{"article":{"annotation":{"article_id":"programgym-training-environments-coding-agents-arxiv-2026","available_at":"2026-10-13T10:04:24.468307Z","geography_revision":"ef6ee0010dda5fe7b240fd514749b43e214c57da73222d2e40ec83b276dd0977","id":"90fd79e95a79674f5a1d5ae3e36207bfd899f28e081df7153e4e768a6a91076e","input_sha256":"b6969cece27c6c966d46aa500bf5ccf604be54b42b2e59f52276d027ec1a58cf","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/2610.13395v1"]}],"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/2610.13395v1"]}],"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/programgym-training-environments-coding-agents-arxiv-2026?lang=en","category":"technology","content":{"key_points":[],"paragraphs":["Researchers from Xiamen University and other institutions introduced ProgramGym, a framework for training coding agents to inspect a program’s behavior and reproduce it from scratch. In the preprint submitted to arXiv on October 9, Program-Env creates training environments with executable reference programs, isolated sandboxes, usage documentation and black-box tests. The authors report building 1,648 environments in C, C++, Go and Rust. Program-Agent then separates analysis, implementation and review into distinct roles, with compilation checks and recovery checkpoints between iterations.","On ProgramBench, the authors report that fine-tuning Qwen3.8-27B on trajectories from the generated environments raised its test-pass result from 34.87% to 42.84%. With Program-Agent, the base model scored 55.71% and the fine-tuned version 56.25%. The paper defines test pass as a macro-average of task-level pass rates, so these figures do not mean every task was solved. The results concern the authors’ program-reproduction tasks and stated tool setup; they are not a measure of general coding ability."],"sections":[]},"corrections":[],"country":"CN","data":{"blocks":[{"columns":[{"key":"measure","label":{"ar":"القياس","de":"Messgröße","en":"Measure","es":"Medición","fr":"Mesure","pt":"Medição","tr":"Ölçüm"},"language":"en","type":"text"},{"key":"value","label":{"ar":"القيمة","de":"Wert","en":"Value","es":"Valor","fr":"Valeur","pt":"Valor","tr":"Değer"},"type":"number","unit":"count"}],"description":{"ar":"بيئات إعادة إنتاج البرامج التي أبلغت عنها الورقة.","de":"In der Studie genannte Umgebungen zur Programmreproduktion.","en":"Program-reproduction environments reported in the paper.","es":"Entornos de reproducción de programas comunicados en el artículo.","fr":"Environnements de reproduction de programmes rapportés dans l’article.","pt":"Ambientes de reprodução de programas informados no artigo.","tr":"Makalenin bildirdiği program yeniden üretme ortamları."},"id":"training-environment-count","methodology":{"ar":"يحصي المؤلفون البيئات المنشأة للغات C وC++ وGo وRust.","de":"Die Autoren zählen Umgebungen für C, C++, Go und Rust.","en":"Authors count environments built for C, C++, Go and Rust.","es":"Los autores cuentan los entornos creados para C, C++, Go y Rust.","fr":"Les auteurs comptent les environnements créés pour C, C++, Go et Rust.","pt":"Os autores contam ambientes criados para C, C++, Go e Rust.","tr":"Yazarlar C, C++, Go ve Rust dillerinde oluşturulmuş ortamları sayıyor."},"rowTranslations":[{"measure":{"ar":"بيئات تدريب Program-Env","de":"Program-Env-Trainingsumgebungen","en":"Program-Env training environments","es":"Entornos de entrenamiento de Program-Env","fr":"Environnements d’entraînement Program-Env","pt":"Ambientes de treinamento do Program-Env","tr":"Program-Env eğitim ortamları"}}],"rows":[{"measure":"Program-Env training environments","value":1648}],"sourceUrls":["https://arxiv.org/html/2610.13395v1"],"title":{"ar":"عدد بيئات التدريب في ProgramGym","de":"Anzahl der ProgramGym-Trainingsumgebungen","en":"ProgramGym training environment count","es":"Número de entornos de entrenamiento de ProgramGym","fr":"Nombre d’environnements d’entraînement ProgramGym","pt":"Número de ambientes de treinamento do ProgramGym","tr":"ProgramGym eğitim ortamı sayısı"},"type":"table","updatedAt":"2026-10-13T09:31:52Z"},{"columns":[{"key":"configuration","label":{"ar":"النموذج وإطار التشغيل","de":"Modell und Ausführungsrahmen","en":"Model and harness","es":"Modelo y configuración","fr":"Modèle et cadre","pt":"Modelo e estrutura","tr":"Model ve düzenek"},"language":"en","type":"text"},{"key":"test_pass","label":{"ar":"اجتياز الاختبارات (%)","de":"Tests bestanden (%)","en":"Tests Passed (%)","es":"Pruebas superadas (%)","fr":"Tests réussis (%)","pt":"Testes aprovados (%)","tr":"Test geçme (%)"},"type":"number","unit":"percent"}],"description":{"ar":"معدلات اجتياز الاختبارات لتركيبات النموذج وإطار الوكيل التي أبلغ عنها المؤلفون.","de":"Berichtete Test-Pass-Raten für die Modell- und Agentengerüst-Kombinationen der Autoren.","en":"Test-pass rates reported for the authors’ model and agent-scaffold pairings.","es":"Tasas de pruebas superadas comunicadas para las combinaciones de modelo y estructura de agente.","fr":"Taux de réussite rapportés pour les associations modèle et cadre d’agent des auteurs.","pt":"Taxas de aprovação em testes informadas para as combinações de modelo e estrutura de agente dos autores.","tr":"Yazarların model ve ajan-iskeleti eşleştirmeleri için bildirdiği test-geçme oranları."