Research map / Reinforcement learning / RL from feedback
Draft-and-Audit Reinforcement Learning for Optimization Modeling
Zeping Min, Weihang Xu, Zhengzhong You, Wotao Yin, Xinshang Wang
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 the Draft-Audit prompt to the base model without RL (ablation) | Qwen2.5-7B-Instruct · MAMO-Complex / IndustryOR | pass@1 accuracy (%): +2.4/+0.0 pp | No clear effect |
| add an audit prompt to single-turn RL model (ablation) | single-turn RL · MAMO-Complex / IndustryOR | pass@1 accuracy (%): 76.7 -> 79.1 and 20.0 -> 20.0 | No clear effect |
| train two-turn Draft-Audit policy instead of single-turn RL (ablation) | DA-RL · MAMO-Complex / IndustryOR | pass@1 accuracy (%): +7.8/+16.7 pp | Improved |
| train with the audit turn so Turn-1 drafts improve (ablation) | DA-RL Turn-1 draft · MAMO-Complex / IndustryOR | pass@1 accuracy (%): +3.9/+15.0 pp over single-turn RL | Improved |
| double generation length for single-turn RL (ablation) | single-turn RL · MAMO-Complex / IndustryOR | pass@1 accuracy (%): 76.7/20.0 | No clear effect |
| use structured audit rubric instead of generic prompt (ablation) | DA-RL · MAMO-Complex | pass@1 accuracy (%): 79.1 -> 84.5 (-5.4 without) | Improved |
| use structured audit rubric instead of generic prompt (ablation) | DA-RL · IndustryOR | pass@1 accuracy (%): 20.0 -> 36.7 (-16.7 without) | Improved |
| extend rollout from two to three turns (ablation) | DA-RL · MAMO-Complex / IndustryOR | pass@1 accuracy (%): 84.5 -> 74.4 and 36.7 -> 15.0 (-10.1/-21.7) | Hurt |
| add few-shot examples to the prompts (ablation) | DA-RL · MAMO-Complex | pass@1 accuracy (%): 76.7% -> 84.5% | Improved |
| add few-shot examples to the prompts (ablation) | DA-RL · OptMATH-Bench | pass@1 accuracy (%): 11.8% -> 30.4% | Improved |
| add Draft-Audit prompting to frontier LLMs | Gemini-3-Pro · Macro AVG | pass@1 accuracy (%): 8.4 -> 42.1 (+33.7) | Improved |
| add Draft-Audit prompting to frontier LLMs | Claude-Opus-4.5 · Macro AVG | pass@1 accuracy (%): 57.4 -> 67.0 (+9.6) | Improved |
Closest papers by meaning
- Solver-Informed RL: Grounding Large Language Models for Authentic Optimization Modeling (NeurIPS 2025 · similarity 0.89)
- EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling (NeurIPS 2026 · similarity 0.87)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.87)
- $\textbf{Re}^{2}$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving (ICLR 2026 · similarity 0.86)
- $\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis (ICLR 2024 · similarity 0.86)
- Revolve: Optimizing AI Systems by Tracking Response Evolution in Textual Optimization (ICML 2025 · similarity 0.86)
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.
