revisionKind":"sha256-source-text","title":"GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents","url":"https://arxiv.org/abs/2610.36877"},"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":"795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087","title":"arXiv: GUI-JEPA: Candidate-Conditioned JEPA Latents for Mobile GUI Agents","argument":"In a paper by Bingchao Wang and coauthors, the authors introduce GUI-JEPA, a mobile GUI agent policy that selects actions using latent representations rather than generating action text token by token. A latent representation is an internal encoding used by a model; JEPA is the architecture named in the paper’s title and abstract.\n\nThe authors say the policy predicts a representation for each candidate action and scores the candidates in one batched pass, without decoding tokens. They report that, on AndroidControl and on the same GPU, GUI-JEPA was 2.8–4.0x faster than a Qwen2.5-VL-7B agent they fine-tuned with the same inputs. They also report 86–92% of that agent’s in-distribution accuracy using 6% of its parameters.\n\nThese results suggest a possible speed and memory advantage for this approach in the tested setting, but the abstract does not establish how it performs across other apps, devices or conditions. Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The reported speed and accuracy comparison in the abstract is on AndroidControl and against one fine-tuned Qwen2.5-VL-7B agent using the same inputs. The abstract reports in-distribution accuracy; it does not establish performance across other apps, devices or conditions.","speaker":{"name":"Wang, Bingchao, Jing, Yongcheng, Liu, Shunyu, Wang, Chengyu, Huang, Jun, Yang, Jie, Liu, Wei, Tao, Dacheng","roleAtSource":"Article author"},"source":{"title":"GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2610.36877","publishedAt":"2026-10-27","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors present GUI-JEPA as a decoder-free policy for selecting actions for mobile GUI agents using JEPA latents.","excerpt":"we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA)","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"},{"text":"The policy predicts a latent for each candidate action and scores the candidates together without token decoding.","excerpt":"letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token.","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"},{"text":"On AndroidControl, the authors report GUI-JEPA ran 2.8–4.0x faster on the same GPU than their fine-tuned Qwen2.5-VL-7B comparison agent.","excerpt":"On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"},{"text":"The authors report GUI-JEPA reached 86–92% of the comparison agent’s in-distribution accuracy with 6% of its parameters.","excerpt":"while reaching 86-92% of its in-distribution accuracy with 6% of its parameters.","url":"https://arxiv.org/abs/2610.36877","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2610.36877","checkedAt":"2026-11-06T01:04:07.737Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"arXiv: GUI-JEPA: Candidate-Conditioned JEPA Latents for Mobile GUI Agents","summary":"In a paper by Bingchao Wang and coauthors, the authors introduce GUI-JEPA, a mobile GUI agent policy that selects actions using latent representations rather than generating action text token by token. A latent representation is an internal encoding used by a model; JEPA is the architecture named in the paper’s title and abstract.\n\nThe authors say the policy predicts a representation for each candidate action and scores the candidates in one batched pass, without decoding tokens. They report that, on AndroidControl and on the same GPU, GUI-JEPA was 2.8–4.0x faster than a Qwen2.5-VL-7B agent they fine-tuned with the same inputs. They also report 86–92% of that agent’s in-distribution accuracy using 6% of its parameters.\n\nThese results suggest a possible speed and memory advantage for this approach in the tested setting, but the abstract does not establish how it performs across other apps, devices or conditions. Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?","question":"Does GUI-JEPA retain its reported speed and accuracy when tested beyond AndroidControl and the paper’s in-distribution setting?","claims":[{"text":"The authors present GUI-JEPA as a decoder-free policy for selecting actions for mobile GUI agents using JEPA latents.","excerpt":"we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA)","support":"source-stated"},{"text":"The policy predicts a latent for each candidate action and scores the candidates together without token decoding.","excerpt":"letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token.","support":"source-stated"},{"text":"On AndroidControl, the authors report GUI-JEPA ran 2.8–4.0x faster on the same GPU than their fine-tuned Qwen2.5-VL-7B comparison agent.","excerpt":"On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs","support":"source-stated"},{"text":"The authors report GUI-JEPA reached 86–92% of the comparison agent’s in-distribution accuracy with 6% of its parameters.","excerpt":"while reaching 86-92% of its in-distribution accuracy with 6% of its parameters.","support":"source-stated"}],"limitations":["The reported speed and accuracy comparison in the abstract is on AndroidControl and against one fine-tuned Qwen2.5-VL-7B agent using the same inputs.","The abstract reports in-distribution accuracy; it does not establish performance across other apps, devices or conditions."],"mediaId":"2a957930710eebf4bcf015228466006c9e5b5e1b317b14ccf510e0fc4f784b17","correctionReason":null,"id":"795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087","publicationId":"389b9f8d-2c27-43a2-8d31-432649c1700b","version":1,"sourceUrl":"https://arxiv.org/abs/2610.36877","sourceTitle":"GUI-JEPA: Candidate-Conditioned JEPA Latents as a Fast Action Policy for Mobile GUI