{"article":{"annotation":{"article_id":"rollout-time-intervention-improves-agent-training-on-hard-tasks","available_at":"2026-10-13T00:52:24.215093Z","geography_revision":"ef6ee0010dda5fe7b240fd514749b43e214c57da73222d2e40ec83b276dd0977","id":"aaee7a574d4aaf39621352ec2c27e4f6a5fdab7c538b6a7bdb4138ade3e12d65","input_sha256":"2c2d29c858240978a5ad57477490a68536cd4e4cec375dc00d9f3764115b839f","limitations":["Rule-derived subjects are candidates, not verified editorial labels or calibrated model probabilities.","Coverage is distinct from event location. Unknown country/city/coordinates remain absent; no automatic geocoding.","Article geometries keep their stated precision and methodology; reproducing them is not independent coordinate verification."],"method":"controlled-rules-v1","places":[{"admin1":null,"city":null,"coordinate_note":null,"coordinate_reference_system":"OGC:CRS84","coordinate_source":null,"country":"CN","evidence":[{"field":"country","language":"en","note":"Explicit preserved editorial country/region; no city or national centroid inferred","source_urls":["https://arxiv.org/html/2610.13438v1"]}],"geometry":null,"id":"iso3166-1:CN","kind":"country-or-area","labels":{"ar":"الصين","en":"China","es":"China","fr":"Chine"},"match_status":"candidate","precision":"regional","role":"coverage","uncertainty_m":null}],"rules_sha256":"fb5ad3c27c91cb840f7a0a05f7f64a5cb26ad0374409ff9457f277c7092e22b4","schema_version":1,"subjects":[{"concept_id":"medtop:13000000","confidence":null,"confidence_method":null,"evidence":[{"field":"category+topics","language":"en","note":"Preserved editorial category/topic; broad candidate only","source_urls":["https://arxiv.org/html/2610.13438v1"]}],"method":"controlled-rules-v1","rule_id":"technology-broad","status":"candidate"}],"taxonomy_revision":"1f2db31d811fca07e99b16d402c10e2cc2965c9df1c7b41d9be348b40a24e7fb","taxonomy_version":"IPTC Media Topics 2026-Q2"},"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=fr","category":"technology","content":{"key_points":[],"paragraphs":["Des chercheurs affiliés à Alibaba et à la Hong Kong University of Science and Technology ont présenté Rollout-Time Intervention (RTI). Pendant l’entraînement, RTI applique des corrections multi-agents aux tentatives échouées à l’aide de grilles propres à chaque tâche, puis filtre les trajectoires d’intervention trop éloignées de la distribution du modèle et pondère les données conservées.","Sur le test hors distribution Claw-Eval-Hard20, l’article rapporte un taux Pass@1 de 62,92 % pour le modèle entraîné avec RTI. Les familles de tâches de cette évaluation n’ont pas servi à la génération des données d’entraînement, mais l’expérience reste limitée à des tâches créées par les chercheurs et à des simulations d’API; l’article donne aussi un résultat dans la distribution lié aux mêmes 20 familles de tâches. Ces résultats n’établissent pas la performance d’agents généralistes."],"sections":[]},"corrections":[],"country":"CN","data":{"blocks":[{"columns":[{"key":"metric","label":{"ar":"المقياس","de":"Messgröße","en":"Measure","es":"Medida","fr":"Mesure","pt":"Medida","tr":"Ölçüt"},"language":"en","type":"text"},{"key":"value","label":{"ar":"النسبة المئوية","de":"Prozent","en":"Percent","es":"Porcentaje","fr":"Pourcentage","pt":"Percentual","tr":"Yüzde"},"type":"number","unit":"percent"},{"key":"scope","label":{"ar":"النطاق","de":"Umfang","en":"Scope","es":"Alcance","fr":"Périmètre","pt":"Escopo","tr":"Kapsam"},"language":"en","type":"text"}],"description":{"ar":"Claw-Eval-Hard20؛ عائلات مهام مستبعدة من توليد التدريب؛ Pass@1","de":"Claw-Eval-Hard20; Aufgabenfamilien nicht in der Trainingssynthese; Pass@1","en":"Claw-Eval-Hard20; task families excluded from training synthesis; Pass@1","es":"Claw-Eval-Hard20; familias excluidas de la síntesis de entrenamiento; Pass@1","fr":"Claw-Eval-Hard20; familles de tâches exclues de la synthèse d’entraînement; Pass@1","pt":"Claw-Eval-Hard20; famílias de tarefas excluídas