Database checks become training rewards: 50 GRPO steps lift Qwen3.8-27B's all-20-pass tasks from 38 to 107
Synopsis
Agents often claim a task is done while leaving the wrong database record. ThinkingBox turns deterministic checks of terminal database state into rewards and trains Qwen3.8-27B on tasks outside its benchmark: after 50 GRPO steps, average success rose from 51.70% to 60.78% and tasks passing all twenty attempts rose from 38 to 107, while tasks passing at least once fell from 453 to 446. The training pool was selected offline by policy-relative difficulty, and none of the 507 benchmark tasks were used for training.
Interpretation
The training reward now reads the database state a task actually leaves behind, rather than which tools the agent called or what it said. Earlier grading focused on whether the agent claimed completion, and could miss a case of nine plausible tool calls leaving one wrong database row. On the 507 ThinkingBox-Bench tasks, each task starts a real backend and runs executable checks against the backend record that needs to be correct.
The training pool keeps only tasks the base model sometimes solves, because all-pass and all-fail rollout groups give GRPO no advantage signal. GRPO centers rewards within a group of eight rollouts, so an all-zero or all-one reward vector offers no contrast to learn from. Each candidate task is probed repeatedly against the base policy, its success probability estimated with a Jeffreys posterior Beta(k+½, n−k+½) and bucketed at 0.08, 0.40, 0.60 and 0.90; the final Qwen3.8-27B pool held 187 tasks.
On a benchmark it never trained on, the trained checkpoint gains repeatable success while at-least-once coverage slips slightly. Base Qwen3.8-27B reached 51.70% pass@1, 38 tasks passing all twenty attempts and 453 tasks passing at least once. After 50 GRPO steps, pass@1 was 60.78%, 107 tasks passed all twenty attempts and 446 passed at least once; pass@1 and 20/20 rose in all five domains, while pass@20 rose in neobank and consulting and fell in retail, travel and auto insurance.
Perspective
The demonstrated setting is 50 steps of full-parameter GRPO on Qwen3.8-27B over 187 ThinkingBox-format training tasks, evaluated 20 times on each of the 507 ThinkingBox-Bench tasks excluded from training; the training pool shares the ThinkingBox framework, simulated-user machinery, executable terminal-state grading and a few common tools with the benchmark.
The training pool shares the ThinkingBox framework, simulated-user machinery, executable terminal-state grading and a few common tools with the benchmark, which limits the scope of the transfer claim; re-probing difficulty during training was the path not taken; the dollar figures are a comparative serving index that excludes training-run and policy-probe cost.