Microsoft has published ThinkingBox-Training on GitHub, a follow-up to its ThinkingBox benchmark. The benchmark grades an agent by comparing the final database state with the expected outcome. The new release reuses that kind of executable check as a reward signal for post-training, on tasks outside the benchmark.
What was measured
Microsoft fine-tuned Qwen3.8-27B with GRPO for 50 steps, then evaluated it on the 507-task ThinkingBox-Bench, which it had not trained on. Each task was run 20 times.
- Average success (pass@1) rose from 51.70% to 60.78%.
- Tasks completed on all 20 attempts rose from 38 (7.50%) to 107 (21.10%).
- Tasks solved at least once (pass@20) fell from 453 (89.35%) to 446 (87.97%).
Microsoft reads this as sharpening: the model became more repeatable, not broader. Gains in pass@1 and all-20 success appeared in all five benchmark domains. Pass@20 was mixed, with auto insurance showing the largest loss at 7.0 points.
One example: on a hotel-booking task that required manager approval before a downgrade, the base model changed the booking without approval in 15 of 20 attempts. The tuned model preserved the booking and payment state in all 20.
Cost per dependable task
Using the same estimated-cost method as its first post, Microsoft puts the cost per task passed on every attempt at $8.00 for the tuned model, against $24.36 for the base model. It lists GPT-5.4 at $6.80, GPT-6 Astra at $7.45 and Claude Opus 5.5 at $7.80. These figures are a comparative serving index, not cloud bills, and they exclude the cost of the training run and of the probing used to pick training tasks.
How the training works
Most rewards are deterministic checks on terminal database state: the reward is 1 or 0, and the path the agent took does not matter. A minority of narrow requirements use an LLM judge to answer a binary rubric question. Training tasks were filtered offline: each candidate was probed against the base model, and tasks it always failed or always passed were dropped, because GRPO learns only from groups with mixed outcomes. The final pool held 187 tasks.
Limits
The benchmark and the training pool share the ThinkingBox framework, simulated-user setup and grading style, which Microsoft says limits the scope of the transfer claim. The 187-task pool is private; the repository ships a public sample list that only exercises the interface. The reference run used 24 H100 GPUs across three nodes, on Linux with Python 3.12 and Verl 0.9.0.
The benchmark side is also useful on its own. Claude Opus 5.5 scored 0.66 points higher than Claude Opus 5 on pass@1 yet passed the same 241 tasks on all 20 attempts, so a higher headline score did not mean more dependable behaviour.