SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination Yunhao Liang1 Xianqi Cao1 Pujun Zhang∗,1,2 Yuan Qu∗,1,2 Yongzhi Qi3 Ningxuan Kang3 Max Z.J. Shen1,2 1
The University of Hong Kong, Hong Kong SAR, China 2 OptiMax AI, Hong Kong SAR, China 3 JD.com, China
arXiv:2607.28488v1 [cs.AI] 30 Jul 2026
Abstract Can supply-chain AI move beyond isolated decision modules toward unified operational planning? A complete replenishment plan specifies which products each location carries, which upstream facility supplies it, how often it is replenished, and how deliveries are routed. These decisions are operationally coupled: the selected assortment changes the demand and load passed to later stages; source assignment and replenishment frequency reshape the delivery requests; and route feasibility and cost, in turn, determine the system value of the earlier choices. Yet in modern supply chains, these decisions are often handled by separate departments and optimized through separate systems, which can lead to stockouts, inventory exposure, and avoidable transportation. We propose SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination, a composite policy model that represents supply-chain entities as tokens, contextualizes them through a shared operational representation, and maps each token type to the corresponding decision interface. Each decision builds on the partial plan formed by earlier decisions while the completed plan is evaluated using a shared systemlevel utility. We instantiate this framework in urban fresh-retail replenishment, where service frequency, assortment, capacity pressure, and road-network routing interact strongly, and evaluate it on real operational data from Dingdong and JD.com, two large-scale supply chains operating at different replenishment echelons. Across both settings, SCOPE consistently outperforms methods that optimize each decision stage separately, as well as practice-oriented baselines commonly used in supply-chain operations. These results show that learning and coordinating cross-department operational couplings lead to more effective end-to-end supply-chain decisions.
1
Introduction
Supply chains are the operating systems of modern society: they coordinate products, capacity, inventory, and service commitments across organizations, facilities, and regions (Mentzer et al. 2001), and their failures threaten economic security, public health, and the availability of essential goods (Baumgartner, Malik, and Padhi 2020; The White House 2021; Christopher and Peck 2004; Tang 2006). Yet the computational tools that operate these systems remain dominated by an older paradigm: each department solves its own optimization program, a demand forecaster here, a route solver there, while the system-level consequences of their interacting decisions are left unmodeled. Recent literature has there-
demand
demand
Manufacturer
replenishment
Assortment
Distributor
demand
replenishment
Replenish Cycle
Retailer
Routing Plan
replenishment
Customer
Delivery
Latent Operational Couplings
Figure 1: Latent operational coupling in a replenishment supply chain. Assortment and replenishment-cycle decisions shape the demand released to logistics and hence the routing plan and delivery outcome; downstream execution costs, in turn, determine the system value of upstream choices.
fore called for integrating machine learning into supply-chain operations (Baryannis, Dani, and Antoniou 2019; Brintrup et al. 2020). We argue that the central AI challenge is not to forecast demand or optimize a downstream route more accurately, but to learn the couplings through which departmental decisions determine each other’s outcomes and coordinate them under system-level objective. In the real modern supply chain, platforms are organized into functional decision layers: merchandising teams decide what products to carry, replenishment teams decide when and how often inventory moves, and logistics teams decide how vehicles and routes serve the resulting demand. This decomposition is organizationally natural but algorithmically incomplete, echoing a long literature on decentralized supply chains and incentive misalignment (Cachon 2003; Cachon and Zipkin 1999; Lariviere and Porteus 2001): decisions that are locally optimal for one layer routinely shift cost, risk, or capacity burden onto another. Existing tools mirror this decomposition: assortment models, partial inventory–routing programs, and neural route solvers optimize local decision surfaces under fixed or hand-specified cross-stage relationships(Kök, Fisher, and Vaidyanathan 2015; Coelho, Cordeau, and Laporte 2014; Shaabani 2022). What they lack is a system view of the decision pipeline: existing methods optimize departments in isolation, ignoring upstream effects on downstream problems and whole system value. We call this missing cross-stage dependency latent operational coupling (Figure 1). For exam-
ple, a high-demand product may become unattractive once its capacity consumption and routing burden are priced in, while a longer replenishment cycle may reduce transportation cost but increase inventory exposure and lost-sales risk; similar tradeoffs arise throughout the pipeline. Because they are not directly observed, locally reasonable decisions can silently transfer cost and risk across the system (Forrester 1997; Sterman 1989; Lee, Padmanabhan, and Whang 1997). Existing methods either optimize within a fixed downstream instance or hand-specify fragments of the coupling; they do not learn and coordinate it across the whole supply-chain pipeline. The present work. We formulate retail replenishment as a decision-learning problem with latent operational coupling. Assortment, source assignment, and replenishment interval decisions determine which demand is served, where it originates, and when it is released. Through known operational rules, these decisions induce the routing instances and feasible route sets that logistics must handle. We