[Paper Review] DIFUSCO: Graph-based Diffusion Solvers for Combinatorial Optimization
DIFUSCO introduces graph-based discrete diffusion models to solve NP-hard combinatorial optimization problems, achieving state-of-the-art results on TSP and MIS by parallel denoising and multimodal solution generation.
Neural network-based Combinatorial Optimization (CO) methods have shown promising results in solving various NP-complete (NPC) problems without relying on hand-crafted domain knowledge. This paper broadens the current scope of neural solvers for NPC problems by introducing a new graph-based diffusion framework, namely DIFUSCO. Our framework casts NPC problems as discrete {0, 1}-vector optimization problems and leverages graph-based denoising diffusion models to generate high-quality solutions. We investigate two types of diffusion models with Gaussian and Bernoulli noise, respectively, and devise an effective inference schedule to enhance the solution quality. We evaluate our methods on two well-studied NPC combinatorial optimization problems: Traveling Salesman Problem (TSP) and Maximal Independent Set (MIS). Experimental results show that DIFUSCO strongly outperforms the previous state-of-the-art neural solvers, improving the performance gap between ground-truth and neural solvers from 1.76% to 0.46% on TSP-500, from 2.46% to 1.17% on TSP-1000, and from 3.19% to 2.58% on TSP10000. For the MIS problem, DIFUSCO outperforms the previous state-of-the-art neural solver on the challenging SATLIB benchmark.
Motivation & Objective
- Broaden neural CO solutions beyond autoregressive and RL-based methods by introducing a graph-based diffusion framework.
- Cast NPC problems as {0,1}-valued vectors and use diffusion denoising on graphs to generate high-quality solutions.
- Compare Gaussian (continuous) vs Bernoulli (discrete) diffusion and identify effective inference schedules.
- Demonstrate scalability and strong performance on large TSP instances and MIS benchmarks.
Proposed method
- Formulate NPC problems as {0,1}-vector optimization over graph-structured instances.
- Use a graph neural network (Anisotropic GNN) to encode instances and denoise corrupted {0,1} variables in a diffusion framework.
- Explore discrete diffusion with Bernoulli noise and continuous diffusion with Gaussian noise; train with supervised denoising (not RL).
- Design fast inference via DDIM-like schedules (linear and cosine) for discrete diffusion to reduce steps while preserving quality.
- Decode solutions with heatmaps and problem-specific strategies (greedy with 2-opt for TSP; greedy for MIS) and perform multiple samples.

Experimental results
Research questions
- RQ1Can graph-based diffusion models generate high-quality, multimodal solutions for NPC problems?
- RQ2How do discrete Bernoulli diffusion and continuous Gaussian diffusion compare on graph-based CO tasks?
- RQ3What inference schedule and decoding strategy yield best performance and efficiency for TSP and MIS?
- RQ4Do DIFUSCO models trained on one scale generalize to larger TSP instances or different MIS benchmarks?
Key findings
- Discrete diffusion substantially outperforms continuous diffusion for NPC problems when using more than a few diffusion steps.
- Cosine denoising schedule superior to linear for discrete diffusion; enables fast inference with good quality.
- For TSP, DIFUSCO closes the gap to ground-truth from 1.76% to 0.46% (with MCTS) on TSP-500, from 2.46% to 1.17% (TSP-1000), and from 3.19% to 2.58% (TSP-10000).
- For MIS on SATLIB benchmark, DIFUSCO achieves state-of-the-art performance compared to previous neural solvers.
- A single Anisotropic GNN backbone can solve both TSP and MIS within the same framework.
- DIFUSCO generalizes well across TSP scales, often outperforming RL- or SL-trained non-autoregressive methods in cross-scale tests.

Read this paper
Open the full text in the Nubint reader and ask about any passage as you read.
Read this paper on NubintThis review was generated by AI from the original paper. Please check the paper before citing it.