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

Judgment

Teams picking a post-training recipe must decide whether to spend engineering effort on reward design — executable database checks plus offline policy-relative difficulty selection — rather than on more model scale, because the ThinkingBox gains arrived alongside a 7-task drop in pass@20 coverage.

Basis

The Microsoft and Hugging Face joint post reports Qwen3.8-27B fine-tuned with GRPO for 50 steps on a training pool selected by offline policy-relative difficulty, using executable checks of terminal database state as the reward. Average success rose from 51.70% to 60.78% (5,242 to 6,163 successful attempts across 10,140 attempts on 507 tasks); tasks completed correctly on all twenty attempts rose from 38 (7.50%) to 107 (21.10%); pass@20 fell from 453 (89.35%) to 446 (87.97%). The mechanism matters because difficulty is policy-relative: each task's success probability was estimated once with the base policy and the pool was held fixed for the whole run, so whoever tunes that selection shapes what the model can learn cheaply. On one hotel-booking task, the base model skipped required manager approval in 15 of 20 attempts; the tuned model preserved booking and payment state in all 20.