pith. sign in

arxiv: 2606.01084 · v1 · pith:WZFCQRL6new · submitted 2026-05-31 · 💻 cs.LG · cs.AI

MViewRouter: Internalizing Geometric Equivariance via Multi-view Alternating Attention for Combinatorial Routing

Shiyan Liu , Bohan Tan , Yaoxin Wu , Yan Jin This is my paper

Pith reviewed 2026-06-28 17:45 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords combinatorial routingtraveling salesman problemcapacitated vehicle routinggeometric equivariancemulti-view attentiondeep reinforcement learningzero-shot generalization
0
0 comments X

The pith

MViewRouter internalizes geometric equivariance via multi-view alternating attention to produce invariant routing decisions

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents MViewRouter as a way to handle geometric symmetries in combinatorial routing problems such as TSP and CVRP without depending on data augmentation. Instead of post-hoc fixes that can lead to inconsistent outputs, the model builds equivariance directly into its structure through a Multi-view Alternating Attention mechanism. This mechanism runs parallel computations over the four rotations and reflections of the D4 group, switching between modeling relations inside each view and aligning information across views. A Collective Policy Gradient Aggregation step then pools gradients from the symmetric views to train a more stable policy. The result is claimed to be decisions that stay the same under geometric transformations while still delivering competitive solution quality and better generalization to new instances.

Core claim

MViewRouter achieves invariant decision-making for routing problems by internalizing geometric equivariance as a structural inductive bias through its Multi-view Alternating Attention mechanism that enables parallel processing over the D4 symmetry group, combined with Collective Policy Gradient Aggregation for training.

What carries the argument

Multi-view Alternating Attention (MAA) mechanism that alternates between intra-view relational modeling and inter-view feature alignment for parallel processing over the D4 symmetry group

If this is right

  • Decisions remain consistent across all D4 transformations of the same routing instance without requiring data augmentation.
  • Training converges faster and more stably because gradients are aggregated from multiple symmetric views.
  • Zero-shot performance improves on unseen TSP and CVRP instances, including real-world TSPLIB benchmarks.
  • The same framework applies to both TSP and CVRP while maintaining competitive solution quality.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same multi-view structure could be tested on other geometric combinatorial problems such as the knapsack or scheduling tasks that also admit rotations and reflections.
  • If the invariance holds, the method would reduce the volume of augmented training data needed for routing models in practice.
  • Extending the symmetry group beyond D4 might reveal whether the approach scales to problems with continuous rotations or 3D symmetries.

Load-bearing premise

That parallel processing over the D4 symmetry group via alternating intra- and inter-view attention is sufficient to internalize geometric equivariance as a structural inductive bias that produces invariant decisions for routing problems.

What would settle it

A direct test in which MViewRouter outputs different routes for an instance and its 90-degree rotation or reflection would show that the claimed invariance does not hold.

Figures

Figures reproduced from arXiv: 2606.01084 by Bohan Tan, Shiyan Liu, Yan Jin, Yaoxin Wu.

Figure 1
Figure 1. Figure 1: Overview of the MViewRouter architecture. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training curves on TSP-100: POMO and MViewRouter under original view and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Zero-shot generalization gap relative to POMO. TSP. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-view tour lengths across the D4 symmetry group on TSP-100. We apply all 8 D4 transformations to 10,000 un￾seen TSP-100 instances and measure variance across orientations as a proxy for internalized geometric invariance. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of view quantity K ∈ {1, 2, 4, 8} on TSP-100. Left: solution quality; center: alignment gap A; right: cosine similarity C. Solution quality ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on TSP-100. Left: solution quality; center: alignment gap [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Training curves on CVRP-100: POMO and MViewRouter under original view and [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
read the original abstract

