acceptodds
Under review as a conference paper at ICLR 2027

LaT: LLM-as-Trainer for Multi-Task Vehicle Routing Solvers

Abstract

Multi-task neural solvers aim to handle multiple Vehicle Routing Problem (VRP) variants within a unified model, avoiding separate training for each constraint combination. However, VRP variants differ in optimization difficulty, while existing methods lack stage-wise feedback on their training status, making the model biased to some specific variants. Although meta-learning can support adaptive training, it typically requires bi-level optimization and additional gradient updates, increasing computational cost. To address this limitation, we propose LLM-as-Trainer (LaT), a plug-and-play training paradigm that uses a pretrained large language model as an external trainer. LaT periodically analyzes cross-task validation metrics to generate a stage-wise guidance vector. This vector is combined with the current task’s constraint vector and injected into each encoder layer, providing the neural solver with additional training information during subsequent policy optimization. Experiments on 16 VRP variants show that LaT improves the solution quality of several state-of-the-art multi-task neural solvers on both trained and unseen variants, supporting the effectiveness and generality of the proposed training paradigm.

Primary area:
optimization
Keywords:
Vehicle Routing Problem, Large Language Models, Multi-Task Learning, Neural Combinatorial Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qpU1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes using a pretrained LLM as a high-level controller for multi-task VRP training. Every 500 epochs, recent validation statistics are summarized and sent to GLM-5.1. The model outputs five scalar values, one for each elementary constraint. These values are then concatenated with the binary task vector and injected into the encoder through learnable side branches. The proposed mechanism consistently improves several multi-task neural solvers, although the magnitude of improvement is usually moderate.

Critical strengths
  1. Existing multi-task VRP methods usually focus on architectures, task representations, experts, or gradients. It is a reasonable direction to explicitly process the state of multi-task training as an input signal.

  2. The method is demonstrated on several backbones rather than one specific architecture. This supports the claimed plug-and-play nature to some extent.

Critical weaknesses
  1. The training procedure explicitly states that, at every control round, the current solver is evaluated on all 16 validation tasks, and these validation results are supplied to the LLM to construct the subsequent guidance vector. However, only six variants are described as training variants, while the remaining ten are subsequently evaluated as “unseen VRPs” and claimed to demonstrate zero-shot generalization. Therefore the supposedly unseen task distributions have already influenced the training trajectory.

  2. I am not convinced that an LLM is necessary for this problem. The MLP baseline does not resolve this concern. The appendix reports a difference of only 0.1 percentage points. This is small, particularly because the paper does not provide variance over independent training runs. Furthermore, the proposed MLP baseline is somewhat unusual. Its parameters are updated only once per control round through the first training batch of the subsequent interval. This does not strike me as the strongest learned controller baseline. Moreover, the guidance-gap correlation is descriptive, but not mechanistic. It would be helpful to compare against explicit non-LLM control rules.

  3. Some reported SOTA margins are extremely small. For example, under the CaDA plug-and-play comparison, the improvement is only 0.0039 percentage points. Without repeated runs, I do not view this as convincing evidence of superiority. Moreover, the main experiments compare LaT against several recent state-of-the-art methods, such as ReLD, CaDA, and CCL, the evaluations on CVRPLIB Set-X and Set-Solomon in Appendix F.4 only include POMO, POMO-MTL, and MVMoE as baselines.

  4. The paper reports low additional training time for LaT itself, but LaT requires reference objectives used to compute gaps. For 14 variants these are generated offline using OR-Tools with a time limit of up to 200 seconds per instance, while HGS is used for CVRP/VRPTW.

Questions for authors

See Weaknesses

1 comment

Thank you to the authors for the detailed response and additional experiments. The new results address several of my concerns, but some issues remain unresolved.

W1. I appreciate the authors acknowledging that the original protocol was not strictly zero-shot. The new six-task-only experiment is encouraging, but it is limited to POMO-MTL at N=50. The strict zero-shot results for the main LaT-CaDA and LaT-ReLD variants are still missing. Furthermore, removing both the unseen-task feedback and the history simultaneously makes it difficult to isolate their individual effects.

W2. Contribution of the LLM. The added non-LLM baselines strengthen the evaluation. However, the improvements relative to these baselines remain relatively small and lack multi-seed validation. Therefore, it remains unclear how much of the gain specifically originates from the LLM-generated guidance rather than dynamic conditioning.

