Instance-Conditioned Adaptation for Large-scale Generalization

of Neural Combinatorial Optimization

Changliang Zhou * 1 Xi Lin * 2 Zhenkun Wang 1 Xialiang Tong 3 Mingxuan Yuan 3 Qingfu Zhang 2

Abstract sive heuristic algorithms, such as LKH3 (Helsgaun, 2017)

and HGS (Vidal, 2022), have been proposed to address
The neural combinatorial optimization (NCO)
arXiv:2405.01906v1 [cs.AI] 3 May 2024

different VRP variants. Although these approaches have

approach has shown great potential for solving
shown promising results for specific problems, the algo-
routing problems without the requirement of ex-
rithm designs heavily rely on expert knowledge and a deep
pert knowledge. However, existing constructive
understanding of each problem. It is very difficult to de-
NCO methods cannot directly solve large-scale
sign an efficient algorithm for a newly encountered problem
instances, which significantly limits their appli-
in real-world applications. Moreover, their required run-
cation prospects. To address these crucial short-
time can increase exponentially as the problem size grows.
comings, this work proposes a novel Instance-
As a result, these limitations greatly hinder the practical
Conditioned Adaptation Model (ICAM) for better
application of classical heuristic algorithms.
large-scale generalization of neural combinatorial
optimization. In particular, we design a powerful Over the past few years, different neural combinatorial op-
yet lightweight instance-conditioned adaptation timization (NCO) methods have been explored to solve
module for the NCO model to generate better so- various problems efficiently (Bengio et al., 2021; Li et al.,
lutions for instances across different scales. In 2022). In this work, we focus on the constructive NCO
addition, we develop an efficient three-stage re- method (also known as the end-to-end method) that builds a
inforcement learning-based training scheme that learning-based model to directly construct an approximate
enables the model to learn cross-scale features solution for a given instance without the need for expert
without any labeled optimal solution. Experimen- knowledge (Vinyals et al., 2015; Kool et al., 2019; Kwon
tal results show that our proposed method is ca- et al., 2020). In addition, constructive NCO methods usually
pable of obtaining excellent results with a very have a faster runtime compared to classical heuristic algo-
fast inference time in solving Traveling Salesman rithms, making them a desirable choice to tackle real-world
Problems (TSPs) and Capacitated Vehicle Rout- problems with real-time requirements. Existing constructive
ing Problems (CVRPs) across different scales. To NCO methods can be divided into two categories: super-
the best of our knowledge, our model achieves vised learning (SL)-based (Vinyals et al., 2015; Xiao et al.,
state-of-the-art performance among all RL-based 2023) and reinforcement learning (RL)-based ones (Nazari
constructive methods for TSP and CVRP with up et al., 2018; Bello et al., 2016). The SL-based method re-
to 1,000 nodes. quires a lot of problem instances with labels (i.e., the optimal
solutions of these instances) as its training data. However,
obtaining sufficient optimal solutions for some complex
1. Introduction problems is unavailable, which impedes its practicality. RL-
based methods learn NCO models by interacting with the
The Vehicle Routing Problem (VRP) plays a crucial role in environment without requiring labeled data. Nevertheless,
various logistics and delivery applications, as its solution due to memory and computational constraints, it is unrealis-
directly affects transportation cost and service efficiency. tic to train the RL-based NCO model directly on large-scale
However, efficiently solving VRPs is a challenging task due problem instances.
to their NP-hard nature. Over the past few decades, exten-
Current RL-based NCO methods typically train the model
*
Equal contribution 1 Southern University of Science and on small-scale instances (e.g., with 100 nodes) and then
Technology, Shenzhen, China 2 City University of Hong attempt to generalize it to larger-scale instances (e.g., with
Kong, Hong Kong SAR, China 3 Huawei Noah’s Ark Lab,
Shenzhen, China. Correspondence to: Zhenkun Wang more than 1, 000 nodes) (Kool et al., 2019; Kwon et al.,
<wangzhenkun90@gmail.com>. 2020). Although these models demonstrate good perfor-
mance on instances of similar scales to the ones they were

1
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

trained on, they struggle to generate reasonable good solu- perform well on the instances with the scale they are trained
tions for instances with much larger scales. Recently, two on, but their performance could drop dramatically on in-
different types of attempts have been explored to address the stances with different scales (Kwon et al., 2020; Xin et al.,
crucial limitation of RL-based NCO on large-scale general- 2020; 2021). To mitigate the poor generalization perfor-
ization. The first one is to perform an extra search procedure mance, an extra search procedure is usually required to find
on model inference to improve the quality of solution over a better solution. Some widely used search methods in-
greedy generation (Hottung et al., 2022; Choo et al., 2022). clude beam search (Joshi et al., 2019; Choo et al., 2022),
However, this approach typically requires expert-designed Monte Carlo tree search (MCTS) (Xing & Tu, 2020; Fu
search strategies and can be time-consuming when dealing et al., 2021; Qiu et al., 2022; Sun & Yang, 2023), and active
with large-scale problems. The second approach is to train search (Bello et al., 2016; Hottung et al., 2022). However,
the model on instances of varying scales (Khalil et al., 2017; these procedures are very time-consuming, could still per-
Cao et al., 2021). The challenge of this approach lies in form poorly on instances with quite different scales, and
effectively learning cross-scale features from these varying- might require expert-designed strategies on a specific prob-
scale training data to enhance the model’s generalization lem (e.g., MCTS for TSP). Recently, some two-stage ap-
performance. proaches, such as divide-and-conquer (Kim et al., 2021; Hou
et al., 2022) and local reconstruction (Li et al., 2021; Pan
Some recent works reveal that incorporating auxiliary infor-
et al., 2023; Cheng et al., 2023; Ye et al., 2023), have been
mation (e.g., the scales of the training instances) in training
proposed. Although these methods have a better generaliza-
can improve the model’s convergence efficiency and gen-
tion ability, they usually require expert-designed solvers and
eralization performance. However, these methods incorpo-
ignore the dependency between the two stages, which makes
rate auxiliary information into the decoding phase without
model design difficult, especially for non-expert users.
including the encoding phase. Although these methods
can improve the inference efficiency, the model fails to be
deeply aware of the auxiliary information, resulting in unsat- Varying-scale Training in NCO Directly training the
isfactory generalization performance on large-scale problem NCO model on instances with different scales is another
instances. In this work, we propose a powerful Instance- popular way to improve its generalization performance. This
Conditioned Adaptation Model (ICAM) to improve the straightforward approach can be traced back to Khalil et al.
large-scale generalization performance for RL-based NCO. (2017), which tries to train the model on instances with
Our contributions can be summarized as follows: 50 − 100 nodes to improve its generalization performance
to instances with up to 1, 200 nodes. Furthermore, Joshi et al.
• We design a novel and powerful instance-conditioned (2020) systematically tests the generalization performance
adaptation module for RL-based NCO to efficiently of NCO models by training on different TSP instances
leverage the instance-conditioned information (e.g., in- with 20 − 50 nodes. Subsequently, a series of works have
stance scale and distance between each node pair) to been developed to utilize the varying-scale training scheme
generate better solutions across different scales. The to improve their own NCO models’ generalization perfor-
proposed module is lightweight with low computa- mance (Lisicki et al., 2020; Cao et al., 2021; Manchanda
tional complexity, which can further facilitate training et al., 2022; Gao et al., 2023; Zhou et al., 2023). Similar
on instances with larger scales. to the varying-size training scheme, a few SL-based NCO
methods learn to construct partial solutions with various
• We develop a three-stage RL-based training scheme sizes during training and achieve a robust generalization per-
across instances with different scales, which enables formance on instances with different scales (Drakulic et al.,
our model to learn cross-scale features without any 2023; Luo et al., 2023). Nevertheless, in real-world applica-
labeled optimal solution. tions, it could be very difficult to obtain high-quality labeled
• We conduct various experiments on different routing solutions for SL-based model training. RL-based models
problems to demonstrate that our proposed ICAM can also face the challenge of efficiently capturing cross-scale
generate promising solutions for cross-scale instances features from varying-scale training data, which severely
with a very fast inference time. To the best of our hinders their generalization ability on large-scale problems.
knowledge, it achieves state-of-the-art performance
among all RL-based constructive methods for CVRP Information-aware NCO Recently, several works have
and TSP instances with up to 1, 000 nodes. indicated that incorporating auxiliary information (e.g., the
distance between each pair of nodes) can facilitate model
2. Related Works training and improve generalization performance. In Kim
et al. (2022b), the scale-related feature is added to the de-
Non-conditioned NCO Most NCO methods are trained coder’s context embedding to make the model scale-aware
on a fixed scale (e.g., 100 nodes). These models usually during the decoding phase. Jin et al. (2023), Son et al.