Combinatorial routing problems such as the Traveling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP) are fundamental NP-hard problems with broad real-world applications. While recent deep reinforcement learning methods have shown promising performance, they typically handle geometric symmetries only through data augmentation, resulting in inconsistent decisions and limited generalization. To address this issue, we propose MViewRouter, a multi-view framework that internalizes geometric equivariance as a structural inductive bias to achieve invariant decision-making across routing problem variants. Our approach introduces a Multi-view Alternating Attention (MAA) mechanism that enables parallel processing over the $D_4$ symmetry group, alternating between intra-view relational modeling and inter-view feature alignment. Furthermore, we optimize the policy via Collective Policy Gradient Aggregation (CPGA), leveraging consensus gradients from multiple symmetric views to stabilize training and accelerate convergence. Experiments on TSP and CVRP benchmarks, as well as real-world TSPLIB instances, demonstrate that MViewRouter achieves competitive solution quality and strong zero-shot generalization.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes MViewRouter, a multi-view deep RL framework for combinatorial routing problems (TSP, CVRP). It introduces Multi-view Alternating Attention (MAA) to process inputs in parallel over the D4 symmetry group by alternating intra-view relational modeling and inter-view feature alignment, thereby internalizing geometric equivariance as a structural inductive bias for invariant decisions. A Collective Policy Gradient Aggregation (CPGA) procedure aggregates gradients across symmetric views for training. Experiments on standard benchmarks and TSPLIB instances are reported to yield competitive solution quality and strong zero-shot generalization without data augmentation.

Significance. If the central invariance claim holds with supporting derivations or metrics, the work would demonstrate a structural alternative to data augmentation for symmetry handling in routing RL, potentially improving generalization. The framing as an inductive bias rather than post-hoc augmentation is a constructive direction for combinatorial optimization.

major comments (2)
  1. [Abstract / MAA description] Abstract and method description: the claim that MAA 'internalizes geometric equivariance' and produces 'invariant decision-making' is load-bearing for the contribution, yet no explicit construction is supplied showing how attention weights or embeddings transform under the generators of D4 (90° rotations, reflections). No invariance metric (e.g., output consistency under group actions on the same instance) or proof sketch is referenced.
  2. [Experiments] Experiments section: the reported 'competitive solution quality and strong zero-shot generalization' cannot be evaluated without ablation tables isolating the contribution of MAA versus CPGA, without error bars or statistical tests across random seeds, and without baselines that also incorporate D4 augmentation or equivariant layers.
minor comments (2)
  1. Notation: ensure the precise definition of the D4 group action on node coordinates and edge features is stated before the MAA equations.
  2. Clarify whether CPGA reduces to standard policy gradient when the number of views is one, to separate the symmetry-handling benefit from the aggregation technique.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of the invariance properties and the experimental evaluation.

read point-by-point responses

  1. Referee: [Abstract / MAA description] Abstract and method description: the claim that MAA 'internalizes geometric equivariance' and produces 'invariant decision-making' is load-bearing for the contribution, yet no explicit construction is supplied showing how attention weights or embeddings transform under the generators of D4 (90° rotations, reflections). No invariance metric (e.g., output consistency under group actions on the same instance) or proof sketch is referenced.

    Authors: We agree that the manuscript would be strengthened by an explicit construction. In the revision we will add a subsection deriving the transformation rules for embeddings and attention weights under the generators of D4, together with a quantitative invariance metric that measures output consistency on transformed instances. revision: yes

  2. Referee: [Experiments] Experiments section: the reported 'competitive solution quality and strong zero-shot generalization' cannot be evaluated without ablation tables isolating the contribution of MAA versus CPGA, without error bars or statistical tests across random seeds, and without baselines that also incorporate D4 augmentation or equivariant layers.

    Authors: We concur that these elements are required for a complete evaluation. The revised manuscript will include ablation tables separating MAA and CPGA, report means and standard deviations with statistical tests over multiple seeds, and add baselines that apply D4 data augmentation as well as equivariant layers. revision: yes

Circularity Check

0 steps flagged

No circularity: architectural proposal with external empirical claims

full rationale

The abstract and description frame MViewRouter as introducing a new MAA mechanism for parallel D4 processing and CPGA for gradient aggregation, presented as a structural inductive bias whose performance is evaluated on TSP/CVRP benchmarks and TSPLIB instances. No quoted equations, self-citations, or derivation steps reduce any central claim (equivariant decisions or zero-shot generalization) to fitted parameters or prior outputs by construction. The inductive-bias argument is an architectural hypothesis whose validity is left to external verification rather than internal self-definition or renaming.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

Abstract-only review limits visibility into free parameters or additional axioms; the core modeling choice rests on the domain assumption that D4 symmetries suffice.