formalize an upstream plan through its induced system value, namely the best complete-pipeline utility attainable after completing its remaining decisions. To our best knowledge, SCOPE (Supply-Chain Operations through Coupled Policies for End-to-End Coordination) is the first learned decision framework to jointly reason over supply-chain entities and produce coordinated end-to-end operational plans under a shared system-level objective. It is designed around two objectives: (1) learning the latent operational couplings through which upstream choices reshape downstream problems and attainable system value, and (2) translating these couplings into complete, inspectable operational decisions.SCOPE encodes products, facilities, demand, geography, capacity, and cost context in a shared operational representation, while each policy interface extends the partial plan formed by preceding decisions. This design is related to multi-task representation learning (Caruana 1997; Ruder 2017; Bommasani et al. 2021), but targets end-to-end operational planning. We validate SCOPE on two operational regimes chosen to place the same couplings at different structural positions. The first uses FreshRetailNet-50K, a public fresh-retail dataset (Wang et al. 2025). The second, through collaboration with JD.com, uses a nationwide warehouse network in which selected demand must first be assigned from regional distribution centers (RDCs) to front distribution centers (FDCs). We train separate dataset-specific checkpoints of the same architecture: the experiments test whether the backboneand-interface design absorbs a topological change through a single added interface, so the claim is structural reuse. Every method is evaluated as a complete pipeline against decomposed baselines with rule-based upstream decisions and strong classical, industrial, metaheuristic, and neural route solvers, together with matched ablations and proxy stress tests. In summary, our contributions are: 1. Problem: a new decision-learning formulation of retail replenishment, in which upstream decisions induce downstream routing problems and complete plans are evaluated under a shared system-level utility;
2. Model: we introduce SCOPE model, which represents supply-chain entities as tokens, contextualizes them through a shared operational representation, and maps each token type to a corresponding decision interface for end-to-end coordination; 3. Concept: latent operational coupling, a modeling language for the implicit cross-stage dependencies among decisions over supply-chain entities; 4. Results: on real Dingdong and JD.com data, SCOPE outperforms strong decomposed pipelines at both warehouse and store levels and remains robust across broad proxy settings. Social impact. This work is a collaboration with JD.com, one of the world’s largest e-commerce companies, whose nationwide fulfillment network motivated the problem: departmental optimizers for selection, replenishment, and logistics interact daily, yet no learned representation connects their decisions. Together we identified two objectives grounded in real operational needs. Learning operational couplings improves cross-department visibility, letting planners see how assortment and service-frequency choices propagate into capacity pressure, inventory exposure, and routing burden before costs materialize. Complete-pipeline decision learning supports joint improvement of replenishment frequency, warehouse assignment, and route execution. The FreshRetailNet experiments are fully reproducible from public data, and the JD evaluation preserves an auditable protocol despite data-sharing constraints, so that research on coupled supply-chain decision learning can continue beyond this collaboration.
2
Related Work
Route-centric vehicle-routing solvers. Vehicle routing spans classical formulations, constructive heuristics, and rich VRP variants (Dantzig and Ramser 1959; Clarke and Wright 1964; Laporte 1992; Toth and Vigo 2014; Vidal et al. 2013). Representative route optimizers include nearest-neighbor construction (Rosenkrantz, Stearns, and Lewis 1977), industrial toolkits such as OR-Tools (Furnon and Perron 2025), and metaheuristics such as Hybrid Genetic Search (Vidal 2022). Recent neural methods learn route-construction policies through attention, multi-start reinforcement learning, mixture-of-experts, or unified node–edge representations(Kool, van Hoof, and Welling 2019; Kwon et al. 2020; Zhou et al. 2024; Meng et al. 2025). We include representative methods from these families as route components in our baseline pipelines. We include representative methods from these families in our baseline pipelines. Machine learning for supply-chain operations. Recent surveys document growing applications of machine learning in supply-chain management (Baryannis, Dani, and Antoniou 2019; Younis, Sundarakani, and Alsharairi 2022). Existing work mainly addresses prediction and visibility, including supply-chain risk prediction(Baryannis, Dani, and Antoniou 2019), supplier-disruption prediction(Brintrup et al. 2020), and graph learning for hidden-link prediction and planning benchmarks(Aziz et al. 2021; Kosasih and Brintrup 2021; Wasi, Islam, and Akib 2024). A separate line learns policies
staged training: assortment supervision · cycle counterfactuals · router Adapter Interface
Operational state Demand + SKU features
Shared operational encoder
Aselection
Assortment selection facility × SKU
Selected SKUs
Net system value + Selected Products Value - Transport Cost
Token Type Sequence
Select Depot/RDC
Operational Unit
- Inventory Cost
Store/FDC
- Service Cost
Replenishment cycle Road-network
Feature + Type Embedding (demand, inventory, location, time .....)
SINGLE DEPOT
Adaptive T*
αhold
evaluate the complete pipeline
Transformer Encoder
...
Cost context cveh
Acycle
MULTI-RDC
Aggregate
cveh/α ...
...
E = [ Edepot, Eoperational, Estore ]
Multi-day rollout
Logistics
Shared Representation E
......