2
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

(2023) and Wang et al. (2024) use the distance to bias the
compatibility calculation, thereby guiding the model toward Instance-Conditioned Information
more efficient exploration. Gao et al. (2023) employs a local Scale
Distance
Matrix
policy network to catch distance and scale knowledge and
integrates it into the compatibility calculation. In Li et al. 𝑁=5
(2023), the distance-related feature is utilized to adaptively
refine the node embeddings so as to improve the model ex-
ploration. Overall, these methods all incorporate auxiliary
information into the decoding process to improve the infer-
ence efficiency. However, this additional plugin way cannot Encoder 𝐻 (𝐿) Decoder
efficiently integrate auxiliary information into the encoding
of nodes and fails to enable the model to be deeply aware of L×
the knowledge of distance and scale.
NCO Model
Current node Other nodes High-preference set
Figure 2. Intuitive Idea of Instance-Conditioned Adaptation.
By providing instance-conditioned information in both the en-
coding and decoding processes, the model is expected to better
comprehend and address problem instances.
is. In addition, the nodes of instances with different scales
exhibit significant variations in density. The scale is larger,
and the corresponding node distribution is denser. Therefore,
the model bias should vary according to the scale.
Small-scale instance Large-scale instance
(100 nodes) (1,000 nodes)
Instance-Conditioned Adaptation Function This work
proposes to integrate the scale and node-to-node distances
Figure 1. Node Selection Bias on Instances with Different Scales. via an instance-conditioned adaptation function. We de-
As the scale increases, the density of nodes increases. Therefore,
note the function as f (N, dij ), where N is the scale of the
the model tends to select the next node from a smaller sub-region of
the current node when dealing with large-scale instances compared
problem instance, and dij represents the distance between
to small-scale ones. each node i and each node j. As shown in Figure 2, the
f (N, dij ) aims to capture features related to instance scale
and node-to-node distances, and feed them into the model’s
3. Instance-Conditioned Adaptation encoding and decoding processes, respectively. Based on
the changing instances, the model could dynamically bias
3.1. Motivation and Key Idea the selection of nodes under the effect of f (N, dij ), thereby
Each instance has some specific information that benefits making better decisions in RL-based training. To enable
the adaptability and generalization of models. By providing f (N, dij ) to learn better features of large-scale general-
this instance-conditioned information, the model can bet- ization, we still need improvements in the following two
ter comprehend and address the problems, especially when aspects:
dealing with large-scale problems. The node-to-node dis-
tances and scale are two fundamental kinds of information • Lightweight and Fast Model: As the RL-based
in routing problems, and both types of information are vi- training on large-scale instances consumes enormous
tal. We need to make the model aware of scale changes to computational time and memory, we need a more
improve generalization. Meanwhile, we also need to allow lightweight yet quick model structure so that large-
the model to be aware of the node-to-node distances to en- scale instances can be included in the training data;
hance exploration and reduce the search space, which in • Efficient Training Scheme: We need more efficient
turn improves its training efficiency. training schemes to accelerate model convergence,
As the example of the two TSP instances shown in Figure 1, especially when training on large-scale problem in-
the two adjacent nodes in the optimal solution are normally stances.
within a specific sub-region, and the distance between them
should not be too far. Likewise, the model tends to select
3.2. Instance-Conditioned Adaptation Model
the next node from a sub-region of the current node. For
the node outside the sub-region, the farther it is from the As shown in Figure 3, the proposed ICAM also adopts the
current node, the lower the corresponding selection bias encoder-decoder structure, which is Transformer-like as

3
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Scale Distance Matrix

𝑑11 𝑑12 … 𝑑1𝑁

𝑑21 𝑑22 … 𝑑2𝑁 Adaptation Attention Compatibility with

𝑁=5 𝑓(𝑁, 𝑑𝑖𝑗 )
… … … … Free Module Adaptation Bias
𝑑𝑁1 𝑑𝑁2 … 𝑑𝑁𝑁

𝜽𝑒𝑛𝑐𝑜𝑑𝑒𝑟 𝜽𝑑𝑒𝑐𝑜𝑑𝑒𝑟

Figure 3. The Proposed ICAM. In our model, two essential types of instance-conditioned information are integrated into both the encoder
and the decoder. Specifically, we utilize AAFM to replace all MHA operations in both the encoder and decoder. Moreover, we combine
the adaptation function with the compatibility calculation in the decoder.

QK T
 
many existing NCO models (Kool et al., 2019; Kim et al., √
Attention(Q, K, V ) = softmax V, (2)
2022a; Luo et al., 2023). It is developed from a very well- dk
known NCO model POMO (Kwon et al., 2020), and the
details of POMO are provided in Appendix A. where X represents the input, W Q , W K , and W V are three
learning matrices, dk is the dimension for K. The MHA
Given an instance X = {xi }N i=1 , xi represents the features incurs the primary memory usage and computational cost,
of each node (e.g., the coordinates of each city in TSP). which poses challenges for training on large-scale instances.
These node features are transformed into the initial embed- Moreover, the specific design of the MHA makes it difficult
(0) (0)
dings H (0) = (h1 , . . . , hN ) via a linear projection. The to intuitively integrate the instance-conditioned information
initial embeddings pass through the L attention layers to (i.e., the adaptation function f (N, dij )).
(L) (L)
get the node embeddings H (L) = (h1 , . . . , hN ). The
We propose a novel module called Adaptation Attention
attention layer consists of a Multi-Head Attention (MHA)
Free Module (AAFM), as shown in Figure 4, to replace
sub-layer and a Feed-Forward (FF) sub-layer.
the MHA operation in both the encoder and the decoder.
During the decoding process, the model generates a solution AAFM is based on the AFT-full operation of the Attention
in an autoregressive manner. For the example of TSP, in the Free Transformer (AFT) (Zhai et al., 2021), which has lower
t-step construction, the context embedding is composed of computation and space complexity but can achieve similar
the first visited node embedding and the last visited node performance to the MHA. The details of the AFT are pro-
(L) (L) vided in Appendix B. We substitute the original bias w of
embedding, i.e., ht(C) = [hπ1 , hπt−1 ]. The new context
AFT-full with our adaptation function f (N, dij ), i.e.,
embedding ĥt(C) is then obtained via the MHA operation
on ht(C) and H (L) . Finally, the model yields the selection exp(A) · (exp(K) ⊙ V )
probability for each unvisited node pθ (πt = i | X, π1:t−1 ) AAFM(Q, K, V, A) = σ(Q) ⊙ ,
exp(A) · exp(K)
by calculating compatibility on ĥt(C) and H (L) . (3)
where σ is Sigmoid function, ⊙ represents the element-wise
product, and Aij = f (N, dij ) denotes the adaptation bias
Adaptation Attention Free Module The MHA operation between node i and node j. The detailed calculation of
is the core component of the Transformer-like NCO model. AAFM is shown in Figure 5.
In the mode of self-attention, MHA performs a scaled dot-
In AAFM, instance-conditioned information is integrated
product attention for each head, the self-attention calculation
in a more appropriate and ingenious manner, which enables
can be written as
our model to comprehend knowledge such as distance and
scale more efficiently. Furthermore, our AAFM exhibits
Q = XW Q , K = XW K , V = XW V , (1) lower computation and space complexity than MHA, which

4
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

𝐴 𝐾 𝑉 𝑄
Scale
Node-to-node Distances X ∈ ℝ𝑁×2
+ Mask (opt.)
(ℓ−1)
𝐡𝟏
(ℓ−1)
𝐡𝟐 … (ℓ−1)
𝐡𝑵
Exp Exp Sigmoid

