acceptodds
Under review as a conference paper at ICLR 2027

DCAST: Decoupled Constraint Activation Signature Tokenization for Multi-Task Neural Multi-Objective Vehicle Routing Problems

Abstract

Neural multi-objective combinatorial optimization (NMOCO) has attracted increasing attentions in solving multi-objective vehicle routing problems (MOVRPs) by their fast Pareto-front approximation capability. However, existing NMOCO solvers are typically trained for a single MOVRP variant, which severely restricts their applications in real-world industry problems with diverse characteristics. Therefore, this paper concentrates on developing multi-task NMOCO solvers for MOVRPs. First, we reformulate 16 single-objective VRP variants to bi-objective VRPs. Subsequently, we adapt two state-of-the-art single-task NMOCO solvers, WE-CA and POCCO, to the multi-task setting by extending their node representations and incorporating dynamic routing attributes into their context embeddings, yielding W-T and P-T, respectively. We also adapt MVMoE, a representative multi-task single-objective NCO solver, to the multi-task NMOCO settings by embedding the preference vector as a pseudo-node and using it to modulate node representations via FiLM, yielding MOM, and further strengthening of its transferred model with a decoder-side nonlinear expert layer yields the MOM-M. Furthermore, building on MOM-M, we propose the Decoupled Constraint Activation Signature Tokenization (DCAST), which factorizes a five-dimensional binary constraint activation signature into five learnable constraint identity embeddings and two shared learnable activation-state embeddings to construct state-aware constraint tokens that are jointly encoded with node and preference embeddings. Lastly, we extend the above solvers with a size-agnostic training paradigm to further boost their cross-size generalization. Extensive experiments on the reformulated MOVRPs demonstrate the superiority of the multi-task paradigms over their original single-task counterparts in terms of both cross-problem and cross-size versatility, as well as the superiority of the proposed DCAST among multi-task solvers. The source code is available at: https://github.com/DFSDarren/DCAST.

Primary area:
optimization
Keywords:
Neural multi-objective combinatorial optimization, multi-task neural combinatorial optimization, decoupled constraint activation signature tokenization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vnwiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper mainly focuses on the direction of multi-objective learning in combinatorial optimization. In specifics, it focuses on the multi-task training of different kinds of VRPs in a unified model. They further provide a tokenization mechanism for processing different kinds of the constraints in the VRP family. Extensive experiments validated their proposed model's utility and effectivenss. Also, one section regarding to the ablation stuidies of tokenization mechanism also demonstrates how their tokenization behaves compared to other choices.

Critical strengths
  1. The motivation is clear. The paper mainly generalizes the current single-task training paradigm into the multi-task training. This is a step towards the more challenging and general setting.

  2. The writing of this paper is very easy to be followed.

  3. The experiments are complete, like the main comparisons with the baselines. Also, they include one ablation study for comparing the effectiveness of the proposed tokenization mechanism with other choices.

Critical weaknesses
  1. The framework figure seems not to be the vector figure (when I enlarge it, it seems to degrade the resoluton a lot). I suggest replacing current one with the one drawed in PPT/Latex platform so that the produced figure will be in the vector format.

  2. Based on my knowledge of tokenization, this module should also handle the unseen constraint types. However, your current formulation seems to be restricted in the given constraint types. I am worrying whether this mechanism can be used for handling the types that are not seen in the training stage. For instance, in the paper of Router-Finder, they also have one extra constraint type called "Mixed-Backhauls", which is not covered in the training stage.

  3. As far as I know, the MVMoE's repo also implements quite similar mechanism like CONSTRAINT ACTIVATION SIGNATURE in your paper. They also use 1 and 0 to indicate whether the current constraint type exists. So I think you should somehow cite this from MVMoE.

Questions for authors

Will the tokenization mechanism affect your model's generalization ability to the unseen constraint types? If your model meets one unseen type of constraint, how will it handle the case?

3 comments
AuthorsOpenReview ↗

Response to Reviewer vnwi (1/2): Vector Figure, Relation to MVMoE, and a Tokenization Ablation

We sincerely thank the reviewer for the careful reading and for the three concrete suggestions, which we address one by one.

W1 — the framework figure should be a vector figure. We agree and apologize for the raster version in the submitted PDF. Fig. 1 has been redrawn as a genuine vector graphic (compiled from SVG through LaTeX) and replaced by that version in the updated PDF, where it stays sharp at any zoom level and at print size.

