acceptodds
Under review as a conference paper at ICLR 2027

ViSO: Visual Skill Distillation for Large-scale Routing Optimization

Abstract

Decomposition enables large-scale routing optimization by allowing powerful solvers to focus on tractable subproblems. Yet selecting productive subproblems remains challenging, and existing approaches leave their selection experience largely implicit and difficult to reuse. We propose ViSO, a novel framework that guides routing decomposition with transferable visual skills for solving large-scale routing problems. ViSO distills solver feedback into transferable visual skills that direct optimization toward promising subproblems. A unified decomposition interface couples this guidance with off-the-shelf solvers across symmetric, asymmetric, and capacitated routing problems. By distilling optimization experience into reusable visual skills, ViSO enables decomposition strategies to generalize to unseen instances without updating model parameters. Across four routing families with up to 85,900 nodes, ViSO reduces mean TSP and CVRP gaps by 10.5% and 11.5% relative to LKH-3 and HGS, respectively, and achieves a mean ACVRP objective value 2.72 lower than PyVRP’s. ViSO outperforms neural and LLM/VLM solvers (e.g., achieving a 26.5% improvement on TSP over ViTSP and a 94.8% improvement on CVRP over NCO-LLM), while its skill-distillation procedure is able to improve existing divide-and-conquer frameworks (e.g., DRAGON).

Primary area:
optimization
Keywords:
Combinatorial Optimization Problems, Vision-Language Models, Large-scale Optimization

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_PwQsOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ViSO, a framework that uses distilled visual skills to guide subproblem selection in large-scale routing optimization. Offline, a vision-language model compares successful and unsuccessful region selections using solver feedback, then consolidates the resulting selection rules across instances into a reusable skill bank. During inference, these skills guide the model to identify promising regions in images of the current routing solution. The selected regions are converted into smaller optimization problems by compressing fixed route segments into hypernodes, solved using conventional solvers, and incorporated only when they yield feasible improvements. The framework supports symmetric, asymmetric, and capacitated routing without updating model parameters.

Critical strengths
  1. The framework supports a wide-range of routing problems, including symmetric, asymmetric, and capacitated routing. Feasibility checks and improvement-only acceptance keep the global solution valid despite imperfect VLM proposals. This strengthens its practical relevance: the approach applies across several routing settings, and unreliable proposals cannot degrade the incumbent solution.
  2. ViSO turns solver feedback into explicit, reusable selection rules. Comparing successful and unsuccessful regions provides a sensible basis for extracting skills without updating model parameters. The reported held-out improvements support the usefulness of this mechanism and strengthen the empirical case for the approach.
Critical weaknesses
  1. The central contribution is insufficiently isolated. Visual selection and segment contraction build on prior work, making skill distillation the main novelty. Comparisons against handwritten rules, simple trajectory summaries, and individual distillation stages are needed to establish why the proposed procedure matters.
  2. The overall contribution of this paper is a bit limited. If I understand it correctly, the overall contribution of this paper is to apply the existing skill-distillation to identify which subregions of the solution should be optimized. I think that is not quite enough for a top-tier machine learning conference.
Questions for authors
  1. Can the learned skills transfer across problem families, size tiers, or solver frameworks without re-distillation? Since separate skill banks are constructed for each setting, including DRAGON, could the authors evaluate an existing bank directly in a different setting to distinguish skill transfer from the portability of the distillation procedure?
  2. Could the authors compare ViSO with strong conventional solvers under matched computational or monetary budgets to assess whether its modest absolute gap reductions justify the additional cost of hosted VLM inference and parallel solver calls?
6 comments
AuthorsOpenReview ↗

Response to Reviewer PwQs (Part 1)

We thank the Reviewer for the careful reading and the constructive comments. We respond to each comment as follows and hope our responses properly address your concerns. All changes are highlighted in blue in the revised manuscript.

W1. Visual selection and segment contraction build on prior work, so skill distillation is the main novelty. Comparisons with handwritten rules, simple trajectory summaries, and individual distillation stages are needed.

Ans. We thank the Reviewer for this constructive suggestion. We agree that new comparisons are required to isolate the contribution of the skill distillation procedure, and we report them below. Before presenting the new results, we would like to clarify the scope of our contribution.

The contribution of ViSO is not limited to skill distillation. It has two parts.

  1. A generalized vision-guided decomposition framework for routing, of which ViTSP is a special case.
  2. A skill distillation procedure designed for visual subproblem selection.

The novelty of the framework and the technical contributions are further discussed in General Response (2/3) and in the response to W2. Below we isolate the skill distillation procedure with the three comparisons the Reviewer asks for.

The full two-stage distillation outperforms handwritten rules, simple trajectory summaries, and its own first stage (new experiment). We replaced the distilled bank with four alternatives and kept the rest of ViSO unchanged.

  • Handcrafted rules. As handwritten rules we use the removal operators that experts designed for adaptive large neighborhood search, taken from the operator catalogue of a recent review (Voigt, 2025). Every operator whose definition can be stated on a single TSP tour is included, and none is chosen by its results in our experiments. The same VLM (gpt-5.4) translates each operator definition into one rule in the Name / Cue / Action / Avoid format. No trajectory, image, or result of ours is used.
  • A VLM prior-knowledge bank. The same VLM writes a bank of selection rules zero-shot, from the task definition, the image legend, and the selection interface alone, without seeing any solving trajectory. Both rule banks are used exactly like the distilled bank, through the same visual scheduler.
  • A simple trajectory summary. The VLM reads the same success–failure pairs as our distillation and keeps one free-text strategy note, which it revises batch by batch, with no critic, merge, or audit. The note is added to every selection prompt.
  • The first stage of the distillation alone, without the cross-instance consolidation of the second stage.