Execute
Alogistic
0
CVRP
cost embedding
45
90
Daily Inventory and dispatch induced replenishment instances
Figure 2: SCOPE as a coupled supply-chain policy-optimization framework. A shared operational decision backbone supports department-facing policy interfaces for assortment selection, replenishment-cycle planning, and routing under a unified operational objective.
for individual operational layers, including lost-sales, dualsourcing, multi-echelon, and beer-game inventory problems (Gijsbrechts et al. 2022; Oroojlooyjadid et al. 2022; Boute et al. 2022), as well as end-to-end replenishment(Qi et al. 2020). Chang et al. (2025) learn hidden production functions from transaction-level supply-chain graphs. These methods, however, do not jointly coordinate decisions across supplychain departments. Integrated supply-chain decision models. Operations research has long recognized that supply-chain decisions are coupled, including assortment planning (Kök, Fisher, and Vaidyanathan 2015), joint location–inventory models (Shen, Coullard, and Daskin 2003), inventory-routing (Coelho, Cordeau, and Laporte 2014), perishable inventory-routing (Shaabani 2022), and integrated assortment–inventory or assortment–shelf-space–supplier models (Martínez-de Albéniz and Kunnumkal 2022; Sajadi and Ahmadi 2022). Shen (2007) surveys the broader integrated supply-chain design literature, including models in which facility-location decisions account jointly for inventory and distribution costs. Supply-chain coordination further studies how decentralized decisions can be aligned through contracts or related mechanisms (Cachon 2003). These works establish the value of coordination, but typically hand-specify the coupling, cover only subsets of the operational pipeline, and rely on business quantities unavailable in many operational datasets. Positioning. Unlike route-centric solvers, we treat routing instances as outputs of learned upstream decisions and evaluate the complete supply-chain pipeline. Unlike prior supplychain machine learning, we learn latent operational couplings and place them in joint policy optimization across depart-
ments. Unlike integrated optimization models, we learn a shared operational representation from demand, product, geography, and cost context, while handling unavailable business quantities through explicit proxies. SCOPE therefore formulates supply-chain operations as coupled decision learning, where multiple decision interfaces are built on a shared representation and optimized toward a unified systemlevel objective.
3 3.1
SCOPE Framework
Problem Formulation
Retail replenishment comprises four dependent planning stages. A stage denotes decision dependency, not elapsed time. Stage 1 selects the assortment, which determines the demand and load considered when choosing a supply source. Stage 2 assigns supply origins, which changes the sourcedependent inputs to interval planning and routing. Stage 3 chooses replenishment intervals, where an interval is the number of days between successive planned replenishments. This decision determines service days and visit loads. Stage 4 constructs routes for the resulting requests. Each nonterminal decision therefore changes the problem faced by later stages and the utility attainable by the complete plan. Downstream locations are indexed by i, upstream supply nodes by j, and SKUs by p. A scenario s contains facilities, physical road distances, demand dip , SKU attributes, eligible supply links, vehicle limits, and replenishment units. The cost context c specifies the dispatch, travel, and inventory exposure tradeoffs known at planning time. For each location i, assortment Ai contains exactly K
SKUs, and A collects all local assortments. Source plan σ assigns each location to one eligible upstream node: σi ∈ Ji , where σi = j means that j supplies i. Locations are partitioned into fixed replenishment units Mz . Requests within a unit are released on the same service days. Interval Tz controls their frequency and accumulated demand. We use T for the complete interval plan and T (s) for its feasible set. Let wp collect the available weight and volume load factors or proxies for SKU p. The selected daily load and nominal load for one replenishment visit are X q̄i (A) = dip wp , (1) p∈Ai qi (A, T ) = Tz q̄i (A), i ∈ Mz . A short interval creates frequent deliveries; a long interval creates larger loads and greater inventory exposure. These quantities are planning proxies and may differ from realized freight. Given (A, σ, T ), let R(s; A, σ, T ) be the feasible set of complete route plans over the planning horizon. It enforces the selected requests, assigned origins, service days, physical road distances, and vehicle weight and volume limits. Thus, A determines what demand moves, σ determines where it originates, and T determines when and at what visit load it is delivered. These decisions jointly determine the routing problems and their feasible routes. Let Y(s) contain exactly the complete plans y = (A, σ, T, R) satisfying |Ai | = K and σi ∈ Ji for every i, T ∈ T (s), and R ∈ R(s; A, σ, T ). We evaluate y by the daily proxy utility U (s, c; y) = V (A) − Lcov (A) − Crt (R; c) − Cinv (A, T ; c), (2) where V measures assortment value, Lcov penalizes expected demand outside the selected assortment, Crt is transportation cost, and Cinv measures interval-induced inventory exposure. All horizon totals are divided by the horizon length before entering U . Lcov is an uncovered-demand proxy, distinct from realized lost sales in a dynamic inventory simulation. For each decision prefix below, Y(s | ·) contains the complete plans in Y(s) that preserve the decisions already fixed. We use the same operational completion value at all three boundaries: Qs,c (A) := Qs,c (A, σ) := Qs,c (A, σ, T ) :=
max U (s, c; y), y∈Y(s|A)
max
U (s, c; y),
y∈Y(s|A,σ)
max
(3)
U (s, c; y).
y∈Y(s|A,σ,T )
The three lines evaluate one value concept. The remaining decisions are (σ, T, R), (T, R), and R, respectively. Thus, each earlier decision changes both the remaining completion problem and the complete-plan value attainable from it. We call this dependence operational coupling. Assortment changes the demand and load passed to later decisions; source assignment changes origins and source-dependent inputs; and interval planning changes service days, visit loads, and feasible routes. Alternative completion problems and
their optimized values are not recorded in operational data and are costly to compute. For one scenario, the operational problem is y ∗ (s, c) ∈ arg max U (s, c; y), y∈Y(s)
(4)
whose optimal value is maxA: |Ai |=K ∀i Qs,c (A). Repeating this search as conditions change motivates a conditional policy Πθ : θ∗ ∈ arg max θ
s.t.
