Distilling Arena Trajectories Into Explicit Observe-Think-Act Chains Lifts Qwen3-Coder-30B Past Qwen3 Coder Plus
Coaching Qwen3 Coder 30B to Think Like a CodeClash Arena Agent (arXiv 2609.16096, submitted 14 Sep 2026) starts from the observation that Qwen3 Coder Plus ranked last among the 8 commercial agents in the original CodeClash multi-round tournament work, then diagnoses the open-weight 30B version as producing frequent syntax and protocol-breaking errors with weak cross-round strategic adaptation. Vanilla instruction tuning cannot fix this because offline SFT cannot verify whether an action is valid or beneficial. Their ReAct SFT rewrites teacher trajectories into explicit [obs][thought][act] chains, and trajectory-quality weighted SFT reweights samples toward post-edit checking; the fine-tuned open model then outperforms the original Qwen3 Coder Plus in tournament evaluation.
↳ Follow the thread