Setup. We use the ten TSP instances and the protocol of Table 11. All variants start from the same LKH-3 warm start (mean starting gap 1.23%) and stop after solving the same number of subproblems, with three runs per instance.

Mean final gap (%), lower is better:

Guidance in the selection promptAllSmallLarge
Handcrafted rules (ALNS operators)0.3250.2680.350
VLM prior-knowledge bank0.2990.2380.325
Simple trajectory summary0.5720.3880.651
First stage of the distillation alone0.3470.2910.371
ViSO (full two-stage distillation)0.2710.1910.305

Results. The full two-stage bank attains the lowest gap on all instances and on both tiers. The handcrafted rules reach 0.325%, the VLM prior-knowledge bank 0.299%, and the simple trajectory summary 0.572%. With the first stage alone the gap is 0.347%, and adding the second stage lowers it to 0.271%.

Overall, these comparisons isolate the contribution of the skill distillation procedure. Neither handwritten rules, prior knowledge, a simple summary of the same trajectories, nor the first stage of the distillation alone matches the full two-stage distillation. We have added these comparisons as Table R5 in Appendix E of the revised manuscript.

References

Stefan Voigt. A review and ranking of operators in adaptive large neighborhood search for vehicle routing problems. European Journal of Operational Research, 322(2):357–375, 2025.

AuthorsOpenReview ↗

Response to Reviewer PwQs (Part 2)

W2. The contribution appears limited, since the paper seems to apply existing skill distillation to decide which subregions of the solution to optimize.

Ans. We sincerely thank the Reviewer for raising this important concern. We respectfully clarify that the contribution of this paper goes beyond applying an existing skill-distillation method. ViSO contributes a constraint-aware visual decomposition framework and a mechanism for turning solver feedback into reusable guidance for allocating optimization effort. We describe these technical advances and their supporting evidence below. Please also refer to General Response (2/3) for a comparison with existing methods.

  1. A unified decomposition interface that extends visual optimization beyond symmetric Euclidean TSP.

ViTSP establishes the effectiveness of visual selection for symmetric Euclidean TSP. Extending this approach to asymmetric and capacitated routing introduces additional requirements: matrix-only instances lack native coordinates, and reconnecting selected regions must preserve directed segment costs and resource constraints. ViSO addresses these requirements through a common interface. MDS provides a layout for visual selection, while constraint-aware hypernodes retain each fixed segment’s entry and exit nodes, internal cost, and aggregate demand. Connections between hypernodes use the original endpoint costs. After re-optimization, the segments are expanded and the complete solution is checked for feasibility and improvement against the original instance (Secs. 3.2, 3.5–3.6; Appendix A). Thus, visual selection can operate across different routing families while optimization and verification retain their original costs and constraints.

This framework has value independently of skill distillation: the skill-free variant achieves a mean CVRP gap of 0.243%, compared with 0.261% for HGS under the same runtime budget (Tables 1 and 9).

2. Skill acquisition and deployment specifically designed for visual subproblem selection.

We acknowledge that learning reusable insights from successful and failed experiences is established by methods such as ExpeL. The technical contribution lies in how ViSO constructs and applies this knowledge for visual decomposition. Specifically, ViSO builds on this principle to learn where solver effort produces useful improvements in a visualized solution. Its learning examples associate selected regions with the solver-measured change in the global objective. Because regions differ in their size and optimization difficulty, ViSO compares successful and unsuccessful selections within the same instance, pairing regions with similar numbers of selected nodes. This reduces differences in instance difficulty and neighborhood size when identifying visual patterns associated with productive selections (Sec. 3.4).

Our distillation procedure also separates independent evidence extraction from collective refinement. Whereas ExpeL sequentially updates a shared insight list, ViSO uses independent visual critics to generate proposals before hierarchical merging. The merged bank is then audited against successful selections from distinct training instances. This allows each comparison to be examined without conditioning its initial interpretation on earlier insight edits. The resulting Cue / Action / Avoid rules state when a rule applies (Cue), which region to select (Action), and which failure to avoid (Avoid), while excluding instance identifiers and absolute coordinates.

Deployment also accounts for the visual instance: an image-conditioned scheduler weights the skill bank according to the initial solution. This design has direct empirical support. With the same skill bank, initialization, and number of solved subproblems, image-conditioned scheduling reduces the mean gap from 0.293% to 0.271%, a 7.5% relative reduction, compared with uniform skill sampling (Fig. 4(c)). This demonstrates the value of conditioning skill use on the instance, beyond simply supplying a bank of advice.

3. ViSO outperforms strong solvers on large-scale routing, and its skills generalize beyond ViSO. On 200+ instances with up to 85,900 nodes across four routing variants and under the same wall-clock limit, the mean gap of ViSO is 10.5% lower than that of LKH-3 on TSP and 11.5% lower than that of HGS on CVRP, and its mean objective is 2.72% below PyVRP on ACVRP (Tables 1 and 2). The same procedure also improves DRAGON, a text-only framework (Appendix I), and carries over to 2D bin packing (see General Response (3/3)).

Overall, ViSO contributes a generalized vision-guided decomposition framework for routing and a skill distillation procedure specifically designed for visual subproblem selection. Together, they outperform strong solvers on large-scale routing and generalize beyond ViSO.

