# Microsoft — Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox

- Company: Microsoft (microsoft.com)
- Announced: 2026-10-22T22:24:34+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/microsoft/thinkingbox-training
- Record: https://forck.live/items/26898-turning-database-grading-into-rl-rewards-post-training-agents-with-thinkingbox
- Models affected: Qwen3.8-27B, Kimi-K3, Claude Opus 5, Claude Opus 5.5

Microsoft and Hugging Face are using ThinkingBox's executable database checks as reinforcement-learning rewards to post-train agents, and report that Qwen3.8-27B's average success rose from 51.70% to 60.78% after 50 GRPO steps. ThinkingBox was built to catch agents that claim they are done when the database says otherwise, and its 507-task benchmark hands an agent a real backend, tools and a simulated user before checking the terminal database state. The authors say most tasks are either already solved or hopeless for a given model, so they estimated each task's success probability once with the base policy and kept that training pool fixed, an approach they call offline policy-relative difficulty. On the sealed benchmark, tasks completed correctly on all twenty attempts rose from 38 to 107, while at-least-once coverage declined slightly. In one hotel-booking task, the base model changed a booking without required manager approval in 15 of 20 attempts, while the fine-tuned model preserved booking and payment state in all 20. The authors argue pass@1 should not be read alone for production database work, citing Claude Opus 5.5 scoring 0.66 percentage points higher than Claude Opus 5 on pass@1 yet passing the same 241 tasks on all twenty attempts. ThinkingBox-Training is now available on GitHub.

## Evidence

Verbatim from https://huggingface.co/blog/microsoft/thinkingbox-training:

> For Qwen3.8-27B, average success rose from 51.70% to 60.78%, while tasks completed correctly on all twenty attempts rose from 38 to 107 with 50 GRPO steps.

---

Record: https://forck.live/items/26898-turning-database-grading-into-rl-rewards-post-training-agents-with-thinkingbox
Catalogue: https://forck.live/llms.txt
Current issue: https://forck.live/feed.md