𝐴11 𝐴12 … 𝐴1𝑁 Linear Linear Linear

𝐴21 𝐴22 … 𝐴2𝑁

MatMul Mul
… … … …
𝐴𝑁1 𝐴𝑁2 … 𝐴𝑁𝑁

𝐴 ∈ ℝ𝑁×𝑁 𝐾 ∈ ℝ𝑁×dh 𝑉 ∈ ℝ𝑁×dh 𝑄 ∈ ℝ𝑁×dh

MatMul

Adaptation Attention Free Module Div

𝑌 ∈ ℝ𝑁×dh Mul

Figure 4. The Structure of AAFM. Note that in decoder, dij in Figure 5. The Detailed Calculation Process of AAFM.
Aij is the distance between the current node and each node, and the
node masking state in current step t is added to Aij additionally.
model is trained for several epochs on small-scale instances.
could bring a more lightweight and faster model. For example, we use a total of 256, 000 randomly generated
TSP100 instances for each epoch in the first stage to train
Compatibility with Adaptation Bias We also integrate the model for 100 epochs. A warm-up training can make
the adaptation function f (N, dij ) into the compatibility cal- the model more stable in the subsequent training.
culation. The new compatibility, denoted as uti , can be
expressed as Stage 2: Learning on Varying-scale Instances In the sec-
 ond stage, we train the model on varying-scale instances for
(L) T
ĥt (h ) much longer epochs. We let the scale N be randomly sam-
ξ · tanh( (C)√di + At−1,i ) if i ̸∈ {π1:t−1 }

uti = k , pled from the discrete uniform distribution Unif([100,500])
−∞ otherwise for each batch. Considering the GPU memory constraints,
(4) we decrease the batch size with the scale increases. The loss
t
eui function (denoted as LPOMO ) used in the first and second
pθ (πt = i | X, π1:t−1 ) = PN ut , (5)
j=1 e
j stages is the same as in POMO. The gradient ascent with an
(L)
approximation of the loss function can be written as
where ξ is the clipping parameter, ĥt(C) and hi are cal-
culated via AAFM instead of MHA. At−1,i represents the 1 XX
B N
Gm,i ∇θ log pθ π i | Xm ,


adaptation bias between each remaining node and the cur- ∇θ LPOMO (θ) ≈
BN m=1 i=1
rent node. Finally, the probability of generating a complete
solution π for instance X is calculated as (7)
i i


N
Gm,i = R π | Xm − b (Xm ), (8)
Y
pθ (π | X) = pθ (πt | X, π1:t−1 ). (6) N
1 X
bi (Xm ) = R π j | Xm
t=2


for all i. (9)
N j=1
By integrating the adaptation bias in the compatibility cal-
culation, the model’s performance can be further enhanced.

where R π i | Xm represents the return (e.g., tour length)
of instance Xm given a specific solution π i . Equation (9) is
3.3. Varying-scale Training Scheme a shared baseline as introduced in Kwon et al. (2020).
We develop a three-stage training scheme to enable the
model to be aware of instance-conditioned information more Stage 3: Top-k Elite Training Under the POMO struc-
effectively. We describe the three training stages as follows: ture, N trajectories are constructed in parallel for each in-
stance during training. In the third stage, we want the model
Stage 1: Warming-up on Small-scale Instances We em- to focus more on the best k trajectories among all N trajec-
ploy a warm-up procedure in the first stage. Initially, the tories. To achieve this, we design a new loss LTop , and its

5
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

gradient ascent can be expressed as 2. In the second stage, the scale N is randomly sampled
B k from the discrete uniform distribution Unif([100,500])
1 X X
and optimize memory usage by adjusting batch sizes
Gm,i ∇θ log pθ π i | Xm .


∇θ LTop (θ) ≈
Bk m=1 i=1 according to the changed scales. For TSP, the batch
(10) size bs = 160 × ( 100 2
N ) , with a training duration of
We combine LTop with LPOMO as the joint loss in the 2,200epochs. In the  case of CVRP, the batch size
training of the third stage, i.e., bs = 128 × ( 100N )2
, with a training duration of 700
epochs. Furthermore, the capacity of each batch is
LJoint = LPOMO + β · LTop . (11)
consistently set by random sampling from the discrete
where β ∈ [0, 1] is a coefficient balancing the original loss uniform distribution Unif([50,100]).
and the new loss.
3. In the last stage, we adjust the learning rate η to 10−5
across all models to enhance model convergence and
4. Experiments training stability. The parameter β and λ are set to
In this section, we conduct a comprehensive comparison be- 0.1 and 20, respectively, as specified in Equation (10)
tween our proposed model and other classical and learning- and Equation (11). The training period is standardized
based solvers using Traveling Salesman Problem (TSP) and to 200 epochs for all models, and other settings are
Capacitated Vehicle Routing Problem (CVRP) instances of consistent with the second stage.
different scales.
Overall, we train the TSP model for 2, 500 epochs and the
Problem Setting For TSP and CVRP, the instances of CVRP model for 1, 000 epochs. For more details on model
training and testing are generated randomly, following Kool hyperparameter settings, please refer to Appendix C.
et al. (2019). For the test set, we generate 10, 000 instances
for 100-node, and 128 instances for each of 200-, 500-, and Baseline We compare ICAM with the following methods:
1, 000-node, respectively. Specifically, for CVRP instances
of different scales, we use capacities of 50, 80, 100, and 1. Classical solver: Concorde (Applegate et al., 2006),
250, respectively (Drakulic et al., 2023; Luo et al., 2023). LKH3 (Helsgaun, 2017) and HGS (Vidal, 2022);

Model Setting The adaptation function f (N, dij ) should 2. Constructive NCO: POMO (Kwon et al., 2020),
be problem-depended. For TSP and CVRP, we define it as MDAM (Xin et al., 2021), ELG (Gao et al., 2023),
Pointerformer (Jin et al., 2023), BQ (Drakulic et al.,
f (N, dij ) = −α · log2 N · dij ∀i, j ∈ 1, . . . , N, (12)
2023) and LEHD (Luo et al., 2023);
where α is a learnable parameter, and it is initially set to 1.
3. Two-stage NCO: Att-GCN+MCTS (Fu et al., 2021),
The embedding dimension of our model is set to 128, and DIMES (Qiu et al., 2022), TAM (Hou et al., 2022),
the dimension of the feed-forward layer is set to 512. We SO (Cheng et al., 2023), DIFUSCO (Sun & Yang,
set the number of attention layers in the encoder to 12. The 2023), H-TSP (Pan et al., 2023) and GLOP (Ye et al.,
clipping parameter ξ = 50 in Equation (4) to obtain the 2023).
better training convergence (Jin et al., 2023). We train and
test all experiments using a single NVIDIA GeForce RTX Metrics and Inference We use the solution lengths, op-
3090 GPU with 24GB memory. timality gaps, and total inference times to evaluate the per-
formance of each method. Specifically, the optimality gap
Training For all models, we use Adam (Kingma & Ba, measures the discrepancy between the solutions generated
2014) as the optimizer, initial learning rate η is 10−4 . Every by learning and non-learning methods and the optimal so-
epoch, we process 1, 000 batches for all problems. For each lutions, which are obtained using Concorde for TSP and
instance, N different tours are always generated in parallel, LKH3 for CVRP. Note that the inference times for classi-
each of them starting from a different city (Kwon et al., cal solvers, which run on a single CPU, and for learning-
2020). The rest of the training settings are as follows: based methods, which utilize GPUs, are inherently different.
Therefore, these times should not be directly compared.
1. In the first stage of the process, we set different batch
sizes for different problems due to memory constraints: For most NCO baseline methods, we directly execute the
256 for TSP and 128 for CVRP. We use problem in- source code provided by the authors with default settings.
stances for TSP100 and CVRP100 to train the cor- We report the original results as published in correspond-
responding model for 100 epochs. Additionally, the ing papers for methods like Att-GCN+MCTS, DIMES, SO.
capacity for each CVRP instance is fixed at 50. Following the approach in Kwon et al. (2020), we report