E(s,c)∼P [U (s, c; Πθ (s, c))] Πθ (s, c) ∈ Y(s).
(5)
Equation (5) defines the policy-level target. SCOPE approximates this target with staged surrogates and selects among the resulting composite policies using validation pipeline utility.
3.2
Model Overview
SCOPE operationalizes this coupling with a conditional composite policy. Each stage extends the current decision configuration: A conditions source assignment, interval planning, and routing; σ further conditions interval planning and routing; and T completes the input to routing. SCOPE then groups the selected requests by assigned source and service day, and a neural routing policy produces R. The resulting complete plan is evaluated by the common utility U . Figure 1 links this flow to Eq. (4). Scenario representation and sequential policies. A shared operational backbone Bθ maps facilities and physical road distances from s to E = Bθ (s). This embedding set contains hi for downstream locations and hj for upstream nodes. A second encoder maps c to zc = eθ (c). The shared representation supplies scenario context; cross-stage dependence enters through the explicit conditioning on earlier decisions. We use θ for all trainable parameters. Superscripts A, σ, T , and R label an assortment scorer, source scorer, interval classifier, and autoregressive routing decoder. The complete policy follows the same order as the formulation: A = πθA (s, E, zc ), σ = πθσ (s, E, zc , A), (6) T = πθT (s, E, zc , A, σ), R = πθR (s, E, A, σ, T ). These steps return Πθ (s, c) = (A, σ, T, R). SCOPE does not calculate the ideal completion value in Eq. (3). Instead, stagespecific surrogates train the conditional policies, and validation utility selects their composition. The same construction applies to other staged systems in which each earlier decision reshapes the remaining problem. The assortment scorer combines hi , projected SKU attributes up = Pθ (xp ), cost context, and demand, value, and load features, then selects the top K SKUs. For source assignment, the scorer compares eligible nodes using facility embeddings, link features, and selected load. Interval selection pools these features within Mz and chooses one candidate τ ∈ {1, . . . , Tmax }. Routing problem construction and solution. Before decoding, SCOPE groups the selected requests by service day
and assigned origin. Each resulting routing subproblem specifies its active locations, loads, vehicle limits, and physical distance matrix. City scenarios use fixed spatial grouping; warehouse routing subproblems are partitioned by σ. The decoder uses capacity masks and multi-start decoding, retaining the lowest-cost route among its feasible rollouts under the implemented metric. It therefore serves as a feasible approximate executor, and its realized utility need not attain the maximum in Eq. (3).
3.3
Model Training
SCOPE approximates Eq. (5) with policy-specific surrogates and validation pipeline utility. Upstream supervision. The assortment policy first learns from SKU targets that combine coverage, value, load, and inventory exposure proxies. It then evaluates one-for-one swaps through the complete downstream pipeline and retains a swap only when proxy utility improves while the required coverage and value are preserved. These comparisons provide local preferences among nearby assortments. The source policy considers only eligible nodes and uses the nearest feasible node as a geometric warm-start target. This promotes a physically plausible assignment but does not identify the source plan with the highest operational completion value. For each training scenario and cost context, SCOPE also evaluates candidate intervals with one interval shared by all active units. The best globally fixed interval provides a common warm-start label, while the deployed policy predicts a separate interval for each unit. This scan supplies no unit-specific alternative labels. Assortment replacements and global interval candidates receive pipeline comparisons; source assignment receives feasibility-constrained geometric supervision. Router specialization. For each day-specific routing subproblem ρ, OR-Tools is used offline to produce a reference sequence rρref . The route decoder is trained by imitation: XX ref ref LR = − log πθR (rρ,m | rρ,<m , E, A, σ, T ). (7) ρ
m
Each subproblem contains the source, service day, selected requests, loads, and vehicle limits determined by the earlier decisions. Route imitation supplies no full-pipeline value target for alternative decision configurations, and the resulting routes remain approximate. At inference, SCOPE executes these subproblems without OR-Tools. Full pipeline selection. The implementation keeps a separate loss for each policy. We compare candidate checkpoints C by running the complete policy on the validation set: X 1 θb ∈ arg max U (s, c; Πθ (s, c)) . (8) θ∈C |Dval | (s,c)∈Dval
Branch-specific surrogates fit candidate policies. Validation utility U then selects among the resulting composite checkpoints without supplying gradients to the policy heads. Proxy definitions, normalization, and surrogate targets appear in the experimental protocol or appendix. Thus, SCOPE coordinates the conditional policies using complete-pipeline evaluations without explicitly estimating Qs,c in Eq. (3).
4
Experiments
Replenishment recurs at every echelon of a supply chain in the same abstract form, a shared decision grammar: an upstream supply node serves a set of downstream demand nodes, and the operator must decide what to serve, how often to replenish, and how to route the deliveries. This recurring structure motivates our evaluation along three dimensions. (i) Coupling: do the latent operational couplings among departmental decisions matter, i.e., does learning these couplings and optimizing the complete assortment–cycle– routing pipeline jointly outperform decomposed pipelines that pair rule-based upstream decisions with strong route solvers? (ii) Generality: does the same joint model remain effective across supply-chain scales and echelons? (iii) Mechanism and limits: which learned interfaces drive the gains, and under which cost regimes do they fail? The main results address coupling and generality; the ablations and robustness studies isolate the underlying mechanisms and failure regimes.