},"id":"programbench-test-pass","methodology":{"ar":"يعرّف ProgramBench هذا المقياس بأنه المتوسط الكلي لمعدلات اجتياز الاختبارات لكل مهمة. تثبت مقارنة Program-Env استخدام Claude Code؛ وتقارن نتيجة Program-Agent النموذج نفسه مع Claude Code أو Program-Agent. ولا يعني ذلك حل المهام بالكامل.","de":"ProgramBench definiert diesen Wert als Makromittel der Test-Pass-Raten pro Aufgabe. Beim Program-Env-Vergleich bleibt Claude Code gleich; beim Program-Agent-Vergleich wird dasselbe Modell mit Claude Code oder Program-Agent kombiniert. Die Aufgaben gelten damit nicht als vollständig gelöst.","en":"ProgramBench defines this metric as the macro-average of task-level test-pass rates. The Program-Env comparison holds Claude Code fixed; the Program-Agent comparison pairs the same model with Claude Code or Program-Agent. The tasks are not thereby fully solved.","es":"ProgramBench define esta métrica como el promedio macro de las tasas por tarea. La comparación de Program-Env mantiene Claude Code fijo; la de Program-Agent compara el mismo modelo con Claude Code o Program-Agent. Esto no significa que las tareas estén resueltas por completo.","fr":"ProgramBench définit cette mesure comme la moyenne macro des taux de réussite par tâche. La comparaison Program-Env maintient Claude Code fixe; celle de Program-Agent compare le même modèle avec Claude Code ou Program-Agent. Cela ne signifie pas que les tâches sont entièrement résolues.","pt":"O ProgramBench define essa métrica como a média macro das taxas de aprovação por tarefa. A comparação do Program-Env mantém o Claude Code fixo; a comparação do Program-Agent combina o mesmo modelo com Claude Code ou Program-Agent. Isso não significa que as tarefas tenham sido totalmente resolvidas.","tr":"ProgramBench ölçütü görev başına test-geçme oranlarının makro ortalamasıdır. Program-Env için Claude Code sabit tutulur; Program-Agent karşılaştırmasında aynı modelin Claude Code ve Program-Agent eşleştirmeleri yer alır. Bu görevler tümden çözülmüş sayılmaz."},"rowTranslations":[{"configuration":{"ar":"Qwen3.8-27B مع Claude Code، قبل الضبط الدقيق على مسارات Program-Env","de":"Qwen3.8-27B + Claude Code, vor dem Fine-Tuning mit Program-Env-Verläufen","en":"Qwen3.8-27B + Claude Code, before fine-tuning on Program-Env trajectories","es":"Qwen3.8-27B + Claude Code, antes del ajuste fino con trayectorias de Program-Env","fr":"Qwen3.8-27B + Claude Code, avant réglage fin sur les trajectoires Program-Env","pt":"Qwen3.8-27B + Claude Code, antes do ajuste fino com trajetórias do Program-Env","tr":"Qwen3.8-27B + Claude Code, sem ajuste fino nas trajetórias do Program-Env"}},{"configuration":{"ar":"Qwen3.8-27B-SFT مع Claude Code، بعد الضبط الدقيق على مسارات Program-Env","de":"Qwen3.8-27B-SFT + Claude Code, nach dem Fine-Tuning mit Program-Env-Verläufen","en":"Qwen3.8-27B-SFT + Claude Code, after fine-tuning on Program-Env trajectories","es":"Qwen3.8-27B-SFT + Claude Code, tras el ajuste fino con trayectorias de Program-Env","fr":"Qwen3.8-27B-SFT + Claude Code, après réglage fin sur les trajectoires Program-Env","pt":"Qwen3.8-27B-SFT + Claude Code, após ajuste fino com trajetórias do Program-Env","tr":"Qwen3.8-27B-SFT + Claude Code, após ajuste fino nas trajetórias do Program-Env"}},{"configuration":{"ar":"Qwen3.8-27B مع Program-Agent، النموذج الأساسي نفسه","de":"Qwen3.8-27B + Program-Agent, dasselbe Basismodell","en":"Qwen3.8-27B + Program-Agent, same base model","es":"Qwen3.8-27B + Program-Agent, mismo modelo base","fr":"Qwen3.8-27B + Program-Agent, même modèle de base","pt":"Qwen3.8-27B + Program-Agent, mesmo modelo-base","tr":"Qwen3.8-27B + Program-Agent, aynı temel model"}},{"configuration":{"ar":"Qwen3.8-27B-SFT مع Program-Agent، النموذج المضبوط نفسه","de":"Qwen3.8-27B-SFT + Program-Agent, dasselbe feinabgestimmte Modell","en":"Qwen3.8-27B-SFT + Program-Agent, same fine-tuned model","es":"Qwen3.8-27B-SFT + Program-Agent, mismo modelo ajustado","fr":"Qwen3.8-27B-SFT + Program-Agent, même modèle réglé","pt":"Qwen3.8-27B-SFT + Program-Agent, mesmo modelo ajustado","tr":"Qwen3.8-27B-SFT + Program-Agent, aynı ince ayarlı model"}}],"rows":[{"configuration":"Qwen3.8-27B + Claude Code, before fine-tuning on Program-Env trajectories","test_pass":34.87},{"configuration":"Qwen3.8-27B-SFT + Claude Code, after fine-tuning on Program-Env trajectories","test_pass":42.84},{"configuration":"Qwen3.8-27B + Program-Agent, same base model","test_pass":55.71},{"configuration":"Qwen3.8-27B-SFT + Program-Agent, same fine-tuned model","test_pass":56.25}],"sourceUrls":["https://arxiv.org/html/2610.13395v1"],"title":{"ar":"o","de":"a","en":"r","es":"r","fr":"g","pt":"m","tr":"P"},"type":"table","updatedAt":"2026-10-13T09:31:52Z"}],"version":1},"data_downloads":[{"formats":{"csv":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/training-environment-count?lang=en&format=csv","json":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/training-environment-count?lang=en&format=json"},"id":"training-environment-count"},{"formats":{"csv":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/programbench-test-pass?lang=en&format=csv","json":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026/data/programbench-test-pass?lang=en&format=json"},"id":"programbench-test-pass"}],"data_types":["table"],"editorial_illustration":{"alt":"Conceptual illustration of technology, processors and connected information.","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":"en","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-f09e42b4-97e6-5bc9-b5a7-3b0c4c035839","event_period":"2026-10-09","event_period_basis":"event-or-reference-period","id":"programgym-training-environments-coding-agents-arxiv-2026","image":null,"importance":0.6,"language":"en","modified_at":"2026-10-13T10:04:10+00:00","publication_basis":"publication-record","published_at":"2026-10-13T10:04:10+00:00","reading_minutes":1,"region":"world","section_illustration":{"alt":"Conceptual illustration of technology, processors and connected information.","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":"en","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":"programgym-training-environments-coding-agents-arxiv-2026","sources":[{"checkedAt":"2026-10-13","name":"arXiv preprint by researchers affiliated with Xiamen University, Washington University in St. Louis, the Chinese Academy of Sciences and Unisound","publishedAt":"2026-10-09","url":"https://arxiv.org/html/2610.13395v1"}],"summary":"The October 9 arXiv preprint describes environments for reproducing programs from scratch and a staged agent workflow.","text":"The October 9 arXiv preprint describes environments for reproducing programs from scratch and a staged agent workflow.\n\nResearchers from Xiamen University and other institutions introduced ProgramGym, a framework for training coding agents to inspect a program’s behavior and reproduce it from scratch. In the preprint submitted to arXiv on October 9, Program-Env creates training environments with executable reference programs, isolated sandboxes, usage documentation and black-box tests. The authors report building 1,648 environments in C, C++, Go and Rust. Program-Agent then separates analysis, implementation and review into distinct roles, with compilation checks and recovery checkpoints between iterations.\n\nOn ProgramBench, the authors report that fine-tuning Qwen3.8-27B on trajectories from the generated environments raised its test-pass result from 34.87% to 42.84%. With Program-Agent, the base model scored 55.71% and the fine-tuned version 56.25%. The paper defines test pass as a macro-average of task-level pass rates, so these figures do not mean every task was solved. The results concern the authors’ program-reproduction tasks and stated tool setup; they are not a measure of general coding ability.","title":"ProgramGym builds 1,648 verifiable training environments for coding agents","topics":[],"translations":[{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=tr","language":"tr","url":"https://newsverum.com/tr/news/programgym-training-environments-coding-agents-arxiv-2026"},{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=en","language":"en","url":"https://newsverum.com/en/news/programgym-training-environments-coding-agents-arxiv-2026"},{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=ar","language":"ar","url":"https://newsverum.com/ar/news/programgym-training-environments-coding-agents-arxiv-2026"},{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=fr","language":"fr","url":"https://newsverum.com/fr/news/programgym-training-environments-coding-agents-arxiv-2026"},{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=es","language":"es","url":"https://newsverum.com/es/news/programgym-training-environments-coding-agents-arxiv-2026"},{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=de","language":"de","url":"https://newsverum.com/de/news/programgym-training-environments-coding-agents-arxiv-2026"},{"api_url":"https://newsverum.com/api/v1/articles/programgym-training-environments-coding-agents-arxiv-2026?lang=pt","language":"pt","url":"https://newsverum.com/pt/news/programgym-training-environments-coding-agents-arxiv-2026"}],"url":"https://newsverum.com/en/news/programgym-training-environments-coding-agents-arxiv-2026"},"version":"1"}