axioms (1)
  • domain assumption The D4 symmetry group captures the relevant geometric symmetries for TSP and CVRP instances.
    Paper states it enables parallel processing over the D4 symmetry group to achieve invariant decision-making.
invented entities (2)
  • Multi-view Alternating Attention (MAA) no independent evidence
    purpose: Mechanism to internalize geometric equivariance
    New attention block introduced in the paper; no independent evidence supplied beyond claimed experiments.
  • Collective Policy Gradient Aggregation (CPGA) no independent evidence
    purpose: Stabilize training via consensus gradients from symmetric views
    New training procedure introduced; no independent evidence supplied.

pith-pipeline@v0.9.1-grok · 5717 in / 1398 out tokens · 29119 ms · 2026-06-28T17:45:36.843707+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 7 canonical work pages · 4 internal anchors

  1. [1]

    Concorde tsp solver, 2006

    David Applegate, Ribert Bixby, Vasek Chvatal, and William Cook. Concorde tsp solver, 2006

    2006

  2. [2]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vision transformer. InProceedings of the IEEE/CVF international conference on computer vision, pages 6836–6846, 2021

    2021

  3. [3]

    Some applications of the generalized vehicle routing problem.Journal of the operational research society, 61(7):1072–1077, 2010

    Roberto Baldacci, Enrico Bartolini, and Gilbert Laporte. Some applications of the generalized vehicle routing problem.Journal of the operational research society, 61(7):1072–1077, 2010

    2010

  4. [4]

    Neural Combinatorial Optimization with Reinforcement Learning

    Irwan Bello, Hieu Pham, Quoc V Le, Mohammad Norouzi, and Samy Bengio. Neural combina- torial optimization with reinforcement learning.arXiv preprint arXiv:1611.09940, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  5. [5]

    Machine learning for combinatorial optimization: a methodological tour d’horizon.European Journal of Operational Research, 290(2):405–421, 2021

    Yoshua Bengio, Andrea Lodi, and Antoine Prouvost. Machine learning for combinatorial optimization: a methodological tour d’horizon.European Journal of Operational Research, 290(2):405–421, 2021

    2021

  6. [6]

    Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges

    Michael M Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.arXiv preprint arXiv:2104.13478, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  7. [7]

    Simulation-guided beam search for neural combinatorial optimization

    Jinho Choo, Yeong-Dae Kwon, Jihoon Kim, Jeongwoo Jae, André Hottung, Kevin Tierney, and Youngjune Gwon. Simulation-guided beam search for neural combinatorial optimization. Advances in Neural Information Processing Systems, 35:8760–8772, 2022

    2022

  8. [8]

    Bq-nco: Bisimulation quotienting for efficient neural combinatorial optimization.Advances in Neural Information Processing Systems, 36:77416–77429, 2023

    Darko Drakulic, Sofia Michel, Florian Mai, Arnaud Sors, and Jean-Marc Andreoli. Bq-nco: Bisimulation quotienting for efficient neural combinatorial optimization.Advances in Neural Information Processing Systems, 36:77416–77429, 2023

    2023

  9. [9]

    Invit: A generalizable routing problem solver with invariant nested view transformer

    Han Fang, Zhihao Song, Paul Weng, and Yutong Ban. Invit: A generalizable routing problem solver with invariant nested view transformer. InInternational Conference on Machine Learning, pages 12973–12992. PMLR, 2024

    2024

  10. [10]

    Winner takes it all: Training performant rl populations for combinatorial optimization.Advances in Neural Information Processing Systems, 36:48485–48509, 2023

    Nathan Grinsztajn, Daniel Furelos-Blanco, Shikha Surana, Clément Bonnet, and Tom Barrett. Winner takes it all: Training performant rl populations for combinatorial optimization.Advances in Neural Information Processing Systems, 36:48485–48509, 2023

    2023

  11. [11]

    An extension of the lin-kernighan-helsgaun tsp solver for constrained traveling salesman and vehicle routing problems.Roskilde: Roskilde University, 12:966–980, 2017

    Keld Helsgaun. An extension of the lin-kernighan-helsgaun tsp solver for constrained traveling salesman and vehicle routing problems.Roskilde: Roskilde University, 12:966–980, 2017

    2017

  12. [12]

    Neural large neighborhood search for the capacitated vehicle routing problem.arXiv preprint arXiv:1911.09539, 2019

