Research map / Reinforcement learning / Visual and representation RL
Chain-of-Context Learning: Dynamic Constraint Understanding for Multi-Task VRPs
Shuangchun Gui, Suyu Liu, Xuehe Wang, 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 |
|---|---|---|---|
| add CCL (RGCR + TSNR) on top of CaDA | CaDA · in-distribution VRPTW N=50 | gap to HGS-PyVRP: 1.745% -> 0.979% | Improved |
| add CCL on top of CaDA on N=100 CVRP | CaDA · CVRP N=100 | gap to HGS-PyVRP: 1.412% -> 1.287% | Improved |
| add ReLD context module to CCL | CCL · in-distribution CVRP N=50 | gap: 0.977% -> 0.881% | Improved |
| add TSNR module (ablation) | CCL† · 8 non-TW variants N=50 | gap: 2.005% -> 1.512% | Improved |
| add TSNR module (ablation) | CCL† · 8 TW variants N=50 | gap: 1.189% -> 0.689% | Improved |
| add RGCR module (ablation) | CCL† · 8 non-TW variants N=50 | gap: 1.568% -> 1.512% | Improved |
| add RGCR module on simple-constraint tasks (ablation) | CCL† · CVRP and VRPL N=50 | gap: 0.881% -> 0.874% w/o RGCR on CVRP; 1.027% -> 0.993% on VRPL | Hurt |
| add RGCR and TSNR together versus removing both (ablation) | CCL† · 8 non-TW variants N=50 | gap: 2.070% -> 1.512% | Improved |
| use ReLD on top of CCL for OOD N=100 TW tasks | CCL · unseen TW tasks N=100 | gap: 31.159% -> 34.163% | Hurt |
| use CCL† versus CaDA† on OOD N=50 non-TW tasks | CaDA† · unseen non-TW tasks N=50 | gap: 34.169% -> 25.781% | Improved |
| replace single-step refinement with per-route naive refinement (ASW-TAM) (ablation) | CCL-style refinement · 16 variants N=50 | gap: 1.10 -> 29.93 | Hurt |
| increase network scale of CaDA† (heavy decoder) instead of adding CCL (ablation) | CaDA† · 16 variants N=50 | gap: 1.63 -> 1.53 (CCL 1.28) | Mixed |
Closest papers by meaning
- MTL-KD: Multi-Task Learning Via Knowledge Distillation for Generalizable Neural Vehicle Routing Solver (NeurIPS 2025 · similarity 0.88)
- Multi-Task Vehicle Routing Solver via Mixture of Specialized Experts under State-Decomposable MDP (NeurIPS 2025 · similarity 0.88)
- SHIELD: Multi-task Multi-distribution Vehicle Routing Solver with Sparsity and Hierarchy (ICML 2025 · similarity 0.86)
- Learning to Handle Complex Constraints for Vehicle Routing Problems (NeurIPS 2024 · similarity 0.86)
- Dynamic Latent Routing (NeurIPS 2026 · similarity 0.85)
- Mixture-of-Experts Meets In-Context Reinforcement Learning (NeurIPS 2025 · 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.
