Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox
Source
huggingface.co
Date
Why it matters
Shows how to reuse an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition →agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → that checks terminal database state as a GRPO reward signal, with reported success gains on held-out tasks and open code.
Key takeaways · AI-distilled
Microsoft and Hugging Face reused ThinkingBox's executable checks of final database state as a binary GRPO reward on tasks outside the benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition →. After 50 steps, Qwen3.8-27B passed 107 of 507 benchmark tasks on all 20 attempts, up from 38, and pass@1 rose from 51.70% to 60.78%.
The gain was repeatability, not breadth: tasks solved at least once fell slightly, from 453 to 446. The authors read this as sharpening, with probability shifting toward trajectories that work every time.
Training tasks were chosen once, offline: each was probed against the base model, scored with a Jeffreys Beta posterior, and dropped if mostly in the rarely solved (below 0.08) or nearly always solved (0.90+) band, since all-fail or all-pass groups give GRPO no contrast.
Pass@1 and repeatability can rank models differently: GPT-5.6 Sol scored 61.91% pass@1 against 60.78% for the fine-tuned Qwen but passed only 82 tasks on all 20 attempts against 107. Claude Opus 5.5 and Opus 5 lead that measure at 241.
By the authors' serving-cost index, cost per dependable task fell from $24.36 to $8.00, near GPT-5.4 ($6.80), GPT-6 Astra ($7.45) and Claude Opus 5.5 ($7.80). The figure excludes training and task-probing costs.
Terms in this piece · Glossary
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