    André Hottung and Kevin Tierney. Neural large neighborhood search for the capacitated vehicle routing problem.arXiv preprint arXiv:1911.09539, 2019

    work page arXiv 1911

  13. [13]

    Efficient active search for combinatorial optimization problems.arXiv preprint arXiv:2106.05126, 2021

    André Hottung, Yeong-Dae Kwon, and Kevin Tierney. Efficient active search for combinatorial optimization problems.arXiv preprint arXiv:2106.05126, 2021

    work page arXiv 2021

  14. [14]

    Pointerformer: Deep reinforced multi-pointer transformer for the traveling salesman problem

    Yan Jin, Yuandong Ding, Xuanhao Pan, Kun He, Li Zhao, Tao Qin, Lei Song, and Jiang Bian. Pointerformer: Deep reinforced multi-pointer transformer for the traveling salesman problem. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 8132–8140, 2023

    2023

  15. [15]

    Sym-nco: Leveraging symmetricity for neural combinatorial optimization.Advances in Neural Information Processing Systems, 35:1936– 1949, 2022

    Minsu Kim, Junyoung Park, and Jinkyoo Park. Sym-nco: Leveraging symmetricity for neural combinatorial optimization.Advances in Neural Information Processing Systems, 35:1936– 1949, 2022

    1936

  16. [16]

    Adam: A Method for Stochastic Optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  17. [17]

    Attention, Learn to Solve Routing Problems!

    Wouter Kool, Herke Van Hoof, and Max Welling. Attention, learn to solve routing problems! arXiv preprint arXiv:1803.08475, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  18. [18]

    Pomo: Policy optimization with multiple optima for reinforcement learning.Advances in Neural Information Processing Systems, 33:21188–21198, 2020

    Yeong-Dae Kwon, Jinho Choo, Byoungjip Kim, Iljoo Yoon, Youngjune Gwon, and Seungjai Min. Pomo: Policy optimization with multiple optima for reinforcement learning.Advances in Neural Information Processing Systems, 33:21188–21198, 2020. 10

    2020

  19. [19]

    Bopo: Neural combinatorial optimization via best-anchored and objective-guided preference optimization

    Zijun Liao, Jinbiao Chen, Debing Wang, Zizhen Zhang, and Jiahai Wang. Bopo: Neural combinatorial optimization via best-anchored and objective-guided preference optimization. arXiv preprint arXiv:2503.07580, 2025

    work page arXiv 2025

  20. [20]

    A learning-based iterative method for solving vehicle routing problems

    Hao Lu, Xingwen Zhang, and Shuang Yang. A learning-based iterative method for solving vehicle routing problems. InInternational conference on learning representations, 2019

    2019

  21. [21]

    Learning to search feasible and infeasible regions of routing problems with flexible neural k-opt.Advances in Neural Information Processing Systems, 36:49555–49578, 2023

    Yining Ma, Zhiguang Cao, and Yeow Meng Chee. Learning to search feasible and infeasible regions of routing problems with flexible neural k-opt.Advances in Neural Information Processing Systems, 36:49555–49578, 2023

    2023

  22. [22]

    Reinforcement learning for solving the vehicle routing problem.Advances in neural information processing systems, 31, 2018

    Mohammadreza Nazari, Afshin Oroojlooy, Lawrence Snyder, and Martin Takác. Reinforcement learning for solving the vehicle routing problem.Advances in neural information processing systems, 31, 2018

    2018

  23. [23]

    Preference optimization for combinatorial optimization problems

    Mingjun Pan, Guanquan Lin, You-Wei Luo, Bin Zhu, Zhien Dai, Lijun Sun, and Chun Yuan. Preference optimization for combinatorial optimization problems. InForty-second International Conference on Machine Learning, 2025

    2025

  24. [24]

    H-tsp: Hierarchically solving the large-scale traveling salesman problem

    Xuanhao Pan, Yan Jin, Yuandong Ding, Mingxiao Feng, Li Zhao, Lei Song, and Jiang Bian. H-tsp: Hierarchically solving the large-scale traveling salesman problem. InProceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 9345–9353, 2023

    2023

  25. [25]

    Optimization with gurobi and python.INESC Porto and Universidade do Porto„ Porto, Portugal, 1, 2011

