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

• URL: http://arxiv.org/abs/2609.16096v1
• Date: Mon, 14 Sep 2026 14:50:05 GMT
• Title: Coaching Qwen3 Coder 30B to Think Like a CodeClash Arena Agent
• Abstract summary: We show that Qwen3-Coder-30B is not well optimized for arena-style interaction.<n>We propose ReAct SFT, which rewrites teacher trajectories into explicit [obs]. chains, and trajectoryquality weighted SFT, which reweights samples to encourage post-edit checking.
• Score: 0.0
• License: http://creativecommons.org/licenses/by/4.0/
• 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 question: what can we do to improve the thinking process of a weak code agent? We study this question in CodeClash, a code-arena benchmark where the original work evaluates 8 commercial coding agents across 6 arenas through multi-round tournaments. Since Qwen3 Coder Plus ranks last among them, we take the open-weight Qwen3-Coder-30B as a case study and investigate how to improve it with distilled knowledge from stronger agents. Our analysis shows that Qwen3-Coder-30B is not well optimized for arena-style interaction: it frequently produces syntax and protocol-breaking errors and exhibits weak strategic adaptation across rounds. These failures are difficult to correct with vanilla instruction tuning alone, since offline SFT cannot directly verify whether a generated action is valid or beneficial. To address this, we propose ReAct SFT, which rewrites teacher trajectories into explicit [obs][thought][act] chains, and trajectoryquality weighted SFT, which reweights samples to encourage post-edit checking. ReAct SFT substantially improves strategic behavior, and our fine-tuned model outperforms the original Qwen3 Coder Plus in tournament evaluation.

Related papers

• TRIM: Reducing AI-Generated CodeSlop via Agent Trajectory Minimization [57.14479960309637]
  Coding agents are increasingly used to accelerate code generation in many downstream tasks.<n>Despite their value, agent-generated code tends to be larger and more verbose than the corresponding human-written implementation.<n>We show that the cause lies in the agent's own search process.
  arXiv  Detail & Related papers  (2026-07-20T17:06:19Z)
• Is Agent Code Less Maintainable Than Human Code? [17.226150835020146]
  We investigate how agent code compares to human code in maintenance settings.<n>We find that agents are less effective at resolving tasks when building on agent code compared to human code.
  arXiv  Detail & Related papers  (2026-06-19T23:56:14Z)
• SmellBench: Towards Fine-Grained Evaluation of Code Agents on Refactoring Tasks [34.43588103259487]
  SmellBench proactively injects code smells into clean code from real-world repositories.<n>It contains 294 cases spanning 7 popular smell types, 3 difficulty levels, 2 instruction settings across 7 real-world repositories.<n>Experiments show that the best combination - Qwenwen Code + Sonnet 4.5 - achieved only a 50.34 score of smell elimination.
  arXiv  Detail & Related papers  (2026-06-04T01:49:07Z)
• FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale [73.57353294235223]
  We introduce FrontierSmith, an automated system for evolving open-ended problems from existing closed-ended coding tasks.<n>On two open-ended coding benchmarks, training on our synthesized data yields substantial gains over the base models.<n>The synthesized problems also make agents take more turns and use more tokens, similar to human-curated ones.
  arXiv  Detail & Related papers  (2026-05-14T06:39:42Z)
• SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks [55.76734816061826]
  We introduce SlopCodeBench, a language-agnostic benchmark comprising 20 problems and 93 checkpoints.<n>We track two trajectory-level quality signals: verbosity, the fraction of redundant or duplicated code, and structural erosion.<n>No agent solves any problem end-to-end across 11 models; the highest checkpoint solve rate is 17.2%.
  arXiv  Detail & Related papers  (2026-03-25T19:26:44Z)
• ReLook: Vision-Grounded RL with a Multimodal LLM Critic for Agentic Web Coding [35.30369059154906]
  ReLook is an agentic, vision-grounded reinforcement learning framework.<n>It empowers an agent to close a robust generate--diagnose--refine loop.<n>It consistently outperforms strong baselines in vision-grounded front-end code generation.
  arXiv  Detail & Related papers  (2025-10-13T15:05:50Z)
• Sifting through the Chaff: On Utilizing Execution Feedback for Ranking the Generated Code Candidates [46.74037090843497]
  Large Language Models (LLMs) are transforming the way developers approach programming by automatically generating code based on natural language descriptions. This paper puts forward RankEF, an innovative approach for code ranking that leverages execution feedback. Experiments on three code generation benchmarks demonstrate that RankEF significantly outperforms the state-of-the-art CodeRanker.
  arXiv  Detail & Related papers  (2024-08-26T01:48:57Z)
• Comments as Natural Logic Pivots: Improve Code Generation via Comment Perspective [85.48043537327258]
  We propose MANGO (comMents As Natural loGic pivOts), including a comment contrastive training strategy and a corresponding logical comment decoding strategy. Results indicate that MANGO significantly improves the code pass rate based on the strong baselines. The robustness of the logical comment decoding strategy is notably higher than the Chain-of-thoughts prompting.
  arXiv  Detail & Related papers  (2024-04-11T08:30:46Z)
• CONCORD: Clone-aware Contrastive Learning for Source Code [64.51161487524436]
  Self-supervised pre-training has gained traction for learning generic code representations valuable for many downstream SE tasks. We argue that it is also essential to factor in how developers code day-to-day for general-purpose representation learning. In particular, we propose CONCORD, a self-supervised, contrastive learning strategy to place benign clones closer in the representation space while moving deviants further apart.
  arXiv  Detail & Related papers  (2023-06-05T20:39:08Z)
• Contrastive Code Representation Learning [95.86686147053958]
  We show that the popular reconstruction-based BERT model is sensitive to source code edits, even when the edits preserve semantics. We propose ContraCode: a contrastive pre-training task that learns code functionality, not form.
  arXiv  Detail & Related papers  (2020-07-09T17:59:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.