4.1
Data
We ground this evaluation in fresh retail, where assortment, replenishment frequency, inventory exposure, and distribution are tightly coupled (van Donselaar et al. 2006; Minner and Transchel 2010; Shaabani 2022), using two industrial level operational datasets at different supply-chain echelons. At the depot-to-store echelon, FreshRetailNet-50K (Wang et al. 2025) is a rare public record of large-scale fresh-retail operations from Dingdong, spanning 898 stores, 18 cities, 865 SKUs, and 90 days. At the upper echelon of supply chain scenario, our JD.com data capture a nationwide, industrial multi-warehouse network with 8 RDCs, 43 FDCs, 1,110 SKUs, and 91 dated snapshots. Such facility-level inventory, demand, and network records are rarely available for academic evaluation; they expose warehouse-scale assortment and an assignment-dependent multi-depot topology, providing a stringent test of whether the same model generalizes beyond last-leg replenishment. Under the NDA, identities are masked and business quantities rescaled while the operational decision structure is preserved. Figure 3 previews the public dataset’s heterogeneity in store–SKU coverage, category mix, demand, and stockout signals. Full dataset details are provided in the Supplementary Material.
4.2
Experimental Setup
Every method is evaluated as a complete assortment– cycle–routing pipeline over the full planning horizon, under one shared objective. We report mean reference-adjusted operational value Uref (larger is better), together with transportation-plus-inventory cost, demand coverage, and vehicles per day. SCOPE is one learned joint pipeline; baselines are decomposed pipelines that cross three assortment rules (TopDemand-K, TopValue-K, Random-K), globally scanned fixed cycles (BestFix / Daily), and seven route executors ranging from nearest-neighbor and OR-Tools (Furnon and Perron 2025) to recent neural routers (Vidal 2022; Kool, van Hoof, and Welling 2019; Kwon et al. 2020; Zhou et al. 2024; Meng et al. 2025). The top-K assortment rules mirror JD.com’s deployed prediction-then-ranking practice (Shen et al. 2025), so
Dingdong (depot–store) Pipeline
Assort.
Cycle
SCOPE
✓
✓
Uref ↑
T
✓ 206329.79
JD.com (RDC–FDC) Uref ↑
Gap↓
Cov.↑
T
–
0.948
✓ 293547.12
Gap↓
Cov.↑
–
0.427
TopDemand assortment with best global fixed cycle TopDemand–OR-Tools TopDemand-K BestFix TopDemand–NN TopDemand-K BestFix TopDemand–HGS TopDemand-K BestFix TopDemand–UniteFormer TopDemand-K BestFix TopDemand–AM TopDemand-K BestFix TopDemand–POMO TopDemand-K BestFix TopDemand–MVMoE TopDemand-K BestFix
6 6 7 7 7 7 7
206276.14 206271.86 206253.31 206248.27 206227.23 206197.59 206180.81
53.65 57.93 76.48 81.51 102.56 132.20 148.98
0.948 0.948 0.948 0.948 0.948 0.948 0.948
2 2 2 3 4 4 3
288734.46 4812.65 287021.11 6526.01 288736.72 4810.39 287320.33 6226.78 282564.81 10982.30 285398.48 8148.64 284890.42 8656.70
0.429 0.429 0.429 0.429 0.429 0.429 0.429
Daily replenishment baselines Daily–HGS TopDemand-K Daily–OR-Tools TopDemand-K Daily–NN TopDemand-K
1 205932.38 1 205931.78 1 205911.85
397.41 398.01 417.94
0.948 0.948 0.948
1 283074.28 10472.83 1 283074.28 10472.83 1 280156.60 13390.51
0.429 0.429 0.429
Daily Daily Daily
Alternative assortment policies with best global fixed cycle TopValue–OR-Tools TopValue-K BestFix 6 205569.95 759.84 0.945 TopValue–HGS TopValue-K BestFix 6 205556.16 773.63 0.945 Random–OR-Tools Random-K BestFix 6 170016.63 36313.15 0.888 Random–HGS Random-K BestFix 7 170003.01 36326.78 0.888
2 288973.15 4573.96 0.427 2 288973.15 4573.96 0.427 5 65370.22 228176.90 0.103 5 65371.79 228175.32 0.103
Table 1: Complete-pipeline results on depot-to-store (Dingdong, 18 90-day scenarios) and RDC-to-FDC (JD.com, 18 91-day SCOPE method real-FDC subset scenarios). Uref is larger-is-better; Gap= Uref −Uref . BestFix is one global fixed cycle over T = 1, . . . , 7, not a per-episode oracle; T is the selected cycle on each dataset. A checkmark denotes a learned decision; Gap “–” means not applicable.
they are industrial baselines. Cost coefficients follow reported Chinese cold-chain vehicle costs and textbook carrying-cost conventions (Feng et al. 2024; Pan et al. 2024; Silver, Pyke, and Peterson 1998). Remaining details are in the Supplementary Material.