Agents","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-10-27","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-10-28T08:46:32.247Z","createdAt":"2026-11-06T00:32:06.935Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"2a957930710eebf4bcf015228466006c9e5b5e1b317b14ccf510e0fc4f784b17","kind":"article","url":"https://arxiv.org/abs/2610.36877","sourceUrl":"https://arxiv.org/abs/2610.36877","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-10-30T02:32:59.707Z","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/2610.36877","checkedAt":"2026-11-06T01:04:07.737Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"2a957930710eebf4bcf015228466006c9e5b5e1b317b14ccf510e0fc4f784b17","confidence":0.94,"provider":"typesafe","model":"jev-1.13.0","inputHash":"62b490853595d2e5acf4a4a7690e1cee2157a09f8e6d94d6e60b5aacb2193d7a","decidedAt":"2026-10-30T02:33:00.280Z"},"provenance":{"textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observationId":"7d6ffcab577286b76a5f8e53e569d0800eb48cf6e393d02a813e0e829302fd2a","retrievedAt":"2026-10-30T02:32:59.628Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"e693e725-b32a-4e7f-8e3d-4f7917a7f9a4"},"reviewer":{"model":"gpt-6-luna","runId":"c4fe618f-15fe-4983-8be4-abb84291ae56"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-06T00:32:06.935Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Wang, Bingchao","url":null,"kind":"unknown"},{"name":"Jing, Yongcheng","url":null,"kind":"unknown"},{"name":"Liu, Shunyu","url":null,"kind":"unknown"},{"name":"Wang, Chengyu","url":null,"kind":"unknown"},{"name":"Huang, Jun","url":null,"kind":"unknown"},{"name":"Yang, Jie","url":null,"kind":"unknown"},{"name":"Liu, Wei","url":null,"kind":"unknown"},{"name":"Tao, Dacheng","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Wang, Bingchao","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Bingchao","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Jing, Yongcheng","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Jing, Yongcheng","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Liu, Shunyu","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Liu, Shunyu","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Wang, Chengyu","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Wang, Chengyu","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Huang, Jun","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Huang, Jun","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Yang, Jie","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Yang, Jie","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Liu, Wei","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Liu, Wei","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}},{"actor":{"name":"Tao, Dacheng","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Tao, Dacheng","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2610.36877","locator":"head/meta[name=author|citation_author]"},"assertionId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1:participation:7","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-06T00:48:07.137Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","sourceUrl":"https://arxiv.org/abs/2610.36877","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z"}}],"participationCoverage":{"total":8,"returned":8,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2610.36877","snapshotId":"d312a6cc479a2289a23948f00a6634297c6e7347ee3bb8b080a437198b31dde1","contentSha256":"b438ba27800b6def0cf65e986f5bac1f5002c10139994c208674620856c08119","textHash":"5a6e79d5efabe9e0b2501abb7eda77bac1a63e533c7ac878763c929fde83239a","observedAt":"2026-11-06T00:48:07.137Z","basis":"publisher-page-metadata","sourceLinks":49,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/795c7138f718410cb4c7301206380761833093318252cb8e5cfe4ceff2ed5087/source"}}}],"audits":[{"publicationId":"389b9f8d-2c27-43a2-8d31-432649c1700b","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the work to Bingchao Wang and coauthors, which matches the listed authors. Results are framed as what the authors report, not as established fact.","excerpt":"Authors:Bingchao Wang, Yongcheng Jing, Shunyu Liu, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao"},{"dimension":"summary","status":"uncertain","explanation":"The reading is faithful, but central terms are not explained for a general reader. 'Mobile GUI agent' is never defined (AI that operates phone app screens). JEPA is only called 'the architecture named in the paper's title', which is circular; its full name, Joint-Embedding Predictive Architecture, and its pretraining on unlabelled screenshots are missing. AndroidControl and 'in-distribution' are also unexplained. It also omits the abstract's notable finding that a randomly initialised predictor does as well as a pretrained one. Smallest repair: add one-clause plain definitions of these terms and mention that ablation finding.","excerpt":"a randomly initialised predictor does just as well as one pretrained on screen dynamics"},{"dimension":"date","status":"pass","explanation":"originalPublished 2026-10-27 matches the arXiv v1 submission date, not the later feed date.","excerpt":"[Submitted on 27 Oct 2026]"},{"dimension":"support","status":"pass","explanation":"The speed, accuracy and parameter figures are quoted accurately with their conditions: AndroidControl, same GPU, same inputs, and in-distribution. The 'possible speed and memory advantage' is hedged and consistent with the abstract's memory framing and the 6% parameter figure. The closing question follows from the single-benchmark scope.","excerpt":"On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs, while reaching 86-92% of its in-distribution accuracy with 6% of its parameters."},{"dimension":"media","status":"pass","explanation":"The selected media is the verified arXiv abstract page for this paper, the direct original source.","excerpt":null}],"createdAt":"2026-11-06T00:32:06.935Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}