Research map / Large language models / Instruction and fine-tuning
AesCoder: Code Aesthetics with Agentic Reward Feedback
Bang Xiao, Lingjie Jiang, Shaohan Huang, Tengchao Lv, Yupan Huang, Xun Wu, Lei Cui, Furu Wei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add agentic reward (execution, static, interactive agents) instead of static-only reward model (ablation) | GRPO-AR Qwen3-4B · OpenDesign | InterAes: 0.71 -> 1.04 | Improved |
| add agentic reward instead of static-only reward (ablation) | GRPO-AR Qwen2.5-Coder-7B · OpenDesign | InterAes: 0.72 -> 0.94 | Improved |
| add agentic reward instead of static-only reward (ablation) | GRPO-AR Qwen3-4B · OpenDesign | Align: 29.16 -> 30.42 | Improved |
| add agentic reward instead of static-only reward (ablation) | GRPO-AR Qwen2.5-Coder-7B · OpenDesign | Align: 28.81 -> 30.03 | Improved |
| use GRPO-AR instead of DPO | Qwen3-4B SFT · OpenDesign | InterAes: 0.70 -> 1.04 | Improved |
| use GRPO-AR instead of RFT | Qwen3-4B SFT · OpenDesign | InterAes: 0.71 -> 1.04 | Improved |
| use GRPO-AR instead of DPO | Qwen2.5-Coder-7B SFT · OpenDesign | Align: 29.75 -> 30.03 | Improved |
| use GRPO-AR instead of RFT | Qwen2.5-Coder-7B SFT · OpenDesign | InterAes: 0.75 -> 0.94 | Improved |
| SFT on AesCode-358K plus GRPO-AR | Qwen3-4B-Instruct-2507 · OpenDesign | Total static: 73.26 -> 81.92 | Improved |
| SFT on AesCode-358K plus GRPO-AR | Qwen2.5-Coder-7B-Instruct · OpenDesign | Total static: 46.27 -> 81.23 | Improved |
| SFT on AesCode-358K plus GRPO-AR | Qwen2.5-Coder-7B-Instruct · PandasPlotBench | error rate: 0.22 -> 0.09 | Improved |
| SFT on AesCode-358K plus GRPO-AR | Qwen3-4B-Instruct-2507 · PandasPlotBench | average score: 65 -> 70 | Improved |
Closest papers by meaning
- SlopCodeBench: Benchmarking How Coding Agents Degrade Over Long-Horizon Iterative Tasks (NeurIPS 2026 · similarity 0.86)
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.86)
- RLEF: Grounding Code LLMs in Execution Feedback with Reinforcement Learning (ICML 2025 · similarity 0.86)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.85)
- Training Language Models to Generate Quality Code with Program Analysis Feedback (NeurIPS 2025 · similarity 0.85)
- ReGAL: Refactoring Programs to Discover Generalizable Abstractions (ICML 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