6
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Table 1. Experimental results on TSP and CVRP with uniformly distributed instances. The results marked with an asterisk (*) are directly
obtained from the original papers.
TSP100 TSP200 TSP500 TSP1000
Method Obj. Gap Time Obj. Gap Time Obj. Gap Time Obj. Gap Time
Concorde 7.7632 0.000% 34m 10.7036 0.000% 3m 16.5215 0.000% 32m 23.1199 0.000% 7.8h
LKH3 7.7632 0.000% 56m 10.7036 0.000% 4m 16.5215 0.000% 32m 23.1199 0.000% 8.2h
Att-GCN+MCTS* 7.7638 0.037% 15m 10.8139 0.884% 2m 16.9655 2.537% 6m 23.8634 3.224% 13m
DIMES AS+MCTS* − − − − − − 16.84 1.76% 2.15h 23.69 2.46% 4.62h
SO-mixed* − − − 10.7873 0.636% 21.3m 16.9431 2.401% 32m 23.7656 2.800% 55.5m
DIFUSCO greedy+2-opt* 7.78 0.24% − − − − 16.80 1.49% 3.65m 23.56 1.90% 12.06m
H-TSP − − − − − − 17.549 6.220% 23s 24.7180 6.912% 47s
GLOP (more revisions) 7.7668 0.046% 1.9h 10.7735 0.653% 42s 16.8826 2.186% 1.6m 23.8403 3.116% 3.3m
BQ greedy 7.7903 0.349% 1.8m 10.7644 0.568% 9s 16.7165 1.180% 46s 23.6452 2.272% 1.9m
LEHD greedy 7.8080 0.577% 27s 10.7956 0.859% 2s 16.7792 1.560% 16s 23.8523 3.168% 1.6m
MDAM bs50 7.7933 0.388% 21m 10.9173 1.996% 3m 18.1843 10.065% 11m 27.8306 20.375% 44m
POMO aug×8 7.7736 0.134% 1m 10.8677 1.534% 5s 20.1871 22.187% 1.1m 32.4997 40.570% 8.5m
ELG aug×8 7.7988 0.458% 5.1m 10.8400 1.274% 17s 17.1821 3.998% 2.2m 24.7797 7.179% 13.7m
Pointerformer aug×8 7.7759 0.163% 49s 10.7796 0.710% 11s 17.0854 3.413% 53s 24.7990 7.263% 6.4m
ICAM single trajec. 7.8328 0.897% 2s 10.8255 1.139% <1s 16.7777 1.551% 1s 23.7976 2.931% 2s
ICAM 7.7991 0.462% 5s 10.7753 0.669% <1s 16.6978 1.067% 4s 23.5608 1.907% 28s
ICAM aug×8 7.7747 0.148% 37s 10.7385 0.326% 3s 16.6488 0.771% 38s 23.4854 1.581% 3.8m
CVRP100 CVRP200 CVRP500 CVRP1000
Method Obj. Gap Time Obj. Gap Time Obj. Gap Time Obj. Gap Time
LKH3 15.6465 0.000% 12h 20.1726 0.000% 2.1h 37.2291 0.000% 5.5h 37.0904 0.000% 7.1h
HGS 15.5632 -0.533% 4.5h 19.9455 -1.126% 1.4h 36.5611 -1.794% 4h 36.2884 -2.162% 5.3h
GLOP-G (LKH3) − − − − − − − − − 39.6507 6.903% 1.7m
BQ greedy 16.0730 2.726% 1.8m 20.7722 2.972% 10s 38.4383 3.248% 47s 39.2757 5.892% 1.9m
LEHD greedy 16.2173 3.648% 30s 20.8407 3.312% 2s 38.4125 3.178% 17s 38.9122 4.912% 1.6m
MDAM bs50 15.9924 2.211% 25m 21.0409 4.304% 3m 41.1376 10.498% 12m 47.4068 27.814% 47m
POMO aug×8 15.7544 0.689% 1.2m 21.1542 4.866% 6s 44.6379 19.901% 1.2m 84.8978 128.894% 9.8m
ELG aug×8 15.9973 2.242% 6.3m 20.7361 2.793% 19s 38.3413 2.987% 2.6m 39.5728 6.693% 15.6m
ICAM single trajec. 16.1868 3.453% 2s 20.7509 2.867% <1s 37.9594 1.962% 1s 38.9709 5.070% 2s
ICAM 15.9386 1.867% 7s 20.5185 1.715% 1s 37.6040 1.007% 5s 38.4170 3.577% 35s
ICAM aug×8 15.8720 1.442% 47s 20.4334 1.293% 4s 37.4858 0.689% 42s 38.2370 3.091% 4.5m

GCN+MCTS takes 15 minutes compared to our 37 seconds.

Table 2. Empirical results on CVRPLib Set-X(Uchoa et al., 2017).
On TSP1000, our model impressively reduces the optimality
BKS refers to the “Best Known Solution”. The results marked gap to less than 3% in just 2 seconds. When switching to a
with an asterisk (*) are directly obtained from the original papers. multi-greedy strategy, the optimality gap further narrows to
N ≤ 200 200<N ≤ 500 500<N ≤ 1000 Total Avg.time 1.9% in 30 seconds. With the instance augmentation, ICAM
(22 instances) (46 instances) (32 instances) (100 instances)
can achieve the optimality gap of 1.58% in less than 4 min-
BKS 0.00% 0.00% 0.00% 0.00% −
utes. To the best of our knowledge, for TSP and CVRP up to
LEHD 11.35% 9.45% 17.74% 12.52% 1.58s
BQ* − − − 9.94% − 1, 000 nodes, our model shows state-of-the-art performance
POMO 9.76% 19.12% 57.03% 29.19% 0.35s among all RL-based constructive NCO methods.
ELG 5.50% 5.67% 5.74% 5.66% 0.61s
ICAM 5.14% 4.44% 5.17% 4.83% 0.34s

Results on Benchmark Dataset We further evaluate the

three types of results: those using a single trajectory, the performance of each method using the well-known bench-
best result from multiple trajectories, and results derived mark datasets from CVRPLib Set-X (Uchoa et al., 2017).
from instance augmentation. In these evaluations, instance augmentation is not employed
for any of the methods. The detailed results are presented
Experimental Results The experimental results on uni- in Table 2, showing that our method consistently maintains
formly distributed instances are reported in Table 1. Our the best performance. ICAM achieves the best performance
method stands out for consistently delivering superior infer- across instances of all scale ranges. Among all learning-
ence performance, complemented by remarkably fast infer- based models, ICAM has the fastest inference time. This
ence times, across various problem instances. Although it also shows the outstanding generalization performance of
cannot surpass Att-GCN+MCTS on TSP100 and POMO on ICAM. According to our knowledge, in the Set-X tests, our
CVRP100, the time it consumes is significantly less. Att- method has achieved the best performance to date.

7
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Table 3. Comparative results in the capacity setting of TAM (Hou et al., 2022) with scale ≥ 1000. “Time” represents the per-instance
runtime. The results marked with an asterisk (*) are directly obtained from the original papers. Note that except for CVRP3000 and
CVRP4000, the optimal solutions obtained using LKH3 are from the original TAM paper.
CVRP1K CVRP2K CVRP3K CVRP4K CVRP5K
Method Obj. Gap Time(s) Obj. Gap Time(s) Obj. Gap Time(s) Obj. Gap Time(s) Obj. Gap Time(s)
LKH3 46.44 0.000% 6.15 64.93 0.000% 20.29 89.90 0.000% 41.10 118.03 0.000% 80.24 175.66 0.000% 151.64
TAM-AM* 50.06 7.795% 0.76 74.31 14.446% 2.2 − − − − − − 172.22 -1.958% 11.78
TAM-LKH3* 46.34 -0.215% 1.82 64.78 -0.231% 5.63 − − − − − − 144.64 -17.659% 17.19
TAM-HGS* − − − − − − − − − − − − 142.83 -18.690% 30.23
GLOP-G (LKH3) 45.90 -1.163% 0.92 63.02 -2.942% 1.34 88.32 -1.758% 2.12 114.20 -3.245% 3.25 140.35 -20.101% 4.45
LEHD greedy 43.96 -5.340% 0.79 61.58 -5.159% 5.69 86.96 -3.270% 18.39 112.64 -4.567% 44.28 138.17 -21.342% 87.12
BQ greedy 44.17 -4.886% 0.55 62.59 -3.610% 1.83 88.40 -1.669% 4.65 114.15 -3.287% 11.50 139.84 -20.389% 27.63
ICAM single trajec. 43.58 -6.158% 0.02 62.38 -3.927% 0.04 89.06 -0.934% 0.10 115.09 -2.491% 0.19 140.25 -20.158% 0.28
ICAM 43.07 -7.257% 0.26 61.34 -5.529% 2.20 87.20 -3.003% 6.42 112.20 -4.939% 15.50 136.93 -22.048% 29.16