AuthorsOpenReview ↗

Response to Reviewer PwQs (Part 3)

Q1. Can existing skill banks be used in other problem families, size tiers, or frameworks without re-distillation, to separate skill transfer from the portability of the distillation procedure?

Ans. We thank the Reviewer for this insightful question. Following the Reviewer’s suggestion, we evaluated existing banks directly in a different setting to test the transferability of the skills themselves, without any re-distillation, in two ways.

  • Across problem families. The bank distilled from ATSP trajectories is used to solve TSP instances. Each TSP instance receives the ATSP bank of its own size tier.
  • Across size tiers. The two TSP banks are swapped. The large TSP instances use the bank distilled on the small tier, and the small instances use the bank distilled on the large tier.

Setup. We use the ten TSP instances and the protocol of Table 11. All variants start from the same LKH-3 warm start and stop after solving the same number of subproblems, with three runs per instance. Only the bank is changed.

Mean final gap (%) on TSP, lower is better:

Bank usedAllSmallLarge
Baseline TSP results (TSP bank of the same tier, Table 11)0.2710.1910.305
Across problem families (ATSP bank used on TSP)0.2790.2410.295
Across size tiers (small-tier and large-tier TSP banks swapped)0.3290.2930.345

Results. With the ATSP bank, the mean gap on TSP is 0.279%, compared with 0.271% with the TSP bank itself. With the TSP banks swapped between tiers, the gap rises to 0.329%. The skills therefore transfer across problem families with little loss, whereas a bank distilled on small instances is less suited to large instances, and vice versa. Because instances of different scales have different visual characteristics and call for different skills, ViSO distills one bank per size tier (Sec. 3.4).

Overall, an existing bank can be reused across problem families without re-distillation and with little loss, while one bank per size tier remains preferable. The distillation procedure itself also carries over to other frameworks, as shown for DRAGON. (Appendix I) We have added these results as Table R6 in Appendix E of the revised manuscript.

Reviewer_tQRrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ViSO, a framework for large-scale routing optimization that distills a visual skill bank from training instances to guide subproblem selection. Specifically, starting from a feasible solution, a VLM selects promising regions, which are reformulated through hypernode contraction and optimized using existing solvers. The skills are extracted from successful and unsuccessful selections and consolidated across training instances, without updating model parameters. Experiments cover TSP, ATSP, CVRP, and ACVRP, with additional ablations.

Critical strengths
  1. The proposed framework combines VLM-based selection with established optimization solvers in a reasonable way. Reconstructed solutions are checked against the original problem before acceptance, providing a clear mechanism for maintaining feasibility and improving the incumbent.
  2. The experiments cover four routing families and include strong traditional solvers. The results on large TSP instances and ACVRP are encouraging. The skill ablation also reports improvements on held-out instances across all four families.
Critical weaknesses
  1. The technical novelty over closely related work is not clearly established. ViTSP already proposed to combine visual region selection, segment aggregation and solver-based re-optimization. Meanwhile, ExpeL proposed to extract natural-language insights by comparing successful and failed trajectories from the same task and identifying common patterns across successful trajectories from different tasks. Considering this, this paper should sufficiently demonstrate the additional technical value of its specific distillation design beyond adapting existing ideas.

  2. The ablations do not fully demonstrate the effectiveness of the proposed skill-generation procedure. For examples, Table 9 shows that the distilled skills improve over a skill-free selector, but there are no comparisons with simple experience-summary baseline. Moreover, Table 10 shows that random box selection outperforms the skill-free VLM overall on TSP and even outperforms the full method on the small tier.

  3. The overall performance advantages do not necessarily imply a substantial contribution from the distilled skills, whose incremental benefits appear limited in some settings. For example, on CVRP, the skill-free framework in Table 9 already achieves a mean gap of 0.243%, outperforming HGS and SISRs in Table 1. Adding the distilled skills further reduces this gap to 0.231%.

Questions for authors
  1. How does ViSO's computational cost compare with that of the competing methods, accounting for both VLM inference and parallel solver usage?
7 comments
AuthorsOpenReview ↗

Response to Reviewer tQRr (Part 1)

We thank the Reviewer for the careful reading and the constructive comments. We respond to each comment as follows and hope our responses properly address your concerns. All changes are highlighted in blue in the revised manuscript.

W1. The technical novelty over ViTSP and ExpeL is not clearly established, and the paper should demonstrate the additional technical value of its specific distillation design.

Ans. We sincerely thank the Reviewer for this important and constructive comment, which helps us articulate the technical contributions of ViSO more clearly. We fully agree that the additional technical value of ViSO beyond existing ideas should be demonstrated explicitly. We first present the differences from ViTSP. We then compare our skill distillation with ExpeL, and show the additional technical contributions specifically designed for visual subproblem selection in large-scale routing. See also General Response (2/3).

1. ViSO is a generalized framework for routing, and ViTSP is a special case of it. In ViSO, a VLM guided by distilled skills selects regions of the rendered solution, which off-the-shelf solvers re-optimize, for symmetric, asymmetric, and capacitated problems. ViTSP corresponds to the special case of a symmetric Euclidean instance with no side constraints and no skill bank.

ViTSPViSO
Problem familiesSymmetric Euclidean TSPTSP, ATSP, CVRP, ACVRP
Asymmetric (matrix-only) instancesNot supportedMDS-based rendering
Side constraintsNoneConstraint-aware hypernodes
Reuse of selection experienceRaw per-instance trajectoriesDistilled skill bank, reused on unseen instances
TSP mean gap (33 TSPLIB instances)0.393%0.289% (26.5% lower)
CVRP / ATSP / ACVRP mean gapNot applicable0.231% / 0.207% / −2.717%

2. Comparison with ExpeL: our skill distillation is specifically designed for visual subproblem selection. While ExpeL is an experiential-learning method for text-based agent tasks, our two-stage parallel skill distillation (Sec. 3.4) is specifically designed for visual subproblem selection in large-scale routing. By learning from solver feedback on rendered solutions, ViSO obtains visual skills that transfer to unseen instances.

