Research map / Graph learning / Graph algorithms and clustering
Problem Distributions as Tasks: Repurposing Meta Learning for Generative Combinatorial Optimization towards Multi-task Pretraining and Adaptation
Wenzheng Pan, Jiale Ma, Nuoyan Chen, Yang Li, Junchi Yan
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 OOD fine-tuning on few-shot support data (ablation) | M2GenCO · MIS-BA 200-300 | objective: 65.45 -> 70.01 | Improved |
| add OOD fine-tuning on few-shot support data (ablation) | M2GenCO · MCut-WS 200-300 | objective: 795.63 -> 839.47 | Improved |
| add OOD fine-tuning on few-shot support data (ablation) | M2GenCO · ATSP-SAT-54 | objective: 5.30 -> 2.31 | Improved |
| add few-shot fine-tuning on an unseen task (MVC) (ablation) | M2GenCO · MVC-RB-LARGE | optimality gap: 1.16% -> 0.94% | Improved |
| add gradient normalization across task batches (ablation) | M2GenCO meta-pretraining · MIS RB-200-300 | objective: 16.678 -> 16.732 | Improved |
| add gradient normalization across task batches (ablation) | M2GenCO meta-pretraining · MCut BA-200-300 | objective: 635.296 -> 657.190 | Improved |
| use 5 inner steps instead of 1 (ablation) | M2GenCO meta-pretraining · MCut BA-200-300 | objective: 657.190 -> 632.530 | Hurt |
| use 10 inner steps instead of 1 (ablation) | M2GenCO meta-pretraining · MCut BA-200-300 | objective: 657.190 -> 572.754 | Hurt |
| add MCl to the task pool (MIS, MCut) (ablation) | M2GenCO meta-pretraining · MCl RB-200-300 | objective: 10.30 -> 11.37 | Improved |
| add MIS to the task pool (MCl, MCut) (ablation) | M2GenCO meta-pretraining · MIS RB-200-300 | objective: 14.84 -> 16.33 | Improved |
| add MCl to the task pool (MIS, MCut) (ablation) | M2GenCO meta-pretraining · MCut BA-200-300 | objective: 631.13 -> 621.82 | No clear effect |
| tune on 10% flipped labels instead of best-quality labels (ablation) | M2GenCO fine-tuning · MIS BA-200-300 | objective: 69.76 -> 69.53 | No clear effect |
Closest papers by meaning
- GOAL: A Generalist Combinatorial Optimization Agent Learner (ICLR 2025 · similarity 0.89)
- Unsupervised Learning for Combinatorial Optimization Needs Meta Learning (ICLR 2023 · similarity 0.89)
- T2T: From Distribution Learning in Training to Gradient Search in Testing for Combinatorial Optimization (NeurIPS 2023 · similarity 0.89)
- Can Computational Reducibility Lead to Transferable Models for Graph Combinatorial Optimization? (ICML 2026 · similarity 0.88)
- What’s Wrong with Deep Learning in Tree Search for Combinatorial Optimization (ICLR 2022 · similarity 0.88)
- Fast T2T: Optimization Consistency Speeds Up Diffusion-Based Training-to-Testing Solving for Combinatorial Optimization (NeurIPS 2024 · 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.