Comparison on Larger-scale Instances We also conduct Effects of Adaptation Function Given that we apply the
experiments on instances for TSP and CVRP with larger adaptation function outlined in Equation (12) to both the
scales, the instance augmentation is not employed for all AAFM and the subsequent compatibility calculation, we
methods due to computational efficiency. For CVRP, fol- conducted three different experiments to validate the effi-
lowing Hou et al. (2022), the capacities for instances with cacy of this function. Detailed results of these experiments
1,000, 2,000, and larger scales are set at 200, 300, and 300, are available in Appendix E.3.
respectively. We perform our model on the dataset gen-
erated by the same settings. Except for CVRP3000 and Parameter Settings in Stage 3 In the final stage, we man-
CVRP4000 instances where LKH3 is used to obtain their ually adjust the β and k values as specified in Equation (10).
optimal solution, the optimal solutions of other instances The experimental results for two settings, involving different
are from the original paper. values, are presented in Appendix E.4.
As shown in Table 3, on CVRP instances with scale ≥ 1000,
our method outperforms the other methods, including GLOP Efficient Inference Strategies for Different Models To
with LKH3 solver and all TAM variants, on all problem in- further improve model performance, many inference strate-
stances except for CVRP3000. On CVRP3000, ICAM is gies are developed for different NCO models. For example,
slightly worse than LEHD. LEHD is an SL-based model and BQ employs beam search, while LEHD uses the Random
consumes much more solving time than ICAM. As shown Re-Construct (RRC) in inference. We investigate the effects
in Appendix D, the superiority of ICAM is not so obvious of the inference strategy on different models. The analysis
on TSP instances with scale >1, 000. Its performance is is provided in Appendix E.5.
slightly worse than the two SL-based NCO models, BQ and
LEHD. Nevertheless, on TSP2000 and TSP3000 instances, 6. Conclusion, Limitation, and Future Work
we achieve the best results in RL-based constructive meth-
ods, and on TSP4000 and TSP5000 instances, we are only In this work, we have proposed a novel ICAM to improve
slightly worse than ELG. Overall, our method still has a large-scale generalization for RL-based NCO. The instance-
good large-scale generalization ability. conditioned information is more effectively integrated into
the model’s encoding and decoding via a powerful yet
lightweight AAFM and the new compatibility calculation.
5. Ablation Study In addition, we have developed a three-stage training scheme
ICAM vs. POMO To improve the model’s ability to that enables the model to learn cross-scale features more effi-
be aware of scale, we implement a varying-scale training ciently. The experimental results on various TSP and CVRP
scheme. Given that our model is an advancement over the instances show that ICAM achieves promising generaliza-
POMO framework, we ensure a fair comparison by training tion abilities compared with other representative methods.
a new POMO model using our training settings. The results ICAM demonstrates superior performance with greedy de-
of the two models are reported in Appendix E.1. coding. However, we have observed its poor applicability
to other complex inference strategies (e.g., RRC and beam
search). In the future, we may develop a suitable inference
Effects of Different Stages Our training is divided into strategy for ICAM. Moreover, the generalization perfor-
three different stages, each contributing significantly to mance of ICAM over differently distributed datasets should
the overall effectiveness. The performance improvements be investigated in the future.
achieved at each stage are detailed in Appendix E.2.

8
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Impact Statement Helsgaun, K. An extension of the lin-kernighan-helsgaun

tsp solver for constrained traveling salesman and vehicle
This paper presents work whose goal is to advance the field routing problems. Roskilde: Roskilde University, 12,
of Machine Learning. There are many potential societal 2017.
consequences of our work, none which we feel must be
specifically highlighted here. Hottung, A., Kwon, Y.-D., and Tierney, K. Efficient ac-
tive search for combinatorial optimization problems. In
References International Conference on Learning Representations,
2022.
Applegate, D., Bixby, R., Chvatal, V., and Cook, W. Con-
corde tsp solver, 2006. Hou, Q., Yang, J., Su, Y., Wang, X., and Deng, Y. Generalize
learned heuristics to solve large-scale vehicle routing
Ba, J. L., Kiros, J. R., and Hinton, G. E. Layer normalization. problems in real-time. In The Eleventh International
arXiv preprint arXiv:1607.06450, 2016. Conference on Learning Representations, 2022.
Bello, I., Pham, H., Le, Q. V., Norouzi, M., and Bengio, Ioffe, S. and Szegedy, C. Batch normalization: Accelerating
S. Neural combinatorial optimization with reinforcement deep network training by reducing internal covariate shift.
learning. arXiv preprint arXiv:1611.09940, 2016. In International Conference on Machine Learning, pp.
Bengio, Y., Lodi, A., and Prouvost, A. Machine learning 448–456. PMLR, 2015.
for combinatorial optimization: a methodological tour
Jin, Y., Ding, Y., Pan, X., He, K., Zhao, L., Qin, T., Song, L.,
d’horizon. European Journal of Operational Research,
and Bian, J. Pointerformer: Deep reinforced multi-pointer
290(2):405–421, 2021.
transformer for the traveling salesman problem. In The
Cao, Y., Sun, Z., and Sartoretti, G. Dan: Decentral- Thirty-Seventh AAAI Conference on Artificial Intelligence,
ized attention-based neural network for the minmax 2023.
multiple traveling salesman problem. arXiv preprint
Joshi, C. K., Laurent, T., and Bresson, X. An efficient
arXiv:2109.04205, 2021.
graph convolutional network technique for the travelling
Cheng, H., Zheng, H., Cong, Y., Jiang, W., and Pu, S. Select salesman problem. arXiv preprint arXiv:1906.01227,
and optimize: Learning to aolve large-scale tsp instances. 2019.
In International Conference on Artificial Intelligence and
Statistics, pp. 1219–1231. PMLR, 2023. Joshi, C. K., Cappart, Q., Rousseau, L.-M., and Lau-
rent, T. Learning the travelling salesperson prob-
Choo, J., Kwon, Y.-D., Kim, J., Jae, J., Hottung, A., Tierney, lem requires rethinking generalization. arXiv preprint
K., and Gwon, Y. Simulation-guided beam search for arXiv:2006.07054, 2020.
neural combinatorial optimization. Advances in Neural
Information Processing Systems, 35:8760–8772, 2022. Khalil, E., Dai, H., Zhang, Y., Dilkina, B., and Song,
L. Learning combinatorial optimization algorithms over
Drakulic, D., Michel, S., Mai, F., Sors, A., and Andreoli, J.- graphs. Advances in Neural Information Processing Sys-
M. Bq-nco: Bisimulation quotienting for efficient neural tems, 30, 2017.
combinatorial optimization. In Thirty-seventh Conference
on Neural Information Processing Systems, 2023. Kim, M., Park, J., et al. Learning collaborative policies
to solve np-hard routing problems. Advances in Neural
Fu, Z.-H., Qiu, K.-B., and Zha, H. Generalize a small Information Processing Systems, 34:10418–10430, 2021.
pre-trained model to arbitrarily large tsp instances. In
Proceedings of the AAAI Conference on Artificial Intelli- Kim, M., Park, J., and Park, J. Sym-nco: Leveraging
gence, volume 35, pp. 7474–7482, 2021. symmetricity for neural combinatorial optimization. Ad-
vances in Neural Information Processing Systems, 35:
Gao, C., Shang, H., Xue, K., Li, D., and Qian, C. To- 1936–1949, 2022a.
wards generalizable neural solvers for vehicle routing
problems via ensemble with transferrable local policy. Kim, M., Son, J., Kim, H., and Park, J. Scale-conditioned
arXiv preprint arXiv:2308.14104, 2023. adaptation for large scale combinatorial optimization. In
NeurIPS 2022 Workshop on Distribution Shifts: Connect-
He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learn- ing Methods and Applications, 2022b.
ing for image recognition. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, Kingma, D. P. and Ba, J. Adam: A method for stochastic
pp. 770–778, 2016. optimization. arXiv preprint arXiv:1412.6980, 2014.