4.3
Main Results Across Two Replenishment Echelons
Table 1 reports the complete-pipeline comparison on both echelons. Across the Dingdong scenarios, SCOPE achieves the highest Uref (206329.79), exceeding every decomposed combination of a top-ranked assortment rule, a globally optimized fixed cycle, and a strong classical or neural router. This is a demanding comparison: prediction-then-ranking assortment is used in JD.com’s deployed fulfillment practice (Shen et al. 2025), while OR-Tools, HGS, and the learned routers are competitive routing methods. Yet, with TopDemand-K and BestFix held fixed, replacing OR-Tools with any of the six alternative routers changes the gap only from 53.65 to 148.98; by comparison, imposing daily replenishment increases it to 398.01, and changing the assortment policy increases it further. Thus, optimizing strong departmental policies separately is insufficient: once assortment and cycle have formed the routing instance, a stronger router cannot recover value lost through their interaction. SCOPE’s consistent lead provides direct evidence for the first question, that learning the latent operational couplings among assortment, replenishment frequency, and routing matters. The JD.com results then
test whether this finding persists beyond the last leg. Despite moving to a larger RDC-to-FDC network with warehousescale SKU volumes and assignment-based instance formation, SCOPE again ranks first, reaching 293547.12 and outperforming the strongest decomposed pipeline by 4573.96 (1.58%) at essentially the same coverage. The same pattern is more pronounced there: daily replenishment loses 10472.83, whereas changing the router within a fixed upstream policy cannot close the gap. Standing out on both a public depot-tostore benchmark and a proprietary industrial RDC-to-FDC network shows that the benefit is not tied to one dataset, scale, or replenishment echelon, answering the second question on generality. Full matrices and implementation details are provided in the Supplementary Material.
4.4
Ablations
Table 2 reports matched decision ablations on both datasets, each replacing one learned interface or one training stage while keeping the rest of the pipeline fixed. The decision ablations reveal a consistent pattern across the two replenishment echelons. Removing the learned interfaces produces the largest performance degradation on both Dingdong (−426.36) and JD.com (−14520.74), while removing counterfactual cycle supervision causes a further substantial loss on JD.com (−4687.86). Replacing the learned assortment policy with TopDemand-K has a smaller but consistently negative effect. These results suggest that system value arises from learning the latent operational couplings among assort-
Dingdong (depot–store) Variant
Change
Uref ↑
∆
SCOPE full staged pipeline 206329.79 0.00 w/o assortment interface replace with TopDemand-K 206275.89 -53.90 w/o service-cycle interface fixed daily cycle (T =1) 205903.42 -426.36 w/o learned-router remove reference-route warm-up 206319.11 -10.68 w/o cycle supervision remove counterfactual training 206292.13 -37.65 Direct joint training train all modules from scratch 173907.60 -32422.18
JD.com (RDC–FDC)
Cov.↑ 0.948 0.948 0.948 0.948 0.948 0.894
Uref ↑
∆
293547.12 0.00 293261.53 -285.59 279026.38 -14520.74 292738.46 -808.65 288859.25 -4687.86 293472.61 -74.50
Cov.↑ 0.427 0.429 0.427 0.427 0.427 0.427
Table 2: Decision ablations under the same protocols as Table 1 (Dingdong: 18 90-day scenarios; JD.com: 18 91-day scenarios). Uref is larger-is-better; ∆ is the signed change relative to full SCOPE. Assignment remains part of the learned JD.com pipeline but is not treated as a separate ablation.
magnitude of its effect varies across data regimes.
4.5
Robustness and Failure Regimes
To test whether the gains depend on proxy calibration, we keep each checkpoint fixed and perturb the assortment-value, route-cost, holding-cost, and load scales. SCOPE ranks first in 30 of 35 value–cost settings and all seven load settings on Dingdong, and in all 35 value–cost settings and five of seven load settings on JD.com. Losses on Dingdong occur only at a 0.25× assortment-value scale, where the objective approaches routing alone; those on JD.com occur when load perturbations change the induced capacity regime. Thus, the coupling advantage is robust across broad proxy ranges but can weaken when upstream value becomes negligible or the capacity regime shifts. Complete grids and numerical results are reported in the Supplementary Material.
5 Figure 3: FreshRetailNet-50K exhibits the heterogeneity that couples the three decisions: (a) skewed store-SKU footprint across 18 cities, (b) city-specific category mix, (c) long-tailed per-series demand volatility, (d) non-stationary demand with a censored stockout signal.
ment, replenishment, and routing decisions. The ablations therefore support the central claim of Table 1: end-to-end performance depends on coordinating these interdependent decisions rather than optimizing each department in isolation. The training ablation highlights a practical difficulty in real supply-chain optimization. Because assortment, replenishment cycle, and routing jointly determine the final system outcome, a performance drop is difficult to attribute to any single decision module or to a poor combination of the three. On Dingdong, direct joint training from scratch converges to a substantially worse, lower-coverage policy, losing 32422.18 value units and reducing coverage from 0.948 to 0.894. On JD.com, however, direct joint training finishes only 74.50 behind staged SCOPE. Together, these results validate the training ablation by showing that the optimization strategy can materially affect the learned joint policy, although the
Conclusion
This paper studies supply-chain operations as a coordinated decision-learning problem and introduces SCOPE, a composite policy model that represents supply-chain entities as typed tokens, contextualizes them through a shared operational representation, and realizes assortment, source assignment, replenishment-cycle planning, and routing as sequentially coupled decision interfacese. SCOPE addresses two central objectives: learning the latent operational couplings through which upstream decisions reshape downstream problems, and using these couplings to produce complete, inspectable plans under a shared system-level utility—whereas prior methods typically optimize a fixed downstream instance or hand-specify only part of the coupling. To demonstrate these capabilities, we evaluated SCOPE on city-scale Dingdong replenishment and a nationwide JD.com warehouse network, where it consistently outperformed decomposed operational pipelines across two replenishment echelons. More broadly, this work suggests a path toward supply-chain AI systems that improve whole-pipeline reliability.