ExpeLViSO
Learning signalText trajectories, labeled by task successRegion selections on rendered solutions, labeled by the solver-measured change in global cost
Distillation procedureOne list of insights, edited sequentially, so a later edit can overwrite an earlier insightIndependent critics run concurrently on success–failure pairs and are merged afterwards, so none is overwritten. The bank is then audited across instances
Skill formatFree-form insightsCue / Action / Avoid skills over visual cues, with no instance-specific reference
The way the skills are employedThe full list of insights in every promptSkills are weighted once per instance from its initial image, and each selected region cites its skill

3. Together, these designs bring additional technical advantages.

  • Interpretability. The skill bank is human-readable, and every selection cites its skill.
  • Stronger performance. With the same VLM, skills reduce the gap on TSP, ATSP, and CVRP by 18.5%, 15.2%, and 4.9%, and increase the advantage over PyVRP on ACVRP by 2.8% (Table 9).
  • Generalization across problem types. Without re-distillation, the ATSP bank solves TSP at 0.279%, close to 0.271% for the TSP bank (see our response to Q1 of Reviewer PwQs).
  • Generalization across frameworks. Applied to DRAGON, a text-only LLM-based framework, the skill-guided interface lowers the mean gap from 15.424% to 15.206% under the same runtime (Fig. 4(d), Appendix I).
  • Generalization beyond routing. On 2D bin packing, the same procedure lowers the gap of the skill-free variant from 18.82% to 15.75% (see our response to W1 of Reviewer FMYw).

In summary, ViSO contributes a vision-guided decomposition framework for symmetric, asymmetric, and capacitated routing, of which ViTSP is a special case, and a distillation procedure designed for visual subproblem selection that yields interpretable and transferable skills. Extensive results show its effectiveness and its generalization across problem types, frameworks, and problems beyond routing. We have added these comparisons as Tables R10 and R11 (Appendix L) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer tQRr (Part 2)

W2. The ablations lack a comparison with a simple experience-summary baseline. In Table 10, random box selection also outperforms the skill-free VLM overall and the full method on the small tier.

Ans. We thank the Reviewer for these two helpful observations. Below we compare the distilled skills with a simple experience-summary baseline and then clarify the role of random box selection (see also General Response (3/3)).

Comparison with a simple experience-summary baseline. We have added this baseline, together with three further alternatives to the distilled bank. In each case only the guidance in the selection prompt is replaced, and the rest of ViSO is unchanged.

  • Simple trajectory summary. The VLM reads success–failure pairs from the same training instances and keeps one free-text strategy note, which it revises batch by batch. There are no per-pair critic calls, no merge, and no cross-instance audit. The note is added to every selection prompt.
  • Handcrafted rules. Rules translated from the removal operators that experts designed for adaptive large neighborhood search.
  • VLM prior-knowledge bank. Rules written zero-shot by the same VLM from the task definition, the image legend, and the selection interface alone, without any solving trajectory.
  • The first stage of the distillation alone, without the cross-instance consolidation of the second stage.

Setup. We use the ten TSP instances and the protocol of Table 11. All variants start from the same LKH-3 warm start (mean starting gap 1.23%) and stop after solving the same number of subproblems, with three runs per instance.

Mean final gap (%), lower is better:

Guidance in the selection promptAllSmallLarge
Simple trajectory summary0.5720.3880.651
Handcrafted rules (ALNS operators)0.3250.2680.350
VLM prior-knowledge bank0.2990.2380.325
First stage of the distillation alone0.3470.2910.371
ViSO (full two-stage distillation)0.2710.1910.305

Results. The distilled bank attains the lowest gap on all instances and on both tiers. The simple trajectory summary reaches 0.572%, compared with 0.271% for ViSO.

The performance gain comes from both the framework and the skills. We would like to emphasize that random box selection is not a separate method. It runs inside the ViSO framework, with the same hypernode contraction, the same sub-solvers, and the same acceptance rule, and only the VLM selection is replaced by a random box. Its result therefore reflects the strength of the decomposition framework itself, which is part of our contribution. On the large tier, the framework with random boxes already reaches 0.663%, compared with 0.746% for LKH-3 under the same runtime, indicating that hypernode decomposition with solver-based re-optimization is effective even before any guidance is added.

As shown in Table 10, skill-guided selection improves further on this. On the large tier, ViSO lowers the gap from 0.663% with random boxes to 0.494%, a 25.5% relative reduction. Over all instances, it lowers the gap from 0.335% to 0.289%. On the small tier, where all methods are already near-optimal, random boxes are slightly better (0.122% vs. 0.156%). Therefore, the benefit of guided selection increases with instance size, which is the regime that large-scale routing targets.

Overall, the distilled skills outperform a simple experience summary and other alternatives, and the strong result of random boxes reflects the strength of the ViSO framework. On top of the ViSO framework, skill-guided selection yields a further gain that grows with instance size. We have added the comparison table as Table R5 in Appendix E of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer tQRr (Part 3)

W3. The benefit of the skills appears limited in some settings. On CVRP, the skill-free framework already reaches 0.243%, below HGS, and the skills only lower it to 0.231%.