9
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Kool, W., van Hoof, H., and Welling, M. Attention, learn to Sun, Z. and Yang, Y. Difusco: Graph-based diffusion
solve routing problems! In International Conference on solvers for combinatorial optimization. arXiv preprint
Learning Representations, 2019. arXiv:2302.08224, 2023.
Kwon, Y.-D., Choo, J., Kim, B., Yoon, I., Gwon, Y., and Uchoa, E., Pecin, D., Pessoa, A., Poggi, M., Vidal, T., and
Min, S. Pomo: Policy optimization with multiple optima Subramanian, A. New benchmark instances for the ca-
for reinforcement learning. Advances in Neural Informa- pacitated vehicle routing problem. European Journal of
tion Processing Systems, 33:21188–21198, 2020. Operational Research, 257(3):845–858, 2017.
Li, B., Wu, G., He, Y., Fan, M., and Pedrycz, W. An Ulyanov, D., Vedaldi, A., and Lempitsky, V. Instance nor-
overview and experimental study of learning-based op- malization: The missing ingredient for fast stylization.
timization algorithms for the vehicle routing problem. arXiv preprint arXiv:1607.08022, 2016.
IEEE/CAA Journal of Automatica Sinica, 9(7):1115–
1138, 2022. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,
L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. Atten-
Li, J., Ma, Y., Cao, Z., Wu, Y., Song, W., Zhang, J., and tion is all you need. Advances in Neural Information
Chee, Y. M. Learning feature embedding refiner for Processing Systems, 30, 2017.
solving vehicle routing problems. IEEE Transactions on
Neural Networks and Learning Systems, 2023. Vidal, T. Hybrid genetic search for the cvrp: Open-source
implementation and swap* neighborhood. Computers &
Li, S., Yan, Z., and Wu, C. Learning to delegate for large-
Operations Research, 140:105643, 2022.
scale vehicle routing. Advances in Neural Information
Processing Systems, 34:26198–26211, 2021. Vinyals, O., Fortunato, M., and Jaitly, N. Pointer networks.
Lisicki, M., Afkanpour, A., and Taylor, G. W. Evaluating Advances in neural information processing systems, 28,
curriculum learning strategies in neural combinatorial 2015.
optimization. arXiv preprint arXiv:2011.06188, 2020. Wang, Y., Jia, Y.-H., Chen, W.-N., and Mei, Y. Distance-
Luo, F., Lin, X., Liu, F., Zhang, Q., and Wang, Z. Neural aware attention reshaping: Enhance generalization of
combinatorial optimization with heavy decoder: Toward neural solver for large-scale vehicle routing problems.
large scale generalization. In Thirty-seventh Conference arXiv preprint arXiv:2401.06979, 2024.
on Neural Information Processing Systems, 2023. Williams, R. J. Simple statistical gradient-following algo-
Manchanda, S., Michel, S., Drakulic, D., and Andreoli, J.-M. rithms for connectionist reinforcement learning. Machine
On the generalization of neural combinatorial optimiza- learning, 8:229–256, 1992.
tion heuristics. In Joint European Conference on Machine
Xiao, Y., Wang, D., Li, B., Wang, M., Wu, X., Zhou, C.,
Learning and Knowledge Discovery in Databases, pp.
and Zhou, Y. Distilling autoregressive models to obtain
426–442. Springer, 2022.
high-performance non-autoregressive solvers for vehi-
Nazari, M., Oroojlooy, A., Snyder, L., and Takác, M. Rein- cle routing problems with faster inference speed. arXiv
forcement learning for solving the vehicle routing prob- preprint arXiv:2312.12469, 2023.
lem. Advances in neural information processing systems,
31, 2018. Xin, L., Song, W., Cao, Z., and Zhang, J. Step-wise deep
learning models for solving routing problems. IEEE
Pan, X., Jin, Y., Ding, Y., Feng, M., Zhao, L., Song, L., Transactions on Industrial Informatics, 17(7):4861–4871,
and Bian, J. H-tsp: Hierarchically solving the large-scale 2020.
travelling salesman problem. In Proceedings of the AAAI
Conference on Artificial Intelligence, 2023. Xin, L., Song, W., Cao, Z., and Zhang, J. Multi-decoder
attention model with embedding glimpse for solving ve-
Qiu, R., Sun, Z., and Yang, Y. Dimes: A differentiable hicle routing problems. In Proceedings of the AAAI Con-
meta solver for combinatorial optimization problems. Ad- ference on Artificial Intelligence, volume 35, pp. 12042–
vances in Neural Information Processing Systems, 35: 12049, 2021.
25531–25546, 2022.
Xing, Z. and Tu, S. A graph neural network assisted monte
Son, J., Kim, M., Kim, H., and Park, J. Meta-sage: Scale carlo tree search approach to traveling salesman problem.
meta-learning scheduled adaptation with guided explo- IEEE Access, 8:108418–108428, 2020.
ration for mitigating scale shift on combinatorial opti-
mization. In International Conference on Machine Learn- Ye, H., Wang, J., Liang, H., Cao, Z., Li, Y., and Li, F.
ing. PMLR, 2023. Glop: Learning global partition and local construction for

10
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

solving large-scale routing problems in real-time. arXiv

preprint arXiv:2312.08224, 2023.
Zhai, S., Talbott, W., Srivastava, N., Huang, C., Goh, H.,
Zhang, R., and Susskind, J. An attention free transformer.
arXiv preprint arXiv:2105.14103, 2021.
Zhou, J., Wu, Y., Song, W., Cao, Z., and Zhang, J. Towards
omni-generalizable neural methods for vehicle routing
problems. In International Conference on Machine Learn-
ing, 2023.

11
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

A. POMO Structure

𝜋11 𝜋21 … 1
𝜋𝑁

𝜋12 𝜋22 … 𝜋𝑁2

x1 x2 … xN …
𝜋1𝑁 𝜋2𝑁 … 𝜋𝑁𝑁
Linear Projection

Multi-Head Attention
𝜋𝑡1 𝜋𝑡2 … 𝜋𝑡𝑁

Autoregressive
Add & Instance Norm Softmax
L×
Feed Forward Compatibility

Add & Instance Norm Multi-Head Attention

𝐡𝟏
(𝑳)
𝐡𝟐
(𝑳) … (𝑳)
𝐡𝑵 Context Embedding

Figure 6. The Illustration of POMO Model. Note that POMO constructs N trajectories for a single instance, so the decoder outputs the
corresponding node selection for each trajectory at the current time step t.

As shown in Figure 6, POMO can be parameterized by θ, although POMO is derived from AM, it still has several differences
compared to AM. The implementation of POMO is shown as follows:

Encoder The encoder generates the embedding of each node based on the node coordinates as well as problem-specific
information (e.g., user demand for CVRP), aiming at embedding the graph information of the problem into high-dimension
vectors.
In the AM structure, the positional encoding is removed. And POMO disuse Layer Normalization (LN) (Ba et al., 2016),
which is used by Transformer, and Batch Normalization (BN)(Ioffe & Szegedy, 2015), which is used by AM, POMO uses
Instance Normalization (IN) (Ulyanov et al., 2016)
Given an instance X = {xi }N i=1 , first of all, the encoder takes node features xi ∈ R
dx
as model input, and transforms
(0) (0)
them to initial embeddings hi ∈ R through a linear projection, i.e., hi = Wi xi + bxi for i = 1, . . . , N . The initial
dh x
(0) (0)
embeddings {h1 , . . . , hN } pass through the L attention layers in turn, and are finally transformed to the final node
(L) (L)
embeddings H (L) = (h1 , . . . , hN ).
Similar to the traditional Transformer, the attention layer of POMO consists of two sub-layers: a Multi-Head Attention
(MHA) sub-layer and a Feed-Forward (FF) sub-layer. Both of them use Instance Normalization and skip-connection (He
(ℓ−1) (ℓ−1)
et al., 2016). Let H (ℓ−1) = (h1 , . . . , hN ) be the input of the ℓ-th attention layer for ℓ = 1, . . . , L. The outputs of its
MHA and FF sub-layer in terms of the i-th node are calculated as:

(ℓ) (ℓ−1) (ℓ−1)

ĥi = IN(ℓ) (hi + MHA(ℓ) (hi , H (ℓ−1) )), (13)

(ℓ) (ℓ) (ℓ)

hi = IN(ℓ) (ĥi + FF(ℓ) (ĥi )), (14)

where IN(·) denotes Instance Normalization, MHA(·) is Multi-Head Attention operation in Equation (13) (Vaswani et al.,
2017), and FF(·) in Equation (14) represents a fully connected neural network with the ReLU activation.

Decoder Based on the node embeddings generated by the encoder, the decoder adopts autoregressive mode to step-by-step
extend the existing partial solution and generate a feasible problem solution based on mask operation.
The decoder constructs the solution based on context embedding and node embeddings from the encoder step-by-step. In
the latest source code provided by the authors of POMO, for each time step, POMO adopts a simpler context embedding
PN (L)
rather than AM, and it has no graph embedding, i.e., h̄(L) = N1 i=1 hi ∈ Rdh .