da síntese de treinamento; Pass@1","tr":"Claw-Eval-Hard20; eğitim sentezinde kullanılmayan görev aileleri; Pass@1"},"id":"reported-result-1","methodology":{"ar":"تقتصر القيمة على تجربة المعيار المذكورة في الورقة؛ وليست نتيجة ميدانية أو ضماناً عاماً للأداء.","de":"Der Wert ist auf das im Artikel beschriebene Benchmark-Experiment beschränkt; er ist kein Feldergebnis und keine allgemeine Leistungsgarantie.","en":"The value is limited to the benchmark experiment reported in the paper; it is not a field result or general performance guarantee.","es":"El valor se limita al experimento de benchmark descrito en el artículo; no es un resultado de campo ni una garantía general.","fr":"La valeur se limite à l’expérience de benchmark décrite dans l’article; ce n’est ni un résultat de terrain ni une garantie générale.","pt":"O valor se limita ao experimento de benchmark descrito no artigo; não é um resultado de campo nem uma garantia geral.","tr":"Değer, makalenin belirttiği benchmark deneyiyle sınırlıdır; saha sonucu veya genel performans garantisi değildir."},"rowTranslations":[{"metric":{"ar":"Pass@1 لنموذج RTI في مهام Claw-Eval-Hard20 خارج التوزيع","de":"Pass@1 des RTI-Modells bei Claw-Eval-Hard20 außerhalb der Verteilung","en":"RTI model Pass@1 on Claw-Eval-Hard20 out-of-distribution tasks","es":"Pass@1 del modelo RTI en tareas Claw-Eval-Hard20 fuera de distribución","fr":"Pass@1 du modèle RTI sur Claw-Eval-Hard20 hors distribution","pt":"Pass@1 do modelo RTI em tarefas Claw-Eval-Hard20 fora da distribuição","tr":"RTI modeli Claw-Eval-Hard20 dış-dağılım Pass@1"},"scope":{"ar":"Claw-Eval-Hard20؛ عائلات مهام مستبعدة من توليد التدريب؛ Pass@1","de":"Claw-Eval-Hard20; Aufgabenfamilien nicht in der Trainingssynthese; Pass@1","en":"Claw-Eval-Hard20; task families excluded from training synthesis; Pass@1","es":"Claw-Eval-Hard20; familias excluidas de la síntesis de entrenamiento; Pass@1","fr":"Claw-Eval-Hard20; familles de tâches exclues de la synthèse d’entraînement; Pass@1","pt":"Claw-Eval-Hard20; famílias de tarefas excluídas da síntese de treinamento; Pass@1","tr":"Claw-Eval-Hard20; eğitim sentezinde kullanılmayan görev aileleri; Pass@1"}}],"rows":[{"metric":"RTI model Pass@1 on Claw-Eval-Hard20 out-of-distribution tasks","scope":"Claw-Eval-Hard20; task families excluded from training synthesis; Pass@1","value":62.92}],"sourceUrls":["https://arxiv.org/html/2610.13438v1"],"title":{"ar":"نتيجة البحث","de":"Forschungsergebnis","en":"Research result","es":"Resultado del estudio","fr":"Résultat de l’étude","pt":"Resultado da pesquisa","tr":"Araştırma sonucu"},"type":"table","updatedAt":"2026-10-13T00:17:05Z"}],"version":1},"data_downloads":[{"formats":{"csv":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks/data/reported-result-1?lang=fr&format=csv","json":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks/data/reported-result-1?lang=fr&format=json"},"id":"reported-result-1"}],"data_types":["table"],"editorial_illustration":{"alt":"Illustration conceptuelle de la technologie, des processeurs et des informations connectées.","asset_id":"newsverum-v1-technology","avif_variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.avif","width":480},{"height":480,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-720.avif","width":720},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.avif","width":960},{"height":800,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1200.avif","width":1200},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.avif","width":1536}],"credit":"Newsverum","documentary":false,"height":1024,"kind":"symbolic-editorial-illustration","language":"fr","scope":"topic","url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.webp","width":480},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.webp","width":960},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","width":1536}],"width":1536},"event_id":"event-4e5b7ba7-4702-5f02-a35b-ad64fbf3aee0","event_period":"2026-10-09","event_period_basis":"event-or-reference-period","id":"rollout-time-intervention-improves-agent-training-on-hard-tasks","image":null,"importance":0.6,"language":"fr","modified_at":"2026-10-13T00:51:41+00:00","publication_basis":"publication-record","published_at":"2026-10-13T00:51:41+00:00","reading_minutes":1,"region":"world","section_illustration":{"alt":"Illustration