Ans. Thank you for the careful reading of Table 9. We agree that the skill-free framework is already strong on CVRP. The distilled skills still contribute a consistent gain, which is statistically significant on TSP, ATSP, ACVRP, and the large CVRP instances.

1. On CVRP, the skills account for about 40% of the margin over HGS. The gap goes from 0.261% (HGS) to 0.243% (ViSO without skills) to 0.231% (ViSO).

2. The gain from skills is statistically significant (new analysis). We ran a one-sided paired t-test over instances, pairing the mean gap of each instance with and without skills:

FamilyInstancesw/o skills (%)ViSO (%)p-value (one-sided)
TSP330.3550.2890.011
ATSP400.2440.2070.039
ACVRP40−2.643−2.7170.001
CVRP, large tier (n>500)320.4920.4590.008
CVRP, small tier (n<500)680.1260.1240.41

On the small CVRP tier, all methods are already close to the best-known solutions, which leaves almost no headroom. On the large tier, where the instances are harder, the skills lower the gap by 6.7%.

3. The strong skill-free result on CVRP comes from the ViSO framework itself, which is part of our main contribution. The 0.243% gap is obtained with the constraint-aware hypernode decomposition of ViSO, a VLM selector, and HGS as the sub-solver, whereas ViTSP cannot be applied to CVRP at all.

Overall, the strong skill-free result on CVRP reflects the ViSO framework. On top of the ViSO framework, the distilled skills add a further gain that is significant wherever headroom remains. See also General Response (3/3). We have added the statistical test and the per-tier breakdown as Table R4 in Appendix E of the revised manuscript.

Reviewer_FMYwOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a novel visual skill distillation framework that leverages offline-collected trajectories to construct a transferable skill bank for identifying promising subproblems in large-scale routing optimization. At inference time, a VLM utilizes this bank to select the most suitable decomposition strategies, pinpointing specific subproblems to be refined by off-the-shelf solvers. Experimental results demonstrate that our method outperforms traditional OR heuristics, as well as state-of-the-art learning-based and LLM/VLM-based solvers. This study characterizes the visual optimization heuristics explicitly and advances the application of VLMs in solving combinatorial optimization problems by bridging visual reasoning with solver feedback.

Critical strengths
  • It advances the visual capability of VLM in combinatorial optimization and builds an explicit decomposition skill bank to make VLM more reliable in assisting the optimization.
  • This study conducted extensive experiments on multiple problems within the routing family, including asymmetric ones, to demonstrate the soundness of the proposed framework.
Critical weaknesses
  • While expanding the experiments to multiple routing problems are appreciated, it did not include any non-routing problems to evaluate the effectiveness of the ViSO.
  • More details should be disclosed for how to determine the runtime limit to ensure that the comparisons across baselines are fair, and for how VLM latency and solver latency are counted explicitly.
  • From a mechanism perspective, the idea is similar to adaptive large neighborhood search (ALNS), which scores each destroy operator used during the iteration. It would be beneficial to incorporate ALNS as one of the baselines.
Questions for authors
  1. When distilling skills, there are existing LLM-based evolutionary methods, such as EoH, to iteratively design code-based functions to execute a skill. What is the motivation of distill skills as text instead of code functions in this study?
  2. In line 242, "the r-th strongest success is paired with the unused failure closest to it in region size", the wording implies that it could be 1st, 2nd, or 3rd strongest success. Could you clarify how you determined which strongest success will be selected?
  3. Please consider comparing with ALNS. Also, when applying the skill bank across instances, is the frequency of each skill calculated based on previous instances be recorded and considered as a factor to weigh the selection?
  4. Please consider comparing with code-based logics such as EoH.
  5. Please report the VLM latency plus solver latency during the test time and discuss how did you determine the runtime limit for each instance.
  6. For highly constrained problems like VRPTW, there is a potential information bottleneck if representing the problem as an image. How do the authors ensure that critical numerical constraints, such as precise time windows or capacity limits, are sufficiently preserved in a2D visual overlay for the VLM to make a feasible decision?
  7. Please clarify what selector is used during the trajectory collection stage. Do you use a VLM to perform as a selector, as in VITSP, or do you use a heuristic method to select the subproblems?

EoH: Evolution of Heuristics: Towards Efficient Automatic Algorithm Design Using Large Language Model. ICML 2024.

ALNS: https://alns.readthedocs.io/en/latest/index.html

10 comments
AuthorsOpenReview ↗

Response to Reviewer FMYw (Part 1)

We thank the Reviewer for the positive assessment and the insightful comments. We respond to each comment as follows and hope our responses properly address your concerns. All changes are highlighted in blue in the revised manuscript.

W1. The evaluation covers only routing problems, and ViSO should also be evaluated on non-routing problems.

Ans. Thank you for this suggestion. We answer in two parts.

1. This paper focuses on routing. As the title indicates (Visual Skill Distillation for Large-Scale Routing Optimization), its scope is routing rather than general combinatorial optimization. Within this scope, ViSO solves symmetric, asymmetric, and capacitated problems, and our evaluation covers 200+ instances with up to 85,900 nodes across four routing variants (TSP, ATSP, CVRP, and ACVRP).

2. ViSO also generalizes to a non-routing problem: 2D bin packing (new experiment). We agree that a non-routing problem is a valuable test of generality, so we applied ViSO to the two-dimensional bin packing problem (2D-BPP). The adaptation of ViSO to 2D-BPP is direct. The framework, the two-stage distillation, and the visual scheduler are unchanged, and only the problem-specific components are replaced: the image shows the bins of the current packing, the VLM selects the bins to re-optimize, and an exact CP-SAT solver repacks their items.

