{"slug": "microsoft-trains-qwen3-8-27b-on-187-tasks-to-improve-agent-repeatability", "title": "Microsoft trains Qwen3.8-27B on 187 tasks to improve agent repeatability", "summary": "Microsoft and Hugging Face published an open-source training system on October 22 that rewards an AI agent for the database state it leaves behind, and fine-tuning Qwen3.8-27B with executable checks raised average success by 9.08 percentage points across a 507-task evaluation, according to Microsoft applied scientist Tuhin Kundu and collaborators. The experiment trained the model on 187 tasks, and the model solved at least one attempt on seven fewer tasks. Microsoft's results show stronger repeatability on its benchmark.", "body_md": "# Microsoft trains Qwen3.8-27B on 187 tasks to improve agent repeatability\n\n**The Microsoft and Hugging Face experiment raised average success by 9.08 percentage points across 507 benchmark tasks, while the model solved at least one attempt on seven fewer tasks.**\n\n        By [Ryan Merket](https://runtimewire.com/author/ryan-merket)\n        · Published \n        · Updated \n\nPrimary source: [Hugging Face Newsroom](https://huggingface.co/blog/microsoft/thinkingbox-training)\n\n## Why it matters\n\nExecutable checks connect agent training to the actual records business systems retain. Microsoft's results show stronger repeatability on its benchmark.\n\nOn October 22nd, Microsoft and [Hugging Face](https://huggingface.co/microsoft?ref=runtimewire) published an open-source training system that rewards an AI agent for the database state it leaves behind. In the [release results](https://huggingface.co/blog/microsoft/thinkingbox-training?ref=runtimewire), Microsoft applied scientist [Tuhin Kundu](https://www.linkedin.com/in/tuhinkundu?ref=runtimewire) and collaborators report that fine-tuning [Qwen3.8-27B](https://runtimewire.com/models/qwen/qwen3.8-27b) with executable checks raised average success by 9.08 percentage points across a 507-task evaluation.…\n\nWhen you purchase through links in our articles, [we may earn a small commission](https://runtimewire.com/affiliate-monetization-standards). This doesn’t affect our editorial independence.", "url": "https://wpnews.pro/news/microsoft-trains-qwen3-8-27b-on-187-tasks-to-improve-agent-repeatability", "canonical_source": "https://runtimewire.com/article/microsoft-thinkingbox-training-database-state-rewards", "published_at": "2026-10-23 02:32:11+00:00", "updated_at": "2026-10-23 05:40:08.881148+00:00", "lang": "en", "topics": ["ai-agents", "ai-research", "large-language-models", "machine-learning", "ai-tools"], "entities": ["Microsoft", "Hugging Face", "Qwen3.8-27B", "Tuhin Kundu"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/microsoft-trains-qwen3-8-27b-on-187-tasks-to-improve-agent-repeatability", "markdown": "https://wpnews.pro/news/microsoft-trains-qwen3-8-27b-on-187-tasks-to-improve-agent-repeatability.md", "text": "https://wpnews.pro/news/microsoft-trains-qwen3-8-27b-on-187-tasks-to-improve-agent-repeatability.txt", "jsonld": "https://wpnews.pro/news/microsoft-trains-qwen3-8-27b-on-187-tasks-to-improve-agent-repeatability.jsonld"}}