conceptuelle de la technologie, des processeurs et des informations connectées.","asset_id":"newsverum-v1-technology","avif_variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.avif","width":480},{"height":480,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-720.avif","width":720},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.avif","width":960},{"height":800,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1200.avif","width":1200},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.avif","width":1536}],"credit":"Newsverum","documentary":false,"height":1024,"kind":"symbolic-editorial-illustration","language":"fr","scope":"section","url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","variants":[{"height":320,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-480.webp","width":480},{"height":640,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-960.webp","width":960},{"height":1024,"url":"https://newsverum.com/assets/newsverum/illustrations/v1/technology-1536.webp","width":1536}],"width":1536},"slug":"rollout-time-intervention-improves-agent-training-on-hard-tasks","sources":[{"checkedAt":"2026-10-13","name":"arXiv preprint by Alibaba and Hong Kong University of Science and Technology researchers","publishedAt":"2026-10-09","url":"https://arxiv.org/html/2610.13438v1"}],"summary":"Le cadre ajoute une correction multi-agent aux trajectoires d’entraînement et filtre les décalages de distribution.","text":"Le cadre ajoute une correction multi-agent aux trajectoires d’entraînement et filtre les décalages de distribution.\n\nDes chercheurs affiliés à Alibaba et à la Hong Kong University of Science and Technology ont présenté Rollout-Time Intervention (RTI). Pendant l’entraînement, RTI applique des corrections multi-agents aux tentatives échouées à l’aide de grilles propres à chaque tâche, puis filtre les trajectoires d’intervention trop éloignées de la distribution du modèle et pondère les données conservées.\n\nSur le test hors distribution Claw-Eval-Hard20, l’article rapporte un taux Pass@1 de 62,92 % pour le modèle entraîné avec RTI. Les familles de tâches de cette évaluation n’ont pas servi à la génération des données d’entraînement, mais l’expérience reste limitée à des tâches créées par les chercheurs et à des simulations d’API; l’article donne aussi un résultat dans la distribution lié aux mêmes 20 familles de tâches. Ces résultats n’établissent pas la performance d’agents généralistes.","title":"Rollout-Time Intervention teste l’entraînement d’agents sur de nouvelles tâches","topics":[],"translations":[{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=tr","language":"tr","url":"https://newsverum.com/tr/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=en","language":"en","url":"https://newsverum.com/en/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=ar","language":"ar","url":"https://newsverum.com/ar/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=fr","language":"fr","url":"https://newsverum.com/fr/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=es","language":"es","url":"https://newsverum.com/es/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=de","language":"de","url":"https://newsverum.com/de/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},{"api_url":"https://newsverum.com/api/v1/articles/rollout-time-intervention-improves-agent-training-on-hard-tasks?lang=pt","language":"pt","url":"https://newsverum.com/pt/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"}],"url":"https://newsverum.com/fr/news/rollout-time-intervention-improves-agent-training-on-hard-tasks"},"version":"1"}