Setup. We use 60 instances of the standard 2D-BPP benchmark of Berkey & Wang (1987) with a roughly 70/30 training/held-out split. Both variants start from the same initial packing and receive the same runtime budget and the same VLM. Gaps are measured against the area lower bound, so they are upper bounds on the true optimality gaps.

Splitw/o skills: binsViSO: binsw/o skills: gap (%)ViSO: gap (%)Rel. gain
All (60)24.7923.9418.8215.7516.3%
Train (42)24.7523.9118.8515.9115.6%
Held-out (18)24.8724.0018.7515.3917.9%

Results. Starting from an initial packing of 33.37 bins on average (63.55% above the lower bound), the skill-free framework reduces the packing to 24.79 bins, and the distilled skills reduce it further to 23.94 bins. As in Table 9 of the paper, we report the relative gain of skill guidance over the skill-free variant. Skills lower the gap from 18.82% to 15.75%, a 16.3% relative reduction. The gain carries over to the held-out instances that were never used for distillation, where the gap falls from 18.75% to 15.39%, a 17.9% relative reduction. The improvement holds in all three classes (0.87, 0.48, and 1.20 fewer bins on Classes 1, 3, and 5).

Overall, this paper focuses on routing, and the new experiment shows that both the framework and the skill distillation carry over to a non-routing problem, with only the problem-specific components replaced. We have added this experiment as Appendix J and refer to it in Section 5 of the revised manuscript.

References

Judith O Berkey and Pearl Y Wang. Two-dimensional finite bin-packing algorithms. Journal of the operational research society, 38(5):423–429, 1987.

AuthorsOpenReview ↗

Response to Reviewer FMYw (Part 2)

W2. How is the runtime limit determined to ensure fair comparisons with the baselines, and how are the VLM latency and the solver latency counted?

Ans. Thank you for emphasizing the importance of the runtime protocol. We clarify it below and have expanded Appendix C accordingly.

The runtime limit is fixed in advance by problem family and instance size. Each instance receives an optimization budget TbudgetT_{\text{budget}} that is fixed in advance by problem family and size tier (Table 4), so that larger instances receive more time:

FamilyInstance size nTbudgetT_{\text{budget}} (s)
TSP1,000–1,889 / 2,103–2,392 / 3,038–7,397 / 11,849–18,512 / 33,810–85,900100 / 200 / 500 / 1,000 / 3,000
ATSP1,000 / 2,000 / 5,000 / 10,000100 / 200 / 500 / 1,000
CVRP101–491 / 502–1,001100 / 200
ACVRP500 / 1,000200

TinitT_{\text{init}} is the time of the warm start. Before any VLM call, ViSO obtains a feasible initial solution from a conventional solver: ten LKH-3 runs for TSP and ATSP, 30 s of HGS for CVRP, and 30 s of PyVRP for ACVRP. TinitT_{\text{init}} is the wall-clock time of this step. It is measured on each instance for TSP and ATSP, and it is 30 s for CVRP and ACVRP. The optimization budget TbudgetT_{\text{budget}} is applied on top of it, so the matched per-instance limit is TmatchedT_{\text{matched}} = TinitT_{\text{init}} + TbudgetT_{\text{budget}}.

All baselines receive the same runtime limit. Every baseline is aligned to TmatchedT_{\text{matched}} according to how it consumes time (Appendix C):

  • Solvers with a native time limit (LKH-3, Concorde, HGS, PyVRP, OR-Tools) receive TmatchedT_{\text{matched}} as their limit, i.e., the warm-start time plus the full optimization budget.
  • Iterative neural methods (GLOP, L2C-Insert, NCO-LLM, DRHG) are given a per-instance wall-clock cap equal to TmatchedT_{\text{matched}}.
  • Methods with fixed decoding procedures (L2R, URS, RADAR, UniCO, RRNCO) have no controllable time limit. We run them in their strongest available configuration and state this explicitly in Appendix C.

Both the VLM latency and the solver latency are counted inside the runtime limit. TbudgetT_{\text{budget}} is an end-to-end wall-clock deadline that starts when the warm start returns. Both the VLM producers and the solver consumers stop at this deadline, and nothing is excluded from it. It covers:

1. all VLM API waiting and latency, including the zoom-in re-selection calls,

2. all sub-solver time (Concorde, HGS, or PyVRP on the selected subproblems),

3. rendering, subproblem construction, verification, and merging.

The “Time” column of Tables 5–8 reports the measured total wall-clock time including initialization, i.e., TinitT_{\text{init}} plus everything above. For example, the 16 TSP instances with a 100 s budget finish in 101–110 s (Table 5), and the ACVRP instances finish in 231–235 s for a 230 s target (Table 8).

Since VLM selection and subproblem solving run asynchronously, their latencies overlap, and the budget bounds the end-to-end time rather than their sum. The only cost outside the per-instance time is the one-time offline skill distillation, which is reported separately (Table 12). We have added these details to the Time alignment paragraph of Appendix C in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer FMYw (Part 3)

W3. The mechanism resembles ALNS, which scores each destroy operator during the search, so ALNS should be included as a baseline.

Ans. Thank you for this suggestion. We have added ALNS (Ropke & Pisinger, 2006) as a baseline, which also addresses the first part of Q3.

