Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
- ID
- 38498
- Status
- summarized
- Published
- 23 Oct 2026, 6:24 AM
- Fetched
- 23 Oct 2026, 7:40 AM
- Provider
- Hugging Face Blog
- Category
- developer-ai
- Original URL
- https://huggingface.co/blog/microsoft/thinkingbox-training
- Source URL
- https://huggingface.co/blog/feed.xml
Summary
- Score
- 7.5
- Created
- 23 Oct 2026, 7:40 AM
- Tags
- Audience
- ai_ml_learnersai_agent_usersdevelopersdatabase_learners
What happened
Microsoft and Hugging Face describe ThinkingBox, a system that grades an agent by comparing the final database state against the expected outcome, across 507 tasks in ThinkingBox-Bench. They reused those 1/0 executable checks as GRPO reward signals on a separate training pool, filtered by 'offline policy-relative difficulty' — estimating each task's success probability with the base policy once, then keeping the pool fixed. After 50 GRPO steps on Qwen3.8-27B, average success across 10,140 attempts rose from 51.70% to 60.78%, and tasks passing all 20 attempts rose from 38 to 107; ThinkingBox-Training is now on GitHub.
Why it matters
If your agents touch a stateful backend, your existing DB assertions are already a reward function — no LLM judge needed. The actionable part is the filtering rule: drop tasks the base model always solves or never solves, because both teach nothing. They measured base-model success once per task and froze the pool for the whole run, which is cheap enough to try on your own test suite before committing to a training run.
Discussion angle
Take a handful of your own integration tests that check final DB state and estimate the base model's pass rate on each — how many land in the useful middle band, and what does that say about whether post-training is worth it versus just fixing your prompts and tool schemas?