Intel / article

Coaching Qwen3 Coder 30B to Think Like a CodeClash Arena Agent

Source: arxiv.org
Author: Ivy Ning Zhang
Date: Published Sep 16, 2026

Context

A coding in a multi-round arena must learn from what happened after each edit. This study trained Qwen3-Coder-30B, an model from Qwen, using stronger agents' examples rewritten into explicit observation, reasoning and action sequences. It also explored weighting training examples to encourage checking after edits. The authors report improved strategic behavior and tournament performance above Qwen3 Coder Plus. The result concerns arena interaction, where repeated diagnosis and adaptation matter; the abstract provides no basis for extending that ranking to ordinary software development.

Key takeaways · AI-distilled

In the CodeClash , where the original study evaluated 8 commercial coding agents across 6 arenas in multi-round tournaments, Qwen3 Coder Plus ranked last; the authors then used the separate open-weight Qwen3-Coder-30B model as their case study for improvement.
The authors found syntax errors, broken interaction protocols, and weak adaptation in Qwen3-Coder-30B. They argue ordinary supervised struggles to fix these because offline examples cannot verify whether an action is valid or useful.
The authors' ReAct SFT technique rewrites teacher-agent trajectories into explicit observation, thought, and action chains to teach the model a clearer reasoning structure before acting.
A second technique, trajectory-quality weighted SFT, reweights training samples to push the model toward checking its own work after edits, and the resulting fine-tuned model beat the original Qwen3 Coder Plus in tournament evaluation.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.

Read the source arxiv.org