用 ReAct SFT 教练 Qwen3 Coder 30B 成为竞技场智能体
原标题：Coaching Qwen3 Coder 30B to Think Like a CodeClash Arena Agent
AI 摘要
研究者在 CodeClash 代码竞技场基准上以开源模型 Qwen3-Coder-30B 为案例，探索如何用更强智能体的知识蒸馏来提升弱代码智能体的长程交互能力。他们发现该模型常出现语法与协议错误、缺乏跨轮次策略调整，于是提出 ReAct SFT（将教师轨迹改写为 [obs][thought][act] 链）和轨迹质量加权 SFT（TQ-SFT，鼓励修改后检查）。实验显示 ReAct SFT 显著改善策略行为，微调后模型在锦标赛评估中超过原始 Qwen3 Coder Plus。
正文节选
Coaching Qwen3 Coder 30B to Think Like a CodeClash Arena Agent Abstract Large language model coding agents have recently become useful for software tasks, but weaker or open-weight agents still struggle to reliably interpret user intent and execute complex multi-step workflows. This gap is especially visible in long-horizon settings, where an agent must repeatedly inspect prior outcomes, diagnose failure, and choose the next code edit under interaction constraints. It motivates a natural questio