Research map / Large language models / Instruction and fine-tuning
CALM Before the STORM: Unlocking Native Reasoning for Optimization Modeling
Zhengyang Tang, Zihan Ye, Chenyu Huang, Xuhan Huang, Chengpeng Li, Sihang Li, Guanhua Chen, Ming Yan, Zizhuo Wang, Hongyuan Zha, Dayiheng Liu, Benyou 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 |
|---|---|---|---|
| fine-tune on OR-Instruct-3K full solutions (ablation) | Qwen3-4B-Thinking-2507 · NL4Opt | pass@1: 85.8 -> 92.9 | Improved |
| fine-tune on OR-Instruct-3K full solutions (ablation) | Qwen3-4B-Thinking-2507 · OptMath | pass@1: 33.1 -> 6.6 | Hurt |
| fine-tune on OR-Instruct-3K full solutions (ablation) | Qwen3-4B-Thinking-2507 · IndustryOR | pass@1: 46.2 -> 27.5 | Hurt |
| fine-tune on OR-Instruct-3K full solutions (ablation) | Qwen3-4B-Thinking-2507 · 5-benchmark macro average | pass@1: 57.1 -> 51.2 | Hurt |
| SFT on CALM-corrected solutions instead of OR-Instruct-3K (ablation) | Qwen3-4B-Thinking-2507 · MAMO Complex | pass@1: 40.5 -> 54.3 (base 46.5) | Improved |
| SFT on CALM-corrected solutions (ablation) | Qwen3-4B-Thinking-2507 · 5-benchmark macro average | pass@1: 57.1 -> 58.7 | Improved |
| SFT on CALM-corrected solutions (ablation) | Qwen3-4B-Thinking-2507 · IndustryOR and OptMath | pass@1: 46.2 -> 44.1; 33.1 -> 30.4 | No clear effect |
| add RL (GRPO) after CALM-SFT (ablation) | CALM-SFT model · 5-benchmark macro average | pass@1: 58.7 -> 68.9 | Improved |
| CALM local correction vs DeepSeek-R1 full-solution distillation before RL (ablation) | STORM pipeline · Hard-3 AVG (matched control) | pass@1: 47.4 -> 54.9 | Improved |
| CALM local correction vs Gemini-2.5-Pro full-solution distillation before RL (ablation) | STORM pipeline · Hard-3 AVG (matched control) | pass@1: 30.8 -> 54.9 | Improved |
| DeepSeek-R1 distillation vs CALM on IndustryOR (ablation) | STORM pipeline · IndustryOR (matched control) | pass@1: 50.4 vs 50.0 | No clear effect |
| swap Gemini-2.5-Pro intervener for open-source DeepSeek-R1-0528 (ablation) | CALM · 5-benchmark macro average | pass@1: 68.9 -> 67.8 | No clear effect |
Closest papers by meaning
- SolverLLM: Leveraging Test-Time Scaling for Optimization Problem via LLM-Guided Search (NeurIPS 2025 · similarity 0.90)
- Native Reasoning Models: Training Language Models to Reason on Unverifiable Data (ICLR 2026 · similarity 0.89)
- Step-by-Step Optimization-like Reasoning in LLMs over Expanding Search Spaces (NeurIPS 2026 · similarity 0.89)
- Rectifying LLM Thought from Lens of Optimization (ICLR 2026 · similarity 0.89)
- How Far Are We from Optimal Reasoning Efficiency? (NeurIPS 2025 · similarity 0.88)
- REASONING COMPILER: LLM-Guided Optimizations for Efficient Model Serving (NeurIPS 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.
