Research map / Large language models / Instruction and fine-tuning
Refining Hybrid Genetic Search for CVRP via Reinforcement Learning-Finetuned LLM
Rongjie Zhu, Cong Zhang, Zhiguang Cao
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 14B LLM with RL to generate HGS crossover operator, vs expert operator (800 iters) | HGS-PyVRP · CVRPLIB X n in [200,400) | gap %: 1.85 -> 1.67 | Improved |
| RL-tuned operator on larger unseen sizes (800 iters) | HGS-PyVRP · CVRPLIB X n in [800,1000] | gap %: 2.32 -> 2.24 | Improved |
| RL-tuned operator on smallest instances (800 iters) | HGS-PyVRP · CVRPLIB X n in [100,200) | gap %: 0.62 -> 0.70 | Hurt |
| RL-tuned operator at 1000 iterations | HGS-PyVRP · CVRPLIB X n in [600,800) | gap %: 2.43 -> 2.35 | Improved |
| prompt GPT-4o or GPT-o3 to modify the crossover operator instead of RL fine-tuning | HGS-PyVRP · CVRPLIB X | gap %: 0.62 -> 0.62 (GPT-4o_800 n 100-200) | No clear effect |
| use continuous relative-improvement reward instead of discrete +1 reward (ablation) | RFTHGS · CVRPLIB X n in [200,400) | gap %: 1.78 -> 1.67 | Improved |
| use continuous reward instead of discrete reward (ablation) | RFTHGS · CVRPLIB X n in [800,1000] | gap %: 2.30 -> 2.24 | Improved |
| RL fine-tune the 14B model instead of prompting GPT models for compilable operators | operator generation · 16 sampled operators | successful compilation rate: 3/16 -> 16/16 | Improved |
Closest papers by meaning
- Large Language Models as End-to-end Combinatorial Optimization Solvers (NeurIPS 2025 · similarity 0.88)
- Improving Generalization of Neural Combinatorial Optimization for Vehicle Routing Problems via Test-Time Projection Learning (NeurIPS 2025 · similarity 0.88)
- MTL-KD: Multi-Task Learning Via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver (NeurIPS 2025 · similarity 0.88)
- Memory-Enhanced Neural Solvers for Routing Problems (NeurIPS 2025 · similarity 0.87)
- Neural Solver Selection for Combinatorial Optimization (ICML 2025 · similarity 0.86)
- Learning to Insert for Constructive Neural Vehicle Routing Solver (NeurIPS 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.
