Verified, then logged.
Sun 22 Nov 2026 · 00:16 KST / 15:16 UTC한국어 | ENSearch /

Breaking02 / IT

Microsoft posts ThinkingBox-Training agent RL tool on GitHub

Company figures show more tasks passed on every attempt and fewer solved at least once, and the licence name is not confirmed.

An open server chassis packed with unbranded GPU accelerator cards, with the blurred back of an engineer facing a monitor showing a grid of colored pass-and-fail cells behind it. AI-generated image

Microsoft said on 22 Oct 2026 that ThinkingBox-Training, a reinforcement learning tool for AI agents, is available on GitHub, though the name of its licence is not confirmed. The company's training run used 24 H100 GPUs.

Specification. According to the company's blog post, the model being trained is Qwen3.8-27B. The method is GRPO, a form of reinforcement learning that learns from differences in reward between attempts. The run covered all of the model's parameters for 50 steps and used 187 tasks.

ThinkingBox-Bench scores agents on 507 tasks. The company said the deciding test is the final state of the database after the work is finished. This time, that same judgement served as the reward on tasks outside the benchmark.

Results. Microsoft reported that average success rose from 51.70% to 60.78%. Tasks passed on all 20 attempts rose from 38 to 107. Tasks solved at least once fell from 453 to 446.

Microsoft said it did not use benchmark tasks in training. The two task sets do share a framework and some tools, and the company added that this narrows the scope of its claim that the gains transfer.

The company's comparison table lists 107 tasks passed on all 20 attempts for the fine-tuned model. Kimi-K3 is lower at 68, while GPT-5.4 has 128 and Claude Opus 5.5 has 241.

The public documents do not confirm weight distribution or the redistribution and commercial terms.