Comment on paper: Position: Rethinking Post-Hoc Search-Based Neural Approaches for Solving Large-Scale Traveling Salesman Problems

TL;DR

This paper re-examines the SoftDist heat map-based approach to solving large-scale TSP by highlighting evaluation flaws. It shows that uneven hardware usage and inconsistent timing definitions undermine the claimed advantages of SoftDist, and that when all steps run on the same hardware the soft dist claim does not hold. Re-analysis with UTSP and data-driven unsupervised neural heat-map methods demonstrates competitive or superior performance within similar time budgets on instances such as TSP-1000. The work underscores the need for fair, hardware-consistent benchmarking and suggests that unsupervised learning approaches hold strong potential for large-scale TSP problems.

Abstract

We identify two major issues in the SoftDist paper (Xia et al.): (1) the failure to run all steps of different baselines on the same hardware environment, and (2) the use of inconsistent time measurements when comparing to other baselines. These issues lead to flawed conclusions. When all steps are executed in the same hardware environment, the primary claim made in SoftDist is no longer supported.

Figures (1)

• Figure 1: The overall scheme of heat map-based search. In the SoftDist paper, when comparing the running time of different models, the authors calculate parts (a) and (c) for SoftDist, while the UTSP running time consists of parts (a), (b), and (c), leading to an unfair comparison.