SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination

Max Z.J. Shen; Ningxuan Kang; Pujun Zhang; Xianqi Cao; Yongzhi Qi; Yuan Qu; Yunhao Liang

REVIEW 4 major objections 5 minor 12 references

SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination

T0 review · 4 major / 5 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Learning the couplings among assortment, replenishment frequency, sourcing, and routing beats optimizing each supply-chain department separately.

desk verdict Solid applied pipeline paper: real two-echelon data and a fair baseline grid, but the wins live on a hand-built proxy utility and the Dingdong margin is tiny. read the letter →

arxiv 2607.28488 v1 pith:EXVI4VBY submitted 2026-07-30 cs.AI cs.LG

Yunhao Liang , Xianqi Cao , Pujun Zhang , Yuan Qu , Yongzhi Qi , Ningxuan Kang , Max Z.J. Shen This is my paper

classification cs.AIcs.LG

keywords supplychainoperationscoupledpoliciesend-to-endcoordinationassortmentplanningreplenishmentvehicleroutinglatentoperationalcouplingfreshretail

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Modern supply chains split assortment, replenishment timing, source assignment, and vehicle routing across departments, each optimizing its own problem. Those choices are coupled: what you stock and how often you restock reshape the loads and origins logistics must serve, and routing cost and feasibility in turn decide whether the upstream plan was worth it. This paper argues that the central AI task is to learn those latent operational couplings and coordinate them under one system-level utility, rather than to improve any single module in isolation. It introduces SCOPE, a composite policy that encodes facilities, products, demand, and cost context in a shared representation and extends a partial plan stage by stage until a complete replenishment plan is scored. On real Dingdong store-level and JD.com warehouse-level data, the coupled policy consistently beats pipelines that pair industrial assortment rules and fixed cycles with strong classical and neural routers, showing that end-to-end coordination can reduce the stockouts, inventory exposure, and wasted transport that siloed optimizers leave on the table.

What carries the argument

SCOPE: a composite conditional policy that tokenizes supply-chain entities, contextualizes them in a shared operational encoder, and maps each token type to a sequential decision interface (assortment, source, cycle, then routing), with the finished plan scored by one shared daily proxy utility.

What would settle it

Re-run the same Dingdong and JD.com complete-pipeline comparison using realized margin, executed freight cost, and true lost sales instead of the proxy utility; if SCOPE no longer leads the best decomposed industrial pipelines, the central claim fails.

Watch

Extended reading notes

Core claim

On real Dingdong depot-to-store and JD.com RDC-to-FDC operational data, a single composite policy that learns latent operational couplings among assortment, source assignment, replenishment cycle, and routing achieves higher complete-pipeline system utility than decomposed pipelines that fix practice-style upstream rules and then apply strong route solvers. The gains persist across two replenishment echelons and are driven more by coordinating upstream decisions than by swapping one router for another.

Load-bearing premise

The hand-built daily proxy that scores assortment value, uncovered demand, transport cost, and inventory exposure is faithful enough to real system value that winning on that proxy means better real replenishment decisions.

Editorial extensions

If this is right

Jointly learned assortment–cycle–routing plans can outperform prediction-then-rank assortment plus fixed cycles plus strong VRP solvers at both store and warehouse echelons.
Once upstream decisions have formed the routing instance, a stronger router alone cannot recover value lost to poor assortment or cycle choices.
The same backbone-plus-interface design can absorb a multi-depot topology change by adding a source-assignment interface rather than redesigning the whole system.
Staged surrogates plus validation on full-pipeline utility can train coupled supply-chain policies when ideal completion values are unobserved and costly to compute.
Planners can inspect how assortment and service-frequency choices propagate into capacity pressure and routing burden before costs materialize.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same staged token-and-interface pattern could extend to other multi-department pipelines where early choices induce later combinatorial instances, such as integrated production–distribution or multi-echelon allocation.
If proxy calibration is the load-bearing hinge, organizations with better ground-truth margin and lost-sales data may see larger or smaller gains than the paper reports, making proxy fidelity an operational A/B test rather than a modeling detail.
Direct joint training’s sharp failure on one echelon and near-parity on the other suggests that curriculum or staged supervision may be as important as architecture when couplings are sparse or noisy.
Cross-department visibility tools built on the shared representation could matter as much for human planners as the automated plan itself.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Punchline: this is a competent industrial decision-learning paper, not a theory breakthrough. What it actually delivers is a shared encoder plus sequential assortment/source/cycle/routing heads, trained with stage surrogates and selected on full-pipeline