Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Using terminal database-state verification as a binary reward signal in GRPO post-training increased Qwen3.8-27B's average agent success from 51.70% to 60.78%, while nearly tripling the number of tasks solved with 100% consistency from 38 to 107. For teams running agents in production, this demonstrates that replacing brittle LLM-as-a-judge evaluators with deterministic database assertions during RL training systematically eliminates "silent failures" where agents falsely report task completion. This methodology and its training stack are now open-source, enabling you to target post-training specifically on tasks within your model's current difficulty threshold.
Conversion of database state checks into RL rewards increased Qwen3.8-27B's average task success rate by 9 percentage points (51.7%→60.78%) and more than doubled the number of tasks solved perfectly on all 20 attempts (38→107). This matters because it enables using existing production database validation logic as automatic reward signals for RL fine-tuning, directly aligning agent behavior with backend correctness without manual reward engineering—critical for agents that interact with real systems where "looks correct" isn't enough.