References Aziz, A.; Kosasih, E. E.; Griffiths, R.-R.; and Brintrup, A. 2021. Data Considerations in Graph Representation Learning for Supply Chain Networks. arXiv preprint arXiv:2107.10609.
Baryannis, G.; Dani, S.; and Antoniou, G. 2019. Predicting Supply Chain Risks Using Machine Learning: The Trade-Off between Performance and Interpretability. Future Generation Computer Systems, 101: 993–1004. Baumgartner, T.; Malik, Y.; and Padhi, A. 2020. Reimagining Industrial Supply Chains. McKinsey & Company. Bommasani, R.; Hudson, D. A.; Adeli, E.; et al. 2021. On the Opportunities and Risks of Foundation Models. arXiv preprint arXiv:2108.07258. Boute, R. N.; Gijsbrechts, J.; van Jaarsveld, W.; and Vanvuchelen, N. 2022. Deep Reinforcement Learning for Inventory Control: A Roadmap. European Journal of Operational Research, 298(2): 401–412. Brintrup, A.; Pak, J.; Ratiney, D.; Pearce, T.; Wichmann, P.; Woodall, P.; and McFarlane, D. 2020. Supply Chain Data Analytics for Predicting Supplier Disruptions: A Case Study in Complex Asset Manufacturing. International Journal of Production Research, 58(11): 3330–3341. Cachon, G. P. 2003. Supply Chain Coordination with Contracts. In Handbooks in Operations Research and Management Science: Supply Chain Management, volume 11, 227–339. Elsevier. Cachon, G. P.; and Zipkin, P. H. 1999. Competitive and Cooperative Inventory Policies in a Two-Stage Supply Chain. Management Science, 45(7): 936–953. Caruana, R. 1997. Multitask Learning. Machine Learning, 28: 41–75. Chang, S.; Lin, Z.; Yan, B.; Bembde, S.; Xiu, Q.; Wong, C. H.; Qin, Y.; Kloster, F.; Luo, X.; Palleti, R.; and Leskovec, J. 2025. Learning Production Functions for Supply Chains with Graph Neural Networks. Proceedings of the AAAI Conference on Artificial Intelligence, 39(27): 27878–27886. Christopher, M.; and Peck, H. 2004. Building the Resilient Supply Chain. The International Journal of Logistics Management, 15(2): 1–14. Clarke, G.; and Wright, J. W. 1964. Scheduling of Vehicles from a Central Depot to a Number of Delivery Points. Operations Research, 12(4): 568–581. Coelho, L. C.; Cordeau, J.-F.; and Laporte, G. 2014. Thirty Years of Inventory Routing. Transportation Science, 48(1): 1–19. Dantzig, G. B.; and Ramser, J. H. 1959. The Truck Dispatching Problem. Management Science, 6(1): 80–91. Feng, Q.; Hua, W.; Chen, C.; Shi, X.; and Ma, J. 2024. Optimization of Cold Chain Logistics Distribution Paths for Community Buying Considering Cost and Time Window: The Case of J Center. Managerial and Decision Economics, 45(8): 5249–5264. Forrester, J. W. 1997. Industrial Dynamics. Journal of the Operational Research Society, 48: 1037–1041. Furnon, V.; and Perron, L. 2025. OR-Tools Routing Library. Gijsbrechts, J.; Boute, R. N.; Van Mieghem, J. A.; and Zhang, D. J. 2022. Can Deep Reinforcement Learning Improve Inventory Management? Performance on Lost Sales, DualSourcing, and Multi-Echelon Problems. Manufacturing & Service Operations Management, 24(3): 1349–1368.