W3 — relation to the binary constraint vector in the MVMoE codebase. We agree that the credit is due, and MVMoE's paper together with its official repository will be cited explicitly in the revised version. The relation is as follows. The released MVMoE implementation accompanies each variant with a binary vector over the same five constraint types (capacity, open route, backhaul, route-length limit, time window), threaded through the code as an optional conditioning input, e.g. for the problem-level gating path. Our signature aτ=[aτC,aτO,aτB,aτL,aτTW]∈0,15\mathbf{a}_{\tau}=[a_{\tau C},a_{\tau O},a_{\tau B},a_{\tau L},a_{\tau TW}]\in\\{0,1\\}^{5} follows the same convention, so the common ground is exactly as the reviewer describes. The difference lies in how the vector is consumed:

  1. Role. In MVMoE the vector conditions the gating path. In DCAST it never enters the expert gates: it only selects one of two shared activation-state embeddings, zτk=Wz[ek,∣,saτk]+bz\mathbf{z}_{\tau k}=\mathbf{W}_z[\mathbf{e}_k\\,\\|\\,\mathbf{s}_{a_{\tau k}}]+\mathbf{b}_z with Eid=[eC,…,eTW]∈R5×d\mathbf{E}_{\mathrm{id}}=[\mathbf{e}_C,\dots,\mathbf{e}_{TW}]\in\mathbb{R}^{5\times d} and Estate=[s0,s1]∈R2×d\mathbf{E}_{\mathrm{state}}=[\mathbf{s}_0,\mathbf{s}_1]\in\mathbb{R}^{2\times d}; the five resulting tokens are encoded jointly with the node and preference embeddings and remain individually addressable in the decoder.
  2. What is learned. Identities are per constraint while activation states are shared by all five, so a variant is factorized into what the constraint is and whether it applies, and an unseen variant becomes a new composition of learned parts. A vector consumed as a plain feature instead ties the representation to the bit patterns seen during training.
  3. Evaluation target. 10 of the 16 combinations are held out from training, so the unseen-variant results are zero-shot on unseen compositions.

Ablation for this point (run for this response). We tested the two designs experimentally. Besides DCAST, whose token is defined above, we trained (i) a direct-signature variant, which appends the single token

zτdir=Wsaτ+bs\mathbf{z}^{\mathrm{dir}}_{\tau}=\mathbf{W}_s\mathbf{a}_{\tau}+\mathbf{b}_s

instead of the five factorized tokens (encoder input length n+3n+3 instead of n+7n+7), and (ii) a per-constraint-state variant, which replaces the shared states by ten per-constraint states and keeps decoupling but removes cross-constraint sharing:

Estate=sk,0,sk,1k∈C,O,B,L,TW∈R5×2×d,zτk=Wz[ek,∣,sk,aτk]+bz\mathbf{E}_{\mathrm{state}}=\\{\mathbf{s}_{k,0},\mathbf{s}_{k,1}\\}_{k\in\\{C,O,B,L,TW\\}}\in\mathbb{R}^{5\times 2\times d},\qquad \mathbf{z}_{\tau k}=\mathbf{W}_z[\mathbf{e}_k\\,\\|\\,\mathbf{s}_{k,a_{\tau k}}]+\mathbf{b}_z

All three are trained identically and evaluated on the 16 variants at n=50n=50:

variantconstraint tokensstate parametersavg. HV
DCAST52×d2\times d0.4014
direct signature100.4013
per-constraint states55×2×d5\times 2\times d0.4013

DCAST is best in aggregate and the most consistent (best or tied-best on 12 of the 16 variants, versus 2 and 3 for the two variants above), while the direct-signature variant is the closest counterpart of consuming the 0/1 vector as a feature. Hence neither the raw vector nor per-constraint states improve over the factorized tokens with shared states.

AuthorsOpenReview ↗

Response to Reviewer vnwi (2/2): Unseen Constraint Compositions and Unseen Constraint Types

Q — Does the tokenization affect the model's generalization to unseen constraint types, and how would it handle such a type if it met one? Two cases need to be distinguished.

(a) Unseen combinations of the five constraint types. This case is supported and is what we evaluate: a single policy trained on six variants is applied zero-shot to the remaining ten at n=50n=50 and n=100n=100 and, with size-agnostic training, at n=200n=200, where the size-agnostic model attains the best average Gap among the compared multi-task solvers (−0.40-0.40%). The reason is visible in the token construction: a bit of aτ\mathbf{a}_{\tau} does only two things — it selects s0\mathbf{s}_0 or s1\mathbf{s}_1 and indexes ek\mathbf{e}_k — so an unseen combination is expressed as a new composition of learned parts rather than as an unseen task-level signature.

(b) A genuinely new constraint type. We are explicit that the current interface is a five-dimensional signature and therefore cannot address a sixth constraint type. A variant that merely relaxes the semantics of an existing constraint (e.g. RouteFinder's Mixed-Backhauls, which still falls under the backhaul bit) remains expressible; a constraint without a bit does not. This is an interface choice rather than an architectural limitation, and the decoupling keeps the extension as small as it can be:

  • one identity row, no new state parameters: a new type changes Eid\mathbf{E}_{\mathrm{id}} from 5×d5\times d to 6×d6\times d (that is +d+d parameters), while s0\mathbf{s}_0 and s1\mathbf{s}_1 are reused unchanged, because they were never tied to a particular constraint identity; the learned notion of "applies / does not apply" therefore transfers to a constraint never seen;
  • architecture unchanged: the token construction and the sequence length n+7n+7 are untouched, so the encoder, the decoder and the decoder-side expert layer need no modification;
  • feasibility is enforced by masks, not by tokens: infeasible moves receive −∞-\infty from the feasibility mask, which the environment constructs deterministically from the constraint definition. A new type therefore needs its mask logic — as for any neural solver — while the token only tells the policy which problem it faces. This separation is also why zero-shot solutions on unseen combinations remain feasible by construction;
  • contrast with the raw-vector design: enlarging a signature that is consumed as a feature changes the input dimension of its embedding layer, and the new entry carries no learned meaning at initialization, on top of the bit-pattern binding discussed above.