    Joo Pedro Pedroso. Optimization with gurobi and python.INESC Porto and Universidade do Porto„ Porto, Portugal, 1, 2011

    2011

  26. [26]

    10,000 optimal cvrp solutions for testing machine learning based heuristics

    Eduardo Queiroga, Ruslan Sadykov, Eduardo Uchoa, and Thibaut Vidal. 10,000 optimal cvrp solutions for testing machine learning based heuristics. InAAAI-22 workshop on machine learning for operations research (ML4OR), 2021

    2021

  27. [27]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  28. [28]

    Tsplib—a traveling salesman problem library.ORSA journal on computing, 3 (4):376–384, 1991

    Gerhard Reinelt. Tsplib—a traveling salesman problem library.ORSA journal on computing, 3 (4):376–384, 1991

    1991

  29. [29]

    E (n) equivariant graph neural networks

    Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. InInternational conference on machine learning, pages 9323–9332. PMLR, 2021

    2021

  30. [30]

    Multi-task learning as multi-objective optimization.Advances in neural information processing systems, 31, 2018

    Ozan Sener and Vladlen Koltun. Multi-task learning as multi-objective optimization.Advances in neural information processing systems, 31, 2018

    2018

  31. [31]

    Difusco: Graph-based diffusion solvers for combinatorial optimization.Advances in neural information processing systems, 36:3706–3731, 2023

    Zhiqing Sun and Yiming Yang. Difusco: Graph-based diffusion solvers for combinatorial optimization.Advances in neural information processing systems, 36:3706–3731, 2023

    2023

  32. [32]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  33. [33]

    Pointer networks.Advances in neural information processing systems, 28, 2015

    Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks.Advances in neural information processing systems, 28, 2015

    2015

  34. [34]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025

    2025

  35. [35]

    Simple statistical gradient-following algorithms for connectionist reinforce- ment learning.Machine learning, 8(3):229–256, 1992

    Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforce- ment learning.Machine learning, 8(3):229–256, 1992

    1992

  36. [36]

    Learning improvement heuristics for solving routing problems.IEEE transactions on neural networks and learning systems, 33(9):5057–5069, 2021

    Yaoxin Wu, Wen Song, Zhiguang Cao, Jie Zhang, and Andrew Lim. Learning improvement heuristics for solving routing problems.IEEE transactions on neural networks and learning systems, 33(9):5057–5069, 2021. 11

    2021

  37. [37]

    Multi-decoder attention model with embedding glimpse for solving vehicle routing problems

    Liang Xin, Wen Song, Zhiguang Cao, and Jie Zhang. Multi-decoder attention model with embedding glimpse for solving vehicle routing problems. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 12042–12049, 2021

    2021

  38. [38]

    Liang Xin, Wen Song, Zhiguang Cao, and Jie Zhang. Neurolkh: Combining deep learning model with lin-kernighan-helsgaun heuristic for solving the traveling salesman problem.Advances in Neural Information Processing Systems, 34:7472–7483, 2021

    2021

  39. [39]

    Gradient surgery for multi-task learning.Advances in neural information processing systems, 33:5824–5836, 2020

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning.Advances in neural information processing systems, 33:5824–5836, 2020

    2020

  40. [40]

    Combining reinforcement learning with lin-kernighan-helsgaun algorithm for the traveling salesman problem

    Jiongzhi Zheng, Kun He, Jianrong Zhou, Yan Jin, and Chu-Min Li. Combining reinforcement learning with lin-kernighan-helsgaun algorithm for the traveling salesman problem. InPro- ceedings of the AAAI conference on artificial intelligence, volume 35, pages 12445–12452, 2021. 12 AD 4 Symmetry Group Transformations The D4 dihedral group consists of 8 isometri...

    2021

  41. [41]

    Overall and per-instance analysis.MViewRouter improves over POMO on 22 of 29 instances, with a 25.5% relative reduction in mean optimality gap (2.373% → 1.769%)

    and MViewRouter use ×8 augmentation with a model trained on TSP-100, evaluated zero-shot without retraining or fine-tuning. Overall and per-instance analysis.MViewRouter improves over POMO on 22 of 29 instances, with a 25.5% relative reduction in mean optimality gap (2.373% → 1.769%). The gains are most pronounced on instances where POMO’s gap is large, s...