ResearchHugging Face

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

#ai#database#training#rewards#microsoft

English

Hugging Face and Microsoft introduced ThinkingBox, a framework for evaluating AI agents by comparing their database changes to expected outcomes. The framework enables the use of executable checks as rewards for training, resulting in improved performance of the Qwen3.8-27B model in task completion rates.

中文

Hugging Face 和微软推出了 ThinkingBox，这是一个通过比较 AI 代理的数据库变化与预期结果来评估它们的框架。该框架使可执行检查作为训练奖励的使用成为可能，从而提高了 Qwen3.8-27B 模型的任务完成率。