ThinkingBox uses database checks as RL rewards

The team reused ThinkingBox's executable checks of final database state as training rewards, building an offline task pool. With 50 GRPO steps on Qwen3.8-27B, average success rose from 51.70% to 60.78%, and tasks solved correctly in all 20 attempts went from 38 to 107.

• Nanbeige releases Nanbeige4.3-33B model2026-10-25
• Five Qwen3.8-27B models match Fable 5 in benchmark tests2026-10-25
• Developer builds Pokemon 2v2 multi-agent LLM test environment2026-10-22
• ISTA-DASLab Uses Prefiller Model to Boost Qwen3.8-27B Low-Quant Accuracy2026-10-20
• Nine models, five coding tasks tested on one 16GB GPU2026-10-18