acceptodds
Withdrawn from ICLR 2027

Rotational Equivariance Improves Neural Routing Solvers

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Although rotational-equivariant neural networks have proven to be an effective architectural bias across domains from chemistry to robotics, such approaches remain largely under-explored in the design of neural solvers for combinatorial optimization. In this work, we focus on Euclidean routing problems—specifically, the traveling salesman problem and the capacitated vehicle routing problem. For these problems, we experimentally demonstrate that converting existing neural solvers into their rotational-equivariant versions can boost their performance and generalization capabilities.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
equivariant neural network, neural combinatorial optimization, traveling salesman problem, vehicle routing problem

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_oCRNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether rotational equivariance can improve neural solvers for Euclidean TSP and CVRP. It presents an equivariant network component and adapts POMO and DRHG to Eq-POMO and Eq-DRHG, covering both construction- and improvement-based models. Experiments across instance sizes, distributions, TSPLIB, and CVRPLIB report improvements over the original solvers in many settings.

Critical strengths
  1. The paper translates SO(2)-equivariant components into concrete neural routing solvers for both POMO and DRHG. The symmetry motivation is well aligned with Euclidean instances.
  2. This evaluation is comprehensive, covering problem scale, synthesis distributions, and TSPLIB/CVRPLIB.
Critical weaknesses
  1. The comparisons do not sufficiently isolate the contribution of equivariance. Eq-POMO also changes preprocessing and distance bias, while Eq-DRHG changes normalization, capacity masks, and optimizer-update schedules. The fiber and bias ablations are useful, but do not separate these effects from the contribution of equivariance.
  2. The Eq-DRHG requires more inference time than DRHG. Without matched wall-clock comparisons and clearly specified repair budgets, it is unclear whether they offer a better quality–time trade-off.
  3. The larger cross-size gains in Table 1 are concentrated on POMO, which generalizes poorly in these settings, while improvements over DRHG are often marginal, and performance degrades in some settings. The evidence that the adaptation provides reliable additional gains on constructive backbones with stronger generalization remains limited.

As an additional suggestion that does not affect my assessment, an overview diagram highlighting the architectural changes from POMO and DRHG to Eq-POMO and Eq-DRHG would make the method easier to understand.

Questions for authors
  1. How do the methods compare under matched inference-time budgets for DRHG?
  2. Could you evaluate the equivariant versions on representative constructive backbones designed for generalization, such as ReLD, ICAM, and LEHD? Given that Eq-POMO already incorporates ICAM’s distance-bias design, it is helpful for clarifying whether equivariance provides consistent additional gains on these backbones.
Reviewer_dLATOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper incorporates the rotation-equivalent mechanism into neural combinatorial optimization. It can be incorporated off-the-shelf neural combinatorial solver such as POMO. Numerical experiments show that proposed method improves solution quality.

Critical strengths

Proposed method is a meta algorithm so that it can incorporated existing neural combinatorial solver.

Critical weaknesses

This paper is not ready for submitting to the conference. For example, abstract is not polished so that proposed method does not explain briefly. Numerical experiments of DIFUSCO and FAST T2T are not fully reported. (in Table1)

The comparison of Sym-NCO and INViT which consider symmetry in routing (Section2) are missing.

Computational overhead is large of proposed method. Table2 show that Eq-DRHG needs more than times computational time of vanilla DRHG.

Questions for authors

Is it possible to compare proposed method to Sym-NCO or INViT? These method considers symmetric of routing.

Is it possible to incorporate proposed method into LEHD? Table1 reported that LEHD works better than Eq-DRHG and Eq-POMO.

Is there any limitation to incorporate proposed method to existing solver?

Reviewer_qKStOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Using equivariant building blocks explored in previous work, the authors adapt the two neural combinatorial optimization solvers POMO and DRHG to be equivariant for solving TSP (traveling salesman problem) and CVRP (capacitated vehicle routing problem).

More precisely, the paper collects (Section 3.3) various tools from the paper by Fuchs et al. 2020 on equivariant attention and the papers by Weiler & Cesa 2019 and by Cesa et al. 2022 on equivariance under Euclidean transformations for CNNs. They are applied to replace the respective steps in the setup of POMO (Section 4.2) and DRHG (Section 4.3). This is all with a focus on TSP (traveling salesman problem) and CVRP (capacitated vehicle routing problem) leading to further adaptations of the embedding (Section 4.1). The closest precursor of this work is the SE(3)-Transformer by Fuchs et al. 2020, and Section 4.4 lists the technical differences in how equivariance is achieved.

Overall, the results are not theoretical but Section 5 gives an overview of the performance in experiments, comparing with some heuristic solvers and several recent neural solvers. The experiments are guided by (1) comparison with existing neural solvers, (2) generalization capabilities and (3) the transfer to real-world data.

Critical strengths
  • The experiments are well-structured and the questions addressed in the evaluation are interesting. The experiments (Table 1,2,3) suggest the derived equivariant solvers have higher accuracy than existing neural solvers in various well-chosen settings.
Critical weaknesses
  1. The setup seems rather restrictive: existing building blocks were applied to make specific neural solvers equivariant for two specific classes of combinatorial optimization problems. Therefore, I have doubts about the wider impact of this work.

  2. Looking at Table 1, I don't see much improvement in the CVRP benchmark results, in contrast to what you write in the paragraph 'Effectiveness of equivariant model' on page 8. Even worse, it seems that Table 2 indicates that the equivariant networks may produce only slightly better results but need significantly more time.

Questions for authors
  1. Why did you focus on POMO / DRHG applied to TSP / CVRP? In the current form, the progress derived in this paper seems rather incremental, since the equivariant building blocks were already there.

  2. As far as I see, you did not give the running times for the real-world data, in particular comparing POMO and eq-POMO as well as DRHG and eq-DRHG. Given that the running time comparisons in Table 2 were not very favourable for the equivariant networks, this would be particularly interesting.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.