12
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

For the TSP, t ∈ {1, . . . , N }, the context embedding ht(C) ∈ R2×dh is expressed as:
(
(L) (L)
t [hπ1 , hπt−1 ] t>1
h(C) = (15)
None t = 1,
(L) (L)
where hπ1 is the embedding of the first visited node, in POMO, each node is considered as the first visited node, hπt−1 is
the embedding of the last visited node. And [·, ·] denotes the concatenation operator.
Let ĥt(C) ∈ Rdh represents the new context embedding. ĥt(C) is calculated via a MHA operation:

ĥt(C) = MHA(C) (ht(C) , H (L) ). (16)

Finally, POMO computes the compatibility uti using Equation (17), the output probabilities pθ (πt |s, π1:t−1 ) is defined as1 :
 (L) T
ĥt (h )
ξ · tanh( (C)√di ) if i ̸∈ {π1:t−1 }

t
ui = k , (17)
−∞ otherwise

where ξ = 10 is a given clipping parameter in POMO, and then the probability pθ (πt = i | X, π1:t−1 ) of the next visiting
node is obtained through Equation (5). Thereby, the probability of generating a complete solution π for instance X can be
calculated in Equation (6).

Training According to Kwon et al. (2020), POMO is trained by the REINFORCE (Williams, 1992), and it uses gradient
ascent with an approximation in Equation (7).

B. Attention Free Transformer

According to Zhai et al. (2021), given the input X, AFT first transforms it to obtain Q, K, V by the corresponding linear
projection operation, respectively. Then, the calculation of AFT is expressed as:
Q = XW Q , K = XW K , V = XW V , (18)
PN
j=1 exp (Kj + wi,j ) ⊙ Vj
Yi = σ (Qi ) ⊙ PN (19)
j=1 exp (Kj + wi,j )

where W Q , W K , W V are three learnable matrices, ⊙ is the element-wise product, σq denotes the nonlinear function applied
to the query Q, the default function is Sigmoid, w ∈ RN ×N is the pair-wise position biases, and each wi,j is a scalar. In
AFT, for every specified target position i, it executes a weighted average of values, and this averaged result is then integrated
with the query by performing an element-wise multiplication.
The basic version of AFT outlined in Equation (19) is called AFT-full, and it is the version that we have adopted. AFT
includes three additional variants: AFT-local, AFT-simple and AFT-conv. Owing to the removal of the multi-head mechanism,
AFT exhibits reduced memory usage and increased speed during both the training and testing, compared to the traditional
Transformer. In fact, AFT can be viewed as a specialized form of MHA, where each feature dimension is treated as an
individual head. A complexity analysis comparing ”AFT-full” with these other variants is provided in Table 4. Further
details are available in the related work section mentioned above.

C. Implementation Details
For the TSP and CVRP, we are adapted from the POMO model (Kwon et al., 2020), and we modify certain settings to suit
our specific requirements. We remove the weight decay method because we observe that the addition of weight decay is
useless for improving the model generalization performance. Note that for the CVRP model, we have implemented the
gradient clipping technique, setting the max norm parameter to 5, to prevent the risk of exploding gradients. The rest of the
settings are consistent with the original POMO model, and detailed information about the hyperparameter settings of our
models can be found in Table 5.
(L)
1
In the latest source code provided by the authors of POMO, ĥt(C) and (hi )T have no imposed learnable matrices.

13
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Table 4. Complexity comparison of AFT-Full and other AFT variants. Here N, d, s denote the sequence length, feature dimension, and
local window size.
Model Time Space
2
Transformer O(N d) O(N 2 + N d)
AFT-full O(N 2 d) O(N d)
AFT-simple O(N d) O(N d)
AFT-local O(N sd), s < N O(N d)
AFT-conv O(N sd), s < N O(N d)

Table 5. Model hyperparameter settings in TSP and CVRP.

TSP CVRP
Optimizer Adam
Clipping parameter 50
Initial learning rate 10−4
Learning rate of stage 3 10−5
Weight decay −
Initial α value 1
Loss function of stage 1 & 2 LPOMO
Loss function of stage 3 LJoint
Parameter β of stage 3 0.1
Parameter k of stage 3 20
The number of encoder layer 12
Embedding dimension 128
Feed forward dimension 512
Scale of stage 1 100
Scale of stage 2 & 3 [100, 500]
Batches of each epoch 1, 000
Epochs of stage 1 100
Epochs of stage 3 200
Epochs of stage 2 2, 200 700
Capacity of stage 1 − 50
Capacity of stage 2 & 3 − [50, 100]
Gradient clipping − max norm=5
Batch size of stage 1 256 128
160 × ( 100 128 × ( 100
 2
  2

Batch size of stage 2 & 3 N ) N )
Total epochs 2, 500 1, 000

D. The Results on TSP Instances with Larger-scale

As shown in Table 6, our performance is slightly worse than the two SL-based NCO models, BQ and LEHD. Nevertheless,
on TSP2000 and TSP3000 instances, we achieve the best results in the RL-based constructive methods. And on TSP4000
and TSP5000 instances, we are only slightly worse than ELG. Overall, our method still has a good large-scale generalization
ability. Meanwhile, this observed trend reveals an important research direction: enhancing existing adaptation bias techniques
to sustain and improve the model’s performance in larger-scale TSP instances. Such advancements would enable the model
to effectively tackle more extensive problem spaces while retaining its efficient solution-generation capabilities.

14
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Table 6. Comparative results on TSP instances with scale >1000. “Time” represents the per-instance runtime.
TSP2K TSP3K TSP4K TSP5K
Method Obj. Gap Time (s) Obj. Gap Time (s) Obj. Gap Time (s) Obj. Gap Time (s)
LKH3 32.45 0.000% 144.67 39.60 0.000% 176.13 45.66 0.000% 455.46 50.94 0.000% 710.39
LEHD greedy 34.71 6.979% 5.60 43.79 10.558% 18.66 51.79 13.428% 43.88 59.21 16.237% 85.78
BQ greedy 34.03 4.859% 1.39 42.69 7.794% 3.95 50.69 11.008% 10.50 58.12 14.106% 25.19
POMO 50.89 56.847% 4.70 65.05 64.252% 14.68 77.33 69.370% 35.12 88.28 73.308% 64.46
ELG 36.14 11.371% 6.68 45.01 13.637% 19.66 52.67 15.361% 44.84 59.47 16.758% 81.63
ICAM 34.37 5.934% 1.80 44.39 12.082% 5.62 53.00 16.075% 12.93 60.28 18.338% 24.51

E. Detailed Ablation Study

Please note that, unless stated otherwise, the results presented in the ablation study reflect the best result from multiple
trajectories. We do not employ instance augmentation in this ablation study, and the performance on TSP instances is used
as the primary criterion for evaluation.

E.1. ICAM vs. POMO

Table 7. Comparison of ICAM and POMO on TSP and CVRP instances with different scales in the same training settings.
TSP100 TSP200 TSP500 TSP1000
Method Obj. Gap Time Obj. Gap Time Obj. Gap Time Obj. Gap Time
Concorde 7.7632 0.000% 34m 10.7036 0.000% 3m 16.5215 0.000% 32m 23.1199 0.000% 7.8h
POMO (original) single trajec. 7.8312 0.876% 2s 11.1710 4.367% <1s 22.1027 33.781% 1s 35.1823 52.173% 2s
POMO (original) 7.7915 0.365% 8s 10.9470 2.274% 1s 20.4955 24.053% 9s 32.8566 42.114% 1.1m
POMO (original) aug×8 7.7736 0.134% 1m 10.8677 1.534% 5s 20.1871 22.187% 1.1m 32.4997 40.570% 8.5m
POMO single trajec. 8.1330 4.763% 2s 11.1578 4.243% <1s 17.3638 5.098% 1s 25.1895 8.952% 2s
POMO 7.8986 1.744% 8s 10.9080 1.910% 1s 17.0568 3.240% 9s 24.6571 6.649% 1.1m
POMO aug×8 7.8179 0.704% 1m 10.8272 1.154% 5s 16.9530 2.612% 1.1m 24.5097 6.011% 8.5m
ICAM single trajec. 7.8328 0.897% 2s 10.8255 1.139% <1s 16.7777 1.551% 1s 23.7976 2.931% 2s
ICAM 7.7991 0.462% 5s 10.7753 0.669% <1s 16.6978 1.067% 4s 23.5608 1.907% 28s
ICAM aug×8 7.7747 0.148% 37s 10.7385 0.326% 3s 16.6488 0.771% 38s 23.4854 1.581% 3.8m
CVRP100 CVRP200 CVRP500 CVRP1000
Method Obj. Gap Time Obj. Gap Time Obj. Gap Time Obj. Gap Time
LKH3 15.6465 0.000% 12h 20.1726 0.000% 2.1h 37.2291 0.000% 5.5h 37.0904 0.000% 7.1h
POMO (original) single trajec. 16.0264 2.428% 2s 21.9590 8.856% <1s 50.2240 34.905% 1s 150.4555 305.645% 2s
POMO (original) 15.8368 1.217% 10s 21.3529 5.851% 1s 48.2247 29.535% 10s 143.1178 285.862% 1.2m
POMO (original) aug×8 15.7544 0.689% 1.2m 21.1542 4.866% 6s 44.6379 19.901% 1.2m 84.8978 128.894% 9.8m
POMO single trajec. 16.3210 4.311% 2s 20.9470 3.839% <1s 38.2987 2.873% 1s 39.6420 6.879% 2s
POMO 16.0200 2.387% 10s 20.6380 2.307% 1s 37.8702 1.722% 10s 39.0244 5.214% 1.2m
POMO aug×8 15.8575 1.348% 1.2m 20.4851 1.549% 6s 37.6902 1.238% 1.2m 38.7652 4.515% 9.8m
ICAM single trajec. 16.1868 3.453% 2s 20.7509 2.867% <1s 37.9594 1.962% 1s 38.9709 5.070% 2s
ICAM 15.9386 1.867% 7s 20.5185 1.715% 1s 37.6040 1.007% 5s 38.4170 3.577% 35s
ICAM aug×8 15.8720 1.442% 47s 20.4334 1.293% 4s 37.4858 0.689% 42s 38.2370 3.091% 4.5m

