Microsoft and Hugging Face Use Database End-State Checks as Agent RL Rewards

Microsoft and Hugging Face describe ThinkingBox-Training, which uses executable checks of final database state as RL rewards for post-training agents. Tasks outside the ThinkingBox-Bench benchmark were selected using offline policy-relative difficulty estimated from the base policy. After 50 GRPO steps, Qwen3.8-27B's average pass@1 rose from 51.70% to 60.78%, and tasks passed on all 20 attempts rose from 38 to 107. Pass@20 declined slightly, from 89.35% to 87.97%, indicating a shift toward repeatable success rather than broader coverage.

Key takeaways

• Qwen3.8-27B average pass@1 rose from 51.70% to 60.78% after 50 GRPO steps.
• Tasks passed on all 20 attempts rose from 38 to 107 on the benchmark.
• Pass@20 coverage fell slightly, from 453 to 446 tasks solved at least once.
• Training pool tasks were selected using success probability estimated once from the base policy.

Why it matters

The work shows that executable database-state checks can serve as training rewards, and that reliability metrics like all-attempts success can diverge from average accuracy. This matters for deploying agents that modify production records.

Read with care

Gains are reported for one model (Qwen3.8-27B) and one benchmark; source claims not independently checked.