Kök, A. G.; Fisher, M. L.; and Vaidyanathan, R. 2015. Assortment Planning: Review of Literature and Industry Practice. In Retail Supply Chain Management, 175–236. Springer. Kool, W.; van Hoof, H.; and Welling, M. 2019. Attention, Learn to Solve Routing Problems! In International Conference on Learning Representations. Kosasih, E. E.; and Brintrup, A. 2021. A Machine Learning Approach for Predicting Hidden Links in Supply Chain with Graph Neural Networks. International Journal of Production Research, 60: 5380–5393. Kwon, Y.-D.; Choo, J.; Kim, B.; Yoon, I.; Gwon, Y.; and Min, S. 2020. POMO: Policy Optimization with Multiple Optima for Reinforcement Learning. In Advances in Neural Information Processing Systems. Laporte, G. 1992. The Vehicle Routing Problem: An Overview of Exact and Approximate Algorithms. European Journal of Operational Research, 59(3): 345–358. Lariviere, M. A.; and Porteus, E. L. 2001. Selling to the Newsvendor: An Analysis of Price-Only Contracts. Manufacturing & Service Operations Management, 3(4): 293–305. Lee, H. L.; Padmanabhan, V.; and Whang, S. 1997. Information Distortion in a Supply Chain: The Bullwhip Effect. Management Science, 43(4): 405–570. Martínez-de Albéniz, V.; and Kunnumkal, S. 2022. A Model for Integrated Inventory and Assortment Planning. Management Science, 68(7): 5049–5067. Meng, D.; Cao, Z.; Gao, J.; Wu, Y.; and Hou, Y. 2025. UniteFormer: Unifying Node and Edge Modalities in Transformers for Vehicle Routing Problems. In Advances in Neural Information Processing Systems. Mentzer, J. T.; DeWitt, W.; Keebler, J. S.; Min, S.; Nix, N. W.; Smith, C. D.; and Zacharia, Z. G. 2001. Defining Supply Chain Management. Journal of Business Logistics, 22(2): 1–25. Minner, S.; and Transchel, S. 2010. Periodic Review Inventory-Control for Perishable Products under ServiceLevel Constraints. OR Spectrum, 32(4): 979–996. Oroojlooyjadid, A.; Nazari, M.; Snyder, L. V.; and Takáč, M. 2022. A Deep Q-Network for the Beer Game: Deep Reinforcement Learning for Inventory Optimization. Manufacturing & Service Operations Management, 24(1): 285–304. Pan, S.; Liao, H.; Zheng, G.; Huang, Q.; and Shan, M. 2024. Cold Chain Distribution Route Optimization for Mixed Vehicle Types of Fresh Agricultural Products Considering Carbon Emissions: A Study Based on a Survey in China. Sustainability, 16(18): 8207. Qi, M.; Shi, Y.; Qi, Y.; Ma, C.; Yuan, R.; Wu, D.; and Shen, Z.-J. M. 2020. A Practical End-to-End Inventory Management Model with Deep Learning. Management Science, 69: 759–773. Rosenkrantz, D. J.; Stearns, R. E.; and Lewis, P. M. 1977. An Analysis of Several Heuristics for the Traveling Salesman Problem. SIAM Journal on Computing, 6(3): 563–581. Ruder, S. 2017. An Overview of Multi-Task Learning in Deep Neural Networks. arXiv preprint arXiv:1706.05098.
Sajadi, S. J.; and Ahmadi, A. 2022. An Integrated Optimization Model and Metaheuristics for Assortment Planning, Shelf Space Allocation, and Inventory Management of Perishable Products: A Real Application. PLOS ONE, 17(3): e0264186. Shaabani, H. 2022. A Literature Review of the Perishable Inventory Routing Problem. The Asian Journal of Shipping and Logistics. Shen, Z.-J. M. 2007. Integrated Supply Chain Design Models: A Survey and Future Research Directions. Journal of Industrial and Management Optimization, 3(1): 1–27. Shen, Z.-J. M.; Coullard, C.; and Daskin, M. S. 2003. A Joint Location-Inventory Model. Transportation Science, 37(1): 40–55. Shen, Z.-J. M.; Sun, S.; Qi, Y.; Hu, H.; Kang, N.; Zhang, J.; Wang, X.; and Lin, X. 2025. JD.com Improves Fulfillment Efficiency with Data-Driven Integrated Assortment Planning and Inventory Allocation. INFORMS Journal on Applied Analytics, 55(5): 386–398. Silver, E. A.; Pyke, D. F.; and Peterson, R. 1998. Inventory Management and Production Planning and Scheduling. John Wiley & Sons, 3rd edition. Sterman, J. D. 1989. Modeling Managerial Behavior: Misperceptions of Feedback in a Dynamic Decision Making Experiment. Management Science. Tang, C. S. 2006. Perspectives in Supply Chain Risk Management. International Journal of Production Economics, 103(2): 451–488. The White House. 2021. Building Resilient Supply Chains, Revitalizing American Manufacturing, and Fostering BroadBased Growth. 100-Day Reviews under Executive Order 14017. Toth, P.; and Vigo, D., eds. 2014. Vehicle Routing: Problems, Methods, and Applications. SIAM. van Donselaar, K.; van Woensel, T.; Broekmeulen, R.; and Fransoo, J. 2006. Inventory Control of Perishables in Supermarkets. International Journal of Production Economics, 104(2): 462–472. Vidal, T. 2022. Hybrid Genetic Search for the CVRP: OpenSource Implementation and SWAP* Neighborhood. Computers & Operations Research, 140: 105643. Vidal, T.; Crainic, T. G.; Gendreau, M.; and Prins, C. 2013. Heuristics for Multi-Attribute Vehicle Routing Problems: A Survey and Synthesis. European Journal of Operational Research, 231(1): 1–21. Wang, Y.; Gu, J.; Long, L.; Li, X.; Shen, L.; Fu, Z.; Zhou, X.; and Jiang, X. 2025. FreshRetailNet-50K: A StockoutAnnotated Censored Demand Dataset for Latent Demand Recovery and Forecasting in Fresh Retail. arXiv preprint arXiv:2505.16319. Wasi, A. T.; Islam, M. S.; and Akib, A. R. 2024. SupplyGraph: A Benchmark Dataset for Supply Chain Planning Using Graph Neural Networks. ArXiv, abs/2401.15299. Younis, H.; Sundarakani, B.; and Alsharairi, M. 2022. Applications of Artificial Intelligence and Machine Learning
within Supply Chains: Systematic Review and Future Research Directions. Journal of Modelling in Management, 17(3): 916–940. Zhou, J.; Cao, Z.; Wu, Y.; Song, W.; Ma, Y.; Zhang, J.; and Chi, X. 2024. MVMoE: Multi-Task Vehicle Routing Solver with Mixture-of-Experts. In International Conference on Machine Learning.