Setup. We use the alns package (Wouda & Lan, 2023) referenced by the Reviewer and evaluate it on the same TSPLIB and CVRPLIB X-set instances under the same matched runtime, with three seeds. We report two settings: ALNS as a standalone solver, and ALNS started from the same warm start as ViSO.

Mean gap (%), lower is better:

MethodTSP AllTSP SmallTSP LargeCVRP AllCVRP n<500CVRP n>500
ALNS (standalone)4.5374.5504.5164.4643.7026.084
Warm start only0.861*0.5601.363*1.4990.9332.702
ALNS (same warm start)0.843*0.5601.315*1.4890.9262.686
ViSO0.2890.1560.4940.2310.1240.459
  • Averaged over 32 of the 33 TSP instances. The warm-started ALNS run on pla85900 is not available.

Results. ViSO attains lower gaps than ALNS in both settings. As a standalone solver, ALNS reaches 4.54% on TSP and 4.46% on CVRP. From the same warm start, ALNS changes the solution only slightly (0.861% to 0.843% on TSP, 1.499% to 1.489% on CVRP), whereas ViSO reaches 0.289% and 0.231%.

Overall, we have incorporated ALNS as a baseline as suggested, and ViSO outperforms it on both TSP and CVRP. We have added ALNS to Table 1 of the revised manuscript, with the full results in Table R8 (Appendix K).

References

Stefan Ropke and David Pisinger. An adaptive large neighborhood search heuristic for the pickup and delivery problem with time windows. Transportation science, 40(4):455–472, 2006.

Niels A Wouda and Leon Lan. ALNS: a Python implementation of the adaptive large neighbourhood search metaheuristic. Journal of Open Source Software, 8(81):5028, 2023.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response (Part 1): Summary of new experiments

Dear Reviewers, ACs, and SACs,

We sincerely thank all reviewers for the constructive feedback. We are encouraged that the reviewers recognize the explicit and reusable skill bank (FMYw, PwQs), the coverage of symmetric, asymmetric, and capacitated routing (FMYw, tQRr, PwQs), the feasibility checks with improvement-only acceptance (tQRr, PwQs), and the gains on held-out instances (tQRr, PwQs).

In response to the reviews, we conducted six sets of new experiments and added them to the revised manuscript, where all changes are highlighted in blue. The key findings of the six sets are as follows:

1. The two-stage distillation outperforms simpler ways of producing the guidance. On TSP, the full two-stage bank reaches a mean gap of 0.271%, compared with 0.572% for a simple trajectory summary, 0.325% for handcrafted rules, 0.299% for a VLM prior-knowledge bank, and 0.347% for the first stage of the distillation alone (Figure 4(e) and Table R5 in Appendix E).

2. The skills transfer across problem families without re-distillation. The ATSP bank solves TSP at 0.279%, compared with 0.271% for the TSP bank itself (Figure 4(f) and Table R6 in Appendix E).

3. The skills carry over to a non-routing problem. On 2D bin packing, they lower the gap from 18.82% to 15.75%, and from 18.75% to 15.39% on held-out instances (Figure 4(g) and Table R7 in Appendix J).

4. ViSO outperforms ALNS. ALNS reaches 4.54% on TSP and 4.46% on CVRP, compared with 0.289% and 0.231% for ViSO (Table 1 and Table R8 in Appendix K).

5. ViSO outperforms the code-evolution methods EoH and EoH-S. They reach 14.8–26.0% on TSP and 15.2–19.7% on CVRP, compared with 0.289% and 0.231% for ViSO (Table 1 and Table R9 in Appendix K).

6. ViSO keeps the lowest gap on TSP even when LKH-3 and Concorde receive its full CPU time. Under this budget, LKH-3 and Concorde reach 0.375% and 0.409%, compared with 0.293% for ViSO (Table R3 in Appendix D).

Below we address the two concerns shared by several reviewers: the contribution and novelty of ViSO (General Response (2/3)) and the role of the distilled skills (General Response (3/3)). The details of each experiment are given in the individual replies.

AuthorsOpenReview ↗

General Response (Part 2): Technical Contribution

We thank reviewers tQRr and PwQs for suggesting that we clarify how ViSO differs technically from ViTSP and ExpeL. We agree that these distinctions will help make ViSO’s contribution clearer. ViSO is a generalized vision-guided decomposition framework for large-scale routing, driven by visual skills that are distilled from solver feedback and reused on unseen instances without any parameter update.

1. ViSO is a generalized framework for routing, and ViTSP is a special case of it. ViTSP is confined to the symmetric Euclidean TSP, with no side constraints and no skill bank. Three components make ViSO general.

  • Constraint-aware hypernodes. Each contracted segment carries its aggregated resources, such as its demand. Constrained problems can therefore be decomposed and re-optimized by an off-the-shelf solver, and every candidate is checked against the original instance before acceptance.
  • Rendering of matrix-only instances. An MDS-based layout lets the VLM reason about asymmetric instances that have no coordinates, while all costs and checks use the true cost matrix.
  • A distilled skill bank. ViTSP reuses raw per-instance trajectories. ViSO distills solver feedback into explicit, reusable skills (see point 2).

As a result, ViSO solves TSP, ATSP, CVRP, and ACVRP in one framework, and its formulation also covers further VRP variants (Appendix A). Specifically, on TSP, its mean gap is 26.5% lower than that of ViTSP (0.289% vs. 0.393%), and on ATSP and ACVRP none of the LLM/VLM-based baselines returns a feasible solution (Table 2).

