Research map / Reinforcement learning / RL from feedback
Towards Better Correctness and Efficiency in Code Generation
Yunlong Feng, Yang Xu, Xiao Xu, Binyuan Hui, Junyang Lin
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 |
|---|---|---|---|
| fine-tune with SFT on the fastest correct code only (ablation) | Qwen2.5-Coder-Instruct-7B · Mercury | Beyond: 74.80 -> 66.36 | Hurt |
| fine-tune with SFT on the fastest correct code only (ablation) | Qwen2.5-Coder-Instruct-7B · Mercury | Pass@1: 77.20 -> 83.20 | Improved |
| use DPO instead of SFT | Qwen2.5-Coder-Instruct-7B · EvalPerf | DPSnorm: 77.85 -> 87.61 | Improved |
| raise proportion of efficiency pairs in DPO data from 0.0 to 0.5 (ablation) | DPO on Qwen2.5-Coder-Instruct-7B · Mercury | Pass@1: 87.11 -> 85.47 | Hurt |
| use only efficiency pairs (proportion 1.0) in DPO data (ablation) | DPO on Qwen2.5-Coder-Instruct-7B · Mercury | Pass@1: 85.47 -> 77.30 | Hurt |
| use high-complexity synthetic inputs for efficiency rewards in online RL (ablation) | GRPO on Qwen2.5-Coder-Instruct-7B · EvalPerf | DPSnorm: 81.94 -> 83.96 | Improved |
| use high-complexity synthetic inputs for efficiency rewards in online RL (ablation) | GRPO on Qwen2.5-Coder-Instruct-7B · Mercury | Pass@1: 90.70 -> 87.50 | Hurt |
| use RLOO instead of GRPO for online RL | Qwen2.5-Coder-Instruct-7B · Mercury | Pass@1: 87.50 -> 89.92 | Improved |
| use RLOO instead of GRPO for online RL | Qwen2.5-Coder-Instruct-7B · Mercury | Beyond: 79.15 -> 76.00 | Hurt |
| initialize online RL from the base model instead of the instruct model (ablation) | online RL · EvalPerf | DPSnorm: 83.96 -> 90.81 | Improved |
| initialize online RL from DPO model with 0.1 efficiency pairs instead of the instruct model (ablation) | online RL · Mercury | Pass@1: 87.50 -> 88.83 | Improved |
| add RLOO efficiency stage after DPO stage (two-stage tuning) (ablation) | Qwen2.5-Coder-Instruct-7B · Mercury | Beyond: 76.66 -> 79.04 | Improved |
Closest papers by meaning
- Afterburner: Reinforcement Learning Facilitates Self-Improving Code Efficiency Optimization (NeurIPS 2025 · similarity 0.92)
- EffiCoder: Enhancing Code Generation in Large Language Models through Efficiency-Aware Fine-tuning (ICML 2025 · similarity 0.89)
- How efficient is LLM-generated code? A rigorous & high-standard benchmark (ICLR 2025 · similarity 0.88)
- CodeScaler: Scaling Code LLM Training and Test-Time Inference via Reward Models (NeurIPS 2026 · similarity 0.88)
- Reinforcement Learning for Code Optimization (NeurIPS 2026 · similarity 0.88)
- Teaching Language Models to Critique via Reinforcement Learning (ICML 2025 · similarity 0.88)
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.