As detailed in Table 7, in our three-stage training scheme, POMO also obtains better performance on large-scale instances
compared to the original model, yet it falls short of ICAM’s performance. In contrast to POMO, ICAM excels in capturing
cross-scale features and perceiving instance-conditioned information. This ability notably enhances model performance in
solving problems across various scales. Detailed information on the model’s performance at different stages can be found in
Appendix E.2.

E.2. The Effects of Different Stages

As illustrated in Table 8, after the first stage, the model performs outstanding performance with small-scale instances but
underperforms when dealing with large-scale instances. After the second stage, there is a marked improvement in the ability
to solve large-scale instances. By the end of the final stage, the overall performance is further improved. Notably, in our

15
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

Table 8. Comparison between different stages on TSP instances with different scales.
TSP100 TSP200 TSP500 TSP1000
After stage 1 0.514% 1.856% 7.732% 12.637%
After stage 2 0.662% 0.993% 1.515% 2.716%
After stage 3 0.462% 0.669% 1.067% 1.907%

ICAM, the capability to tackle small-scale instances is not affected despite the instance scales varying during the training.

E.3. The Effects of Instance-Conditioned Adaptation Function

Table 9. The detailed ablation study on instance-conditioned adaptation function. Here AFM denotes that AAFM removes the adaptation
bias, and CAB is the compatibility with the adaptation bias.
TSP100 TSP200 TSP500 TSP1000
AFM 1.395% 2.280% 4.890% 8.872%
AFM+CAB 0.956% 1.733% 4.081% 7.090%
AAFM 0.514% 0.720% 1.135% 2.241%
AAFM+CAB 0.462% 0.669% 1.067% 1.907%

The data presented in Table 9 indicates a notable enhancement in the solving performance across various scales when
instance-conditioned information, such as scale and node-to-node distances, is integrated into the model. This improvement
emphasizes the importance of including detailed, fine-grained information in the model. It also highlights the critical role of
explicit instance-conditioned information in improving the adaptability and generalization capabilities of RL-based models.
In particular, the incorporation of richer instance-conditioned information allows the model to more effectively comprehend
and address the challenges, especially in the context of large-scale problems.

E.4. Parameter Settings in Stage 3

Table 10. Comparison between different parameters in Stage 3 on TSP instances with 1, 000 nodes.
single trajec. no augment.
β=0 β = 0.1 β = 0.5 β = 0.9 β=0 β = 0.1 β = 0.5 β = 0.9
k = 20 2.996% 2.931% 3.423% 3.480% 2.039% 1.907% 1.859% 1.875%
k = 50 − 3.060% 3.123% 3.328% − 1.935% 1.892% 1.857%
k = 100 − 2.979% 3.201% 3.343% − 1.948% 1.899% 1.899%

As indicated in Table 10, when trained using LJoint as outlined in Equation (11), our model shows further improved
performance. When using the multi-greedy search strategy, we observe no significant performance variation among different
models at various k values. However, increasing the β coefficients, while yielding a marginal improvement in performance
with the multi-greedy strategy, notably diminishes the solving efficiency in the single-trajectory mode. Given the challenges
in generating N trajectories for a single instance as the instance scale increases, we are focusing on optimizing the model
effectiveness specifically in the single trajectory mode to obtain the best possible performance.

E.5. Comparison of Different Inference Strategies

As detailed in Table 11, upon attempting to replace the instance augmentation strategy with alternative inference strategies,
it is observed that there is no significant improvement in the performance of our model. However, the incorporation of
RRC technology into the LEHD model and the implementation of beam search technology into the BQ model both result

16
 Instance-Conditioned Adaptation for Large-scale Generalization of Neural Combinatorial Optimization

in substantial enhancements to the performance of respective models. This highlights a crucial insight: different models
require different inference strategies to optimize their performance. Consequently, it is essential to investigate more effective
strategies to achieve further improvements in the performance of ICAM.

Table 11. Experimental results with different inference strategies on TSP instances.
TSP100 TSP200 TSP500 TSP1000
Method Obj. Gap Time Obj. Gap Time Obj. Gap Time Obj. Gap Time
Concorde 7.7632 0.000% 34m 10.7036 0.000% 3m 16.5215 0.000% 32m 23.1199 0.000% 7.8h
BQ greedy 7.7903 0.349% 1.8m 10.7644 0.568% 9s 16.7165 1.180% 46s 23.6452 2.272% 1.9m
BQ bs16 7.7644 0.016% 27.5m 10.7175 0.130% 2m 16.6171 0.579% 11.9m 23.4323 1.351% 29.4m
LEHD greedy 7.8080 0.577% 27s 10.7956 0.859% 2s 16.7792 1.560% 16s 23.8523 3.168% 1.6m
LEHD RRC100 7.7640 0.010% 16m 10.7096 0.056% 1.2m 16.5784 0.344% 8.7m 23.3971 1.199% 48.6m
ICAM single trajec. 7.8328 0.897% 2s 10.8255 1.139% <1s 16.7777 1.551% 1s 23.7976 2.931% 2s
ICAM 7.7991 0.462% 5s 10.7753 0.669% <1s 16.6978 1.067% 4s 23.5608 1.907% 28s
ICAM RRC100 7.7950 0.409% 2.4m 10.7696 0.616% 14s 16.6886 1.012% 2.4m 23.5488 1.855% 16.8m
ICAM bs16 7.7915 0.365% 1.3m 10.7672 0.594% 14s 16.6889 1.013% 1.5m 23.5436 1.833% 10.5m
ICAM aug×8 7.7747 0.148% 37s 10.7385 0.326% 3s 16.6488 0.771% 38s 23.4854 1.581% 3.8m

F. Licenses for Used Resources

Table 12. List of licenses for the codes and datasets we used in this work
Resource Type Link License
Concorde (Applegate et al., 2006) Code https://github.com/jvkersch/pyconcorde BSD 3-Clause License
LKH3 (Helsgaun, 2017) Code http://webhotel4.ruc.dk/˜keld/research/LKH-3/ Available for academic research use
HGS (Vidal, 2022) Code https://github.com/chkwon/PyHygese MIT License
H-TSP (Pan et al., 2023) Code https://github.com/Learning4Optimization-HUST/H-TSP Available for academic research use
GLOP (Ye et al., 2023) Code https://github.com/henry-yeh/GLOP MIT License
POMO (Kwon et al., 2020) Code https://github.com/yd-kwon/POMO/tree/master/NEW_py_ver MIT License
ELG (Gao et al., 2023) Code https://github.com/gaocrr/ELG MIT License
Pointerformer (Jin et al., 2023) Code https://github.com/pointerformer/pointerformer Available for academic research use
MDAM (Xin et al., 2021) Code https://github.com/liangxinedu/MDAM MIT License
LEHD (Luo et al., 2023) Code https://github.com/CIAM-Group/NCO_code/tree/main/single_objective/LEHD Available for any non-commercial use
BQ (Drakulic et al., 2023) Code https://github.com/naver/bq-nco CC BY-NC-SA 4.0 license
CVRPLib Dataset http://vrp.galgos.inf.puc-rio.br/index.php/en/ Available for academic research use

We list the used existing codes and datasets in Table 12, and all of them are open-sourced resources for academic usage.

17