We would like to be transparent about the scope of this evidence: we have not trained on a constraint type outside the five bits, so our answer here is structural rather than empirical. We nevertheless hope the above shows why the tokenization does not reduce the model's ability to generalize, and how a genuinely new constraint type would be accommodated.

These added responses and experimetns address my concerns. The authors also admit their weakness and limitaitons. I think this is ok. I maintain my score. But I still suggest that authos should pursue how to generalize current paradigms into unseen constraint types, this is a more promising direction.

Reviewer_doxCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multi-task neural multi-objective vehicle routing. The authors formulate 16 bi-objective VRP variants and adapt WE-CA, POCCO, and MVMoE to a shared multi-task setting. Building on MOM-M, they propose DCAST, which represents each routing constraint using a learnable identity embedding and a shared activation-state embedding. The resulting constraint tokens are encoded jointly with customer and preference representations.

The paper also introduces size-agnostic training and evaluates the models on six training variants, ten unseen constraint combinations, and larger instances with 200 customers. The results suggest that multi-task learning offers substantial benefits over task-specific models. DCAST improves performance in several settings, while DCAST-S achieves the lowest average HV Gap in the cross-size experiment.

Critical strengths

S1. The paper addresses a relevant problem.

Existing neural multi-objective solvers often focus on individual routing variants, while multi-task VRP methods primarily address single-objective optimization. Combining multi-task learning with preference-conditioned multi-objective routing is a meaningful research direction.

S2. The experimental framework is useful.

The paper establishes a unified evaluation setting with 16 bi-objective VRP variants and a clear split between observed and unseen constraint combinations. The adaptations of WE-CA, POCCO, and MVMoE also provide useful baselines for studying multi-task neural multi-objective optimization.

S3. The results provide evidence for the value of multi-task learning.

The multi-task models generally outperform the Bi-CVRP-trained single-task checkpoints on unseen variants. DCAST-S achieves an average Gap of -0.40% at N=200, compared with -0.32% for MOM-MS. The inclusion of training-time comparisons and CVRPLIB experiments further strengthens the empirical study.

Critical weaknesses

W1. The novelty and necessity of the proposed tokenization remain insufficiently established.

DCAST extends MOM-M by introducing explicit constraint tokens into the encoder. The proposed factorization combines constraint identity embeddings with shared activation-state embeddings through a linear projection.

The design is reasonable, but its advantage over simpler constraint-conditioning schemes is not clear. Equation (4) can be decomposed into an identity term and a state term. Consequently, the initial token change caused by activation is identical across constraint types. Later attention layers may learn constraint-specific interactions, but this does not establish why the shared factorization itself should improve generalization.

A comparison with direct binary constraint embeddings and constraint-specific state embeddings is needed.

Evidence: Equation (4), p. 5, and Section 4.4, pp. 8–9.

W2. The improvements over the closest baseline are small and inconsistent.

MOM-M is the most direct baseline for evaluating DCAST. According to Table 1, DCAST improves the average Gap by 0.07 and 0.02 percentage points on the training variants at N=50 and N=100. The corresponding differences on unseen variants are 0.15 and -0.01 percentage points. DCAST therefore does not consistently outperform MOM-M.

At N=200, DCAST-S improves the average Gap over MOM-MS by 0.08 percentage points. These differences require statistical validation, yet the paper does not report variability across independent training runs.

Evidence: Table 1, p. 6, Table 4, p. 9, and Appendix E, p. 16.

W3. The ablation study does not isolate the main contribution.

Figure 3 compares DCAST with several alternative architectures, but the evaluation covers only four variants at N=50. It does not directly test whether sharing activation-state embeddings is better than learning separate states for each constraint.

The study also does not establish whether the gains come from the proposed factorization, the additional constraint tokens, or increased representational capacity. Since the central claim concerns generalization to unseen constraint combinations, the ablation should include these settings.

Evidence: Section 4.4 and Figure 3, pp. 8–9.

Questions for authors

Q1. Could the authors compare DCAST with direct five-dimensional binary constraint embeddings, constraint-specific state embeddings, and MOM-M without additional constraint tokens? These comparisons should use matched backbones, model capacities, and training protocols. Results on the ten unseen variants would be particularly helpful.

Q2. In Equation (4), the activation-state difference is shared across all constraint types. What benefit does this shared parameterization provide over constraint-specific representations? Has this been tested directly?

Q3. Could the authors report results across multiple independent random seeds for DCAST, MOM-M, DCAST-S, and MOM-MS? Please also specify the HV reference points and normalization procedure.

Q4. Could the authors verify the MOM-MS results on Set-Solomon? The R103 Gap in Table 11 appears inconsistent with the reported HV values and may explain the discrepancy between Tables 9 and 11.

2 comments
AuthorsOpenReview ↗

Response to Reviewer doxC(part 1/2): two new ablations; W1 / Q1

We thank the reviewer for the careful review. We trained two ablations that isolate the two design choices of Eq. (4) and evaluated both on all 16 variants at n = 50. Their results are in the table below; the HV reference points and the normalization procedure follow in part 2/2, together with the R103 issue.

SigDirect (direct signature embedding). It keeps the MOM-M backbone but removes the factorization of Sec. 3.3: the raw five-dimensional binary signature is projected by a single linear layer Ws∈Rd×5W_s \in \mathbb{R}^{d\times 5} and appended as one token after the preference embedding, H0=[h00,…,hn0,hλ0,Wsaτ+bs]H^0 = [h^0_0, \ldots, h^0_n, h^0_\lambda, W_s a_\tau + b_s] (768 parameters). It is exactly the "signature used directly as a continuous encoder feature" alternative of Sec. 3.2.

StateBind (constraint-specific activation states). It keeps the five tokens, the five identity embeddings and the projection of Eq. (4), but replaces the two shared states Estate∈R2×dE_{\mathrm{state}} \in \mathbb{R}^{2\times d} by a constraint-specific table S^∈R5×2×d\hat{S} \in \mathbb{R}^{5\times 2\times d}, i.e. zτk=Wz[ek∥saτk]+bzz_{\tau k} = W_z[e_k \Vert s_{a_{\tau k}}] + b_z. Each constraint thus owns its own inactive/active pair (ten states instead of two); StateBind differs from DCAST by 1,024 parameters only. Both states of every constraint occur in training — each of C, O, B, L, TW appears active and inactive in the six training variants — so StateBind is not handicapped by untrained states and is a fair constraint-specific baseline.

HV (higher is better) on all 16 variants at n = 50, the first six rows being the training variants and the remaining ten the zero-shot unseen ones; best per row in bold:

VariantDCASTSigDirectStateBind
Bi-CVRP0.39930.39920.3991
Bi-OVRP0.44090.44070.4406
Bi-VRPB0.40700.40670.4069
Bi-VRPTW0.34420.34450.3444
Bi-VRPL0.39750.39740.3975
Bi-OVRPTW0.41460.41460.4149
Bi-OVRPB0.44160.44130.4414
Bi-VRPBL0.42470.42430.4243
Bi-VRPLTW0.34670.34630.3464
Bi-OVRPBTW0.40920.40910.4090
Bi-VRPBLTW0.33950.33900.3395
Bi-OVRPL0.44400.44370.4437
Bi-VRPBTW0.33250.33200.3322
Bi-OVRPBL0.45710.45670.4568
Bi-OVRPLTW0.41260.41280.4126
Bi-OVRPBLTW0.41170.41170.4118
Avg.0.40140.40130.4013

Three observations. (i) Neither simplification improves on DCAST, which is best or tied-best on 12 of 16 variants (SigDirect 2, StateBind 4). (ii) The ranking holds on the ten unseen variants — best or tied-best on 8 of 10, strictly better than SigDirect on 8 and than StateBind on 7 — so it is not a memorization of seen bit patterns. (iii) The three tokenizations are close in absolute terms (largest per-variant margin 0.0005 HV), whereas MOM-M (same backbone, no token) is below DCAST in four of the five aggregate cells of Table 1 (avg. Gap −0.10%/−0.22% vs. −0.17%/−0.37% at n = 50 trained/unseen).

W1 / Q1 — novelty, necessity, and the requested comparisons. We agree with the reviewer's reading of Eq. (4): since WzW_z is linear, zτk=Wzidek+Wzstsaτk+bzz_{\tau k} = W_z^{id} e_k + W_z^{st} s_{a_{\tau k}} + b_z, so activation applies a common shift. This is the intended bias. (1) The identity term eke_k distinguishes constraint types; the state term only says whether a constraint applies, and "active" should mean the same for C, O, B, L, TW, since the six training variants cover only 6 of 32 compositions and a per-constraint activation semantics would be estimated from far fewer variants per state. (2) Constraint-specific behaviour is not removed: each token is a distinct vector, and in every layer the joint self-attention over the n+2+5n+2+5 tokens lets node and preference keys decide how the shared shift affects the representation. (3) This is now tested rather than argued: StateBind removes exactly that sharing and does not improve; SigDirect removes the factorization and the tokens and does not improve either. Since MOM-M stays below DCAST, signature-derived tokens are necessary, and among the tested tokenizations the decoupled-and-shared one is the best-performing and the most parameter-efficient (2 vs. 10 states, 1,024 fewer parameters). The table above contains exactly the direct-binary-embedding, constraint-specific-state and no-token comparisons requested, with matched backbones and matched capacity (at most 1,024 parameters apart).

AuthorsOpenReview ↗

Response to Reviewer doxC(part 2/2): W2, W3, Q2–Q4

W2 — magnitude and consistency. We concede that the margins over MOM-M are small (0.07 and 0.15 points on the trained and unseen variants at n = 50, 0.02 at n = 100, 0.01 at n = 200) and we do not claim a large margin; the claim is the strongest aggregate performance with a single policy across 16 variants. (i) The direction is consistent: DCAST is better in four of the five aggregate cells, the single exception being n = 100 unseen (−0.33% vs. −0.34%), matching the per-variant picture of part 1/2. (ii) The comparison is paired: all methods use the same 200 test instances per variant, the same 101 preference vectors (20,200 solutions per cell), the same reference points and inference code, so the reported differences are paired differences on identical data. (iii) The gain is in transfer: one shared policy is best or tied-best on 12/16 variants, at 43.61 h at n = 50 versus 56.16 h for the six task-specific WE-CA models (Table 8).

W3 — isolating the main contribution. We agree that the comparison in Fig. 3 (four variants, n = 50) is too narrow. The table in part 1/2 covers all 16 variants, including the ten unseen ones, and separates the two design axes with matched backbone, token interface and protocol (capacity differs by at most 1,024 parameters). Combined with the four architecture ablations already reported, the design axis now reads: no constraint conditioning (MOM, MOM-M) → direct signature (SigDirect) → constraint-specific states (StateBind) → decoupled and shared (DCAST).

Q2 — what the shared state parameterization provides. This is tested directly by StateBind, which gives each constraint its own pair of states while keeping the identity embeddings, the token construction and the projection unchanged; it does not improve over the two shared states. Mechanistically, sharing provides (i) one well-estimated notion of "active", updated by every active constraint in every mini-batch (batches are drawn from a single variant, Appendix D), instead of five sparsely updated copies; (ii) a state vocabulary of size 2 rather than 10, so an unseen variant is by construction a new composition of seen primitives; (iii) 2 instead of 10 states, i.e. 1,024 fewer parameters.

Q3 — seeds, reference points and normalization. Every model — single-task, multi-task, size-agnostic and both new ablations — uses the same seed (2024, the MvMoE seed) and the identical recipe (Adam, 200 epochs, lr 3×10−43\times 10^{-4}, weight decay 10−610^{-6}, 100k instances per epoch, batch size 64, Dirichlet(1,1) preferences, the same six training variants and batch composition). No seed is tuned or selected per method, so no method is seed-optimized and the comparison is fair by construction; evaluation is deterministic given the fixed test sets.

Normalization. Both objectives are in normalized units, so that HV is comparable across variants and sizes: coordinates are scaled to [0,1][0,1] (synthetic instances drawn in [0,1]2[0,1]^2, CVRPLIB coordinates divided by 1000, Solomon instances by an instance-specific scaler) and demands are divided by the vehicle capacity. For each test instance, the 101 solutions obtained for the 101 preference vectors are filtered to their non-dominated set, and the exact hypervolume of that set is computed with the WFG algorithm (hvwfg). The hypervolume is then divided by the volume of the reference box, i.e. by the product of the reference-point coordinates with ideal point (0,0)(0,0), yielding an HV in [0,1][0,1]; the reported HV is the mean over the 200 test instances.

Reference points (identical for all methods): (45,4)(45,4) for n = 50, (80,4)(80,4) for n = 100 and (120,6)(120,6) for n = 200; for the CVRPLIB subsets the reference points are set-specific, namely (15,3)(15,3) for Set-A, (10,2)(10,2) for Set-B, (10,2)(10,2) for Set-E, (10,3)(10,3) for Set-P, (300,10)(300,10) for Set-X and (80,4)(80,4) for Set-Solomon. Gap is the normalized HV difference relative to the reference model of that comparison, Gap=100×(HVref−HV)/HVref\mathrm{Gap} = 100 \times (HV_{\mathrm{ref}} - HV)/HV_{\mathrm{ref}}, with HVref=MOMHV_{\mathrm{ref}} = \mathrm{MOM} for n ≤ 100 and HVref=MOM−MHV_{\mathrm{ref}} = \mathrm{MOM-M} for n = 200 and the CVRPLIB tables; a negative Gap means a larger HV than the reference.

Q4 — R103 in Table 11. Thank you for spotting this. It is a transcription error: the MOM-MS entry for R103 should read 0.08%, not 8.00% ((0.5267 − 0.5263)/0.5267 = 0.076%). We have corrected it in the newly uploaded PDF (marked in blue), and we re-checked every HV/Gap pair in Tables 10–15 for the same kind of inconsistency. The aggregates in Table 9 were computed from the per-instance HV values and are therefore unaffected by this transcription error; no conclusion changes.

Reviewer_kG6bOpenReview ↗
Rating
2(+1)
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates multi-task neural multi-objective combinatorial optimization for 16 bi-objective vehicle-routing variants formed from capacity, open-route, backhaul, route-length-limit, and time-window requirements. The two objectives are total routing distance and maximum route length. The authors adapt WE-CA and POCCO to form multi-task baselines (W-T and P-T) and extend MVMoE to preference-conditioned multi-objective models (MOM and MOM-M). Their main proposed component, DCAST, represents each variant with a five-bit constraint-activation signature and constructs a separate token for each constraint by combining a learnable constraint-identity embedding with one of two shared activation-state embeddings. The resulting tokens are jointly encoded with node and preference representations. The experiments train on six variants, evaluate zero-shot transfer to ten unseen constraint combinations at 50 and 100 customers, and assess transfer to 200-customer instances with additional size-agnostic training. Performance is evaluated mainly by hypervolume, HV-based gaps, and runtime.

Critical strengths
  1. A relevant and well-scoped evaluation setting. Unifying multi-task learning and preference-conditioned bi-objective VRP optimization is potentially useful. The benchmark explicitly distinguishes seen and unseen constraint combinations, making cross-variant transfer testable.
  2. Useful breadth of empirical comparisons. The paper adapts multiple existing multi-objective and multi-task neural solvers and includes separate fixed-size and cross-size evaluations.
Critical weaknesses
  1. Insufficient methodological differentiation from existing constraint-aware multi-task NCO. Representing VRP variants as compositions of basic constraints and conditioning a shared encoder on their activation states is already central to prior work. In particular, CaDA (ICML 2025) uses a five-dimensional constraint activation vector and an encoder-side constraint prompt; RouteFinder (TMLR 2025) uses global attribute embeddings for variant-aware representations; and CoEKS (ICLR 2026) exploits combinations of constraint-specific components for zero-shot transfer. DCAST's per-constraint tokenization is architecturally distinct, but the paper does not convincingly explain why the proposed identity/state factorization is preferable to these simpler alternatives. Moreover, Eq. (4) is an affine projection of concatenated embeddings: writing it as z(k,a)=Widek+Wstatesa+bz(k,a) = W_{id} e_k + W_{state} s_a + b gives z(k,1)−z(k,0)=Wstate(s1−s0)z(k,1) - z(k,0) = W_{state}(s_1 - s_0), the same initial activation shift for every constraint. Subsequent attention can model constraint-specific interactions, but that alone does not establish the claimed benefit of decoupling. A matched-backbone comparison with a single global-constraint prompt, independently learned embeddings for each constraint/state pair, and equally sized multi-token controls is essential.
  2. The main component's empirical benefit is small, scale-dependent, and not convincingly separated from other gains. MOM-M already includes the decoder-side nonlinear expert layer, so DCAST should primarily be evaluated against MOM-M rather than against much weaker single-task transfers. In Table 1, DCAST improves unseen-task average HV Gap over MOM-M at n=50 (-0.37% vs. -0.22%), but is slightly worse at n=100 (-0.33% vs. -0.34%); at n=200 the difference is negligible (-0.01% vs. 0.00%). From Table 3's reported HV values, at n=100 DCAST is better than MOM-M on four unseen variants, equal at reported precision on three, and worse on three. While DCAST-S improves on MOM-MS under the size-agnostic protocol (average Gap: -0.40% vs. -0.32%), this is a relatively modest gain and incurs additional inference costs on several tasks (Table 4). Figure 3's architecture ablations cover only four n=50 variants, of which just one is unseen, limiting evidence for the central compositional generalization claim. Variability across training seeds, uncertainty for small differences, and a controlled test of the particular factorization are needed. Further, the large differences relative to WE-CA and POCCO on unseen tasks should be interpreted with caution, as their transferred checkpoints were trained only on Bi-CVRP, whereas the multi-task methods were trained on six variants. As presented, the experiments establish the value of broader multi-task exposure more clearly than the necessity of DCAST itself.
Questions for authors
  1. Mechanism / necessary ablation. Within the same MOM-M backbone, training tasks, preference sampling, and compute budget, can the authors compare (i) a CaDA-style single-constraint prompt from the five-bit signature, (ii) five constraint tokens with independently learned active/inactive embeddings for each constraint, and (iii) DCAST's shared identity/state factorization?
  2. Robustness of the claimed gains. Why does DCAST's clear unseen-task advantage at n=50 largely disappear at n=100 and n=200 (Tables 1 and 3)? Could the authors also report a broader unseen-combination ablation, rather than only the single unseen task in Figure 3?
  3. Scope and comparison fairness. Since CVRP-only single-task checkpoints and six-task multi-task policies have different task exposure, which conclusions in the paper specifically support DCAST's token mechanism rather than multi-task training? Can the authors include or implement a capacity-matched, constraint-aware baseline (e.g., CaDA-style prompting or a CoEKS-style constraint-conditioned module) under the same bi-objective setup? Results from this controlled comparison could materially change the assessment.
5 comments
AuthorsOpenReview ↗

Response to Reviewer kG6b (1/2): Weakness 1 / Question 1

We thank the reviewer for a detailed report. This comment addresses Weakness 1 and Question 1; the second addresses Weakness 2 and Questions 2–3.

Differentiation. We thank the reviewer for pointing to CaDA (Li et al., ICML 2025). CaDA, RouteFinder and CoEKS are single-objective cross-problem solvers that condition a task-agnostic policy on constraint or attribute information and output one tour. Two differences matter. First, our policy is also conditioned on a continuous preference vector and must approximate a Pareto front, so constraint information must be fused with preference conditioning (FiLM, preference token, decoder context), not supplied as one prompt. Second, CaDA enumerates all 32 five-bit combinations as prompts, i.e. it operates on variants it trained on, whereas our policy trains on 6 of 32 signatures and is evaluated zero-shot on the other 10, where a single prompt is a task label, not a compositional parameterization; DCAST instead uses five tokens, each combining a constraint identity with one of two shared activation states.

On Eq. (4). The algebra is correct: since the projection is affine, z(k,1)−z(k,0)=Wstate(s1−s0)z(k,1)-z(k,0)=W_{\mathrm{state}}(s_1-s_0) is the same vector for every kk — the identity term sets where a constraint sits, the shared term moves every constraint along one common activation direction. This is the intended inductive bias: one learned axis for "active", so that 32 signatures are spanned by 7 embedding vectors and an unseen composition is a recombination of individually trained directions. We therefore claim compositionality and parameter economy, not added expressiveness — and the controls test exactly that: both alternatives are strictly more flexible (a free 5-to-dd projection in SigDirect, ten independent states in StateBind) and neither is better.

The three controls of Question 1 share the MOM-M backbone, the six training variants, Dirichlet(1,1) sampling, the optimizer and the budget; only the constraint representation differs.

  • SigDirect: the five-bit signature embedded and supplied as a single constraint prompt — the CaDA-style prompt, i.e. the capacity-matched constraint-aware baseline of Question 3.
  • StateBind: five tokens, but each constraint owns its own inactive/active pair (ten states); it keeps the five-token layout, so it is also the equally sized multi-token control, and uses 15 embeddings instead of 7.
  • DCAST: five tokens, each combining one of five identity embeddings with one of two shared state embeddings.

HV (↑\uparrow), identical backbone, tasks, preference sampling and budget; † marks the ten variants evaluated zero-shot, the other six being the training variants; bold is the best entry per row.

VariantDCASTSigDirectStateBind
Bi-CVRP0.39930.39920.3991
Bi-OVRP0.44090.44070.4406
Bi-VRPB0.40700.40670.4069
Bi-VRPTW0.34420.34450.3444
Bi-VRPL0.39750.39740.3975
Bi-OVRPTW0.41460.41460.4149
Bi-OVRPB †0.44160.44130.4414
Bi-VRPBL †0.42470.42430.4243
Bi-VRPLTW †0.34670.34630.3464
Bi-OVRPBTW †0.40920.40910.4090
Bi-VRPBLTW †0.33950.33900.3395
Bi-OVRPL †0.44400.44370.4437
Bi-VRPBTW †0.33250.33200.3322
Bi-OVRPBL †0.45710.45670.4568
Bi-OVRPLTW †0.41260.41280.4126
Bi-OVRPBLTW †0.41170.41170.4118

This also answers the broader unseen-combination ablation of Question 2: all 16 variants, not a single unseen one. DCAST is best or tied-best on 12 of the 16 (8 of the 10 unseen) with the fewest parameters. On the six trained variants the three lie within 0.6×10−40.6\times10^{-4} HV of each other; on the ten unseen compositions the distance to DCAST grows to 2.7×10−42.7\times10^{-4} (SigDirect) and 1.9×10−41.9\times10^{-4} (StateBind), four- and fifteen-fold. So (a) a single global-constraint prompt is not sufficient: it is the most flexible map from signature to representation, sees the same six signatures, and improves nothing; (b) the shared activation direction is not the bottleneck either: StateBind removes the sharing, adds eight state vectors, and is the weakest in aggregate; (c) what remains distinct is that DCAST's 7 vectors are compositional, and the ordering appears, and is largest, on the compositions never trained. Differences of this size are close to run-to-run variability, and we make no superiority claim at n≥100n\ge100.

AuthorsOpenReview ↗

Response to Reviewer kG6b (2/2): Weakness 2 / Questions 2–3

What this paper contributes, and how the MOM-M comparison should be read. We would like to make the structure explicit, because it determines the meaning of the MOM-M comparison. Before this work, multi-task learning did not exist in NMOCO: WE-CA and POCCO are single-task solvers, while multi-task NCO (MVMoE, RouteFinder, CaDA, CoEKS, POMO-MTL) is single-objective. This paper brings multi-task NCO into the neural multi-objective setting and populates the paradigm with five models that do not exist in prior work: W-T and P-T (multi-task adaptations of WE-CA and POCCO through unified static node features and dynamic routing attributes in the context embedding), MOM (MVMoE transferred to preference-conditioned multi-objective routing, the preference vector being used as a pseudo-node and to modulate node representations via FiLM), MOM-M (MOM plus our decoder-side nonlinear expert layer) and DCAST (the constraint tokenization). MOM-M and DCAST are therefore both contributions of this paper rather than baselines taken from the literature, and comparing DCAST with MOM-M — same backbone, same training tasks, same preference sampling, same budget — is already the strictest available test, since the two differ only in the constraint representation. DCAST changes seven embedding vectors and one projection on an already preference-conditioned MoM-M backbone, so its effect on the aggregate Gap is necessarily modest; the claim it supports is about how a variant is represented, and its evidence is the fixed-backbone ablation in the first comment and the single-variant transfer below.

On scale dependence. We agree that the aggregate differences at larger sizes are not a superiority claim: relative to MOM-M, the unseen-variant Gap is −0.37%-0.37\% versus −0.22%-0.22\% at n=50n=50, −0.01-0.01 pp at n=100n=100 and 0.010.01 pp at n=200n=200. What is stable across variants is the unseen-composition ordering of the ablation table in our first comment, the size-agnostic transfer (DCAST-S attains −0.40%-0.40\% against −0.14%-0.14\% for MOM-S and −0.32%-0.32\% for MOM-MS at n=200n=200, best or tied-best on 12 of 16 variants), and the cost of one shared policy (79.35 h versus 96.41 h for six task-specific WE-CA models and 382.56 h for six POCCO models).

Exposure versus mechanism (Question 3). To separate the tokenization from multi-task exposure, DCAST was trained on Bi-CVRP only — one variant, one signature, the same task exposure as the WE-CA and POCCO checkpoints it is compared with in Table 2 — and evaluated zero-shot on all 16 variants.

The table below reports the resulting average Gap (%, ↓\downarrow) at n=50n=50; reference and protocol are those of Table 1, whose WE-CA and POCCO rows are reproduced exactly, and bold marks the best entry per column.

Method (trained on Bi-CVRP only)Trained variantsZero-shot variants
WE-CA3.316.14
POCCO2.685.05
DCAST-Bi-CVRP2.875.23

Three conclusions follow. (i) With exposure matched, the tokenized solver improves over WE-CA by 0.44 pp on the trained variants and by 0.91 pp on the ten zero-shot variants, and trails the strongest single-task NMOCO solver, POCCO, by only 0.19 pp and 0.18 pp. The three rows differ in nothing but the solver, so this isolates the constraint representation from the number of training variants — exactly the concern raised — and shows that the compositional tokenization carries a large part of the cross-variant behaviour on its own, without multi-task training. (ii) Adding multi-task exposure to the same architecture moves the zero-shot Gap from 5.23% to −0.37%-0.37\%, while one policy covers all 16 variants and, with size-agnostic training, sizes up to 200. The two effects are complementary and separately measurable, which is why we report them as two numbers rather than as one claim about DCAST. (iii) The price of generality is small and explicit: 0.19/0.18 pp relative to the best single-task solver, in exchange for one shared policy instead of six task-specific models whose training takes 382.56 h.

Thank you for the additional experiments. I remain unconvinced that DCAST provides a consistent performance advantage over POCCO-based alternatives, and several reported values require clarification.

  1. Your Bi-CVRP-only experiment at n=50 reports aggregate Gaps of 2.68% and 5.05% for POCCO, versus 2.87% and 5.23% for DCAST-Bi-CVRP. The results as reported favor POCCO in both evaluation groups.

  2. P-T and DCAST use the same six training variants. With training at n=100 and testing at n=200, P-T outperforms DCAST on all 16 variants in Table 4, with average Gaps of −0.33% and −0.01%, respectively.

  3. Table 9 shows a lower average Gap for P-T on five of six subsets. On Set-Solomon, P-T obtains −1.38%, compared with 3.19% for DCAST.

  4. At n=100, Table 8 reports 61.22 hours for P-T and 79.35 hours for DCAST. The cited 382.56 hours covers six separately trained POCCO models; P-T already provides one shared policy.

Please provide the complete per-variant HV, Gap, and inference time for the newly added Bi-CVRP-only comparison on all 16 variants at n=50, including Bi-CVRP itself. Please also report the matched single-task training times.

I acknowledge DCAST’s stronger n=50 results and DCAST-S’s advantage over P-TS in the n=200 mixed-size-training comparison. Please explain the contrasting outcomes, resolve the numerical inconsistencies, and qualify the performance claims accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.