2. ViSO introduces a skill distillation mechanism for visual subproblem selection. While ExpeL distills lessons from text-based agent trajectories, ViSO tackles a different technical problem: turning solver feedback into reusable visual skills that guide which routing subproblems to select. Four design choices specific to this problem distinguish ViSO’s mechanism from ExpeL’s.

  • Different learning signal. ExpeL labels a whole text trajectory by whether the task succeeded. In contrast, ViSO labels each region selection by the solver-measured change in global cost, so the signal is objective and attached to a single decision.
  • Different distillation procedure. ExpeL updates a shared insight list sequentially, so later edits can overwrite earlier insights. In ViSO, independent critics evaluate all success–failure pairs in parallel. ViSO collects all proposals before merging them, preventing one critic from overwriting another’s proposal during evaluation. It then audits the consolidated skill bank against successful selections from other instances.
  • Different skill format. ViSO structures each skill around a region-selection decision: Cue specifies what visual pattern the VLM should look for, Action specifies which region to select, and Avoid identifies the failure mode to guard against. Each skill abstracts away instance-specific details so it can be applied to unseen routing instances.
  • ViSO employs distilled skills differently from ExpeL. ExpeL includes the full list of insights in every prompt. ViSO uses the initial image to assign skill weights once per instance, then samples a skill from that distribution for each selection. The VLM cites the sampled skill when reporting the selected region.

By this design, the skills generalize to unseen instances and transfer across problem types and sizes (see General Response (3/3)).

3. Together, these designs lead to the following advantages:

  • Interpretability. The selection knowledge is an explicit, human-readable skill bank instead of model weights or raw trajectories, and every selection can be traced to the skill it cites.
  • Stronger performance. On 200+ instances with up to 85,900 nodes across four routing variants and under the same runtime, the mean gap of ViSO is 10.5% lower than that of LKH-3 on TSP and 11.5% lower than that of HGS on CVRP, and its mean objective is 2.72% below PyVRP on ACVRP.
  • Generalization beyond ViSO. The same distillation procedure also improves DRAGON, a text-only framework, and carries over to 2D bin packing, a non-routing problem (see General Response (3/3)).

In summary, ViSO makes two unique technical advances. First, it generalizes the vision-guided decomposition used in ViTSP to a routing framework that includes ViTSP as a special case. Second, it introduces a skill distillation procedure that uses solver-measured changes in global routing cost to learn which visual cues should guide region selection. The resulting skills transfer to unseen instances without updating model parameters. A detailed comparison with ViTSP and ExpeL is given in Tables R10 and R11 (Appendix L) of the revised manuscript.

AuthorsOpenReview ↗

General Response (Part 3): Role of the distilled skills

Reviewers tQRr and PwQs suggest isolating the contribution of the distilled skills from that of the framework, and validating the distillation procedure against simpler alternatives.

We present four lines of evidence that distilled skills are needed to achieve ViSO’s full performance gains. We then show that the vision-guided decomposition framework makes an independent contribution on its own.

1. With the same VLM, skills improve all four families, also on held-out instances. Skills lower the gap in every family, and the gains hold on instances that were never used for distillation (Table 9). On TSP, the benefit grows with instance size, from an 11% relative reduction on the small tier to 22% on the large tier (Table 10), which is particularly relevant to large-scale routing.

2. Alternative forms of guidance are consistently less effective than the full two-stage distilled skills (new experiment). We replaced the bank of the full two-stage distillation with four alternatives and kept the rest of ViSO unchanged:

  • Handcrafted rules, translated from the removal operators that experts designed for adaptive large neighborhood search.
  • A VLM prior-knowledge bank, written zero-shot by the same VLM from the task definition, the image legend, and the selection interface alone, without seeing any solving trajectory.
  • A simple trajectory summary, in which the VLM reads the same success–failure pairs as our distillation and keeps one free-text strategy note, with no critic, merge, or audit.
  • The first stage of the distillation alone, without the cross-instance consolidation of the second stage.

The full two-stage bank attains the lowest gap in every comparison (the numbers are in finding 1 of our summary of new experiments and in Table R5 of Appendix E).

3. The skills transfer to other problem types and sizes without re-distillation (new experiment). The ATSP bank solves TSP at 0.279%, and the TSP bank of the other size tier at 0.329%, compared with 0.271% for the matching bank. The loss is small across problem types and larger across size tiers, which is why ViSO distills one bank per tier.

4. The distillation procedure generalizes to a non-routing problem (new experiment) and to another LLM-based routing framework. On 2D bin packing, it lowers the gap of the skill-free variant from 18.82% to 15.75%. Applied to DRAGON, a text-only LLM-based framework, the skill-guided interface lowers the mean gap from 15.424% to 15.206% under the same runtime (Fig. 4(d), Appendix I).

5. Even without skills, the generalized framework is itself a contribution. Without any skill, ViSO reaches 0.243% on CVRP, below HGS at 0.261% under the same runtime (Tables 1 and 9).

Overall, these results show that the distilled skills are both necessary and transferable. Removing them, or replacing them with alternative forms of guidance, degrades performance. The skills carry over to unseen instances and to other problem types, and the distillation procedure carries over to a non-routing problem and to another LLM-based framework. More details are given in our responses to Reviewers tQRr (W2, W3), PwQs (W1, Q1), and FMYw (W1).

Summary

In summary, ViSO contributes a generalized vision-guided decomposition framework for symmetric, asymmetric, and capacitated routing, of which ViTSP is a special case, and a skill distillation procedure designed for visual subproblem selection. The new experiments clearly establish the effectiveness of the skill distillation procedure.

We hope that our responses resolve your concerns. If so, we would be grateful if you would consider updating your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.