W3. Limited methodological insights. The performance improvements over strong baselines are often marginal, and the paper provides limited insight into why the LLM-generated guidance improves the optimization process. The guidance values are introduced as additional conditioning inputs, but it is not clearly explained how they reduce cross-task interference or enhance task-specific learning.

Overall, I acknowledge the authors' efforts and the improved empirical evidence. However, the remaining issues prevent me from increasing my score. I will maintain my rating.

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

This paper proposes LaT, using frozen LLM as an external trainer for multi-task VRP neural solvers: existing methods lack stage feedback and tend to be biased towards some variants. Every 500 epochs, LLM reads the validation gaps of 16 variants, outputs a 5-dimensional guidance vector, and concatenates it with the task constraint vector and then injects it into each encoder layer. After training, the LLM is removed. On 16 variants (6 used for training, 10 zero-shot), N=50/100, LaT lowers the average gaps of ReLD and CaDA, LaT-CaDA outperforms CaDA on all 20 unseen setups, and the paper also reports CVRPLIB Set-X and Solomon results.

Critical strengths
  1. LaT reduces the gap of both ReLD and CaDA on all 12 training settings, adds only about 0.2M parameters, and no LLM is required for inference.
  2. Covering 16 variants with two scales (including unseen) and complementing CVRPLIB Set-X (5.198% vs. 9.820%) and Solomon results.
Critical weaknesses
  1. No meaning is given for the five guidance values: the prompt does not explain how it affects training, and the network can interpret this set of inputs arbitrarily. The positive correlation in Fig. 4 only shows that LLM converts the gap into a number; 16 tasks share a γ, which cannot reflect the differences in tasks.
  2. The technical advance is somehow limited: it only adds a lightweight side branch driven by 5 scalars (about 0.2M parameters) to the existing backbone. It is only 0.1 percentage points ahead of the MLP generator, and the contribution is mainly empirical.
Questions for authors
  1. Please add a rule-based γ or random γ control given by the gap mean to distinguish whether the gains are from LLM judgments or slowly changing conditioning.
  2. Please report the validation set size |D_k| for each task and the total time spent on offline reference solutions, and give the guidance results using only 6 training variant gaps.
  3. Please give a measure of bias (such as gap dispersion between variants) and indicate whether LaT reduces said bias.
  4. After changing to another LLM or rewriting prompt, will the guidance trajectory and final gap be stable?
Reviewer_zhUEOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LLM-as-Trainer (LaT), a plug-and-play training method for multi-task neural VRP solvers. Every 500 epochs, LaT evaluates the solver on 16 VRP variants and gives the validation gaps to a frozen LLM (GLM-5.1). The LLM returns a 5-dim guidance vector, with one value for each elementary constraint (C, O, B, L, TW). This vector is concatenated with the task constraint vector and fed to zero-initialized side branches after each encoder layer. After training, the final vector is fixed and the LLM is removed, so inference needs no LLM calls. The method is applied to ReLD, CaDA, POMO-MTL, and MVMoE under the MVMoE protocol (training on 6 variants, zero-shot on 10). The paper reports lower average gaps than the backbones on both trained and unseen variants, with small training overhead.

Critical strengths

The paper reports its results honestly. It shows the settings where CCL or the backbone beats LaT (Tables 1, 2), and it states that the gains depend on the backbone (§5). The code is also released, covering all five backbones, the data, and the reference solutions.

Critical weaknesses

W1. The contribution of the LLM is not verified. All ablations use POMO-MTL only. None use CaDA or ReLD. A natural baseline is missing: skip the LLM and feed the normalized raw constraint-wise validation gaps directly as the guidance vector. The LLM guidance already tracks these gaps closely (ρ ≈ 0.81, App. C.3), so this baseline would likely perform similarly. The released code includes a related gap-based rule (_build_stat_control_guidance, used as a fallback). The paper reports neither this rule nor the simpler baseline.

W2. Small, inconsistent gains. The gains are marginal. On unseen VRPs, CaDA goes from 5.490% to 5.067% and ReLD from 5.328% to 5.240%. LaT-ReLD is worse in 6 of 20 settings (Table 2). The margin over PoMtVRS-CaDA is 0.004%p (Table 4). All results are single runs with no variance reported.

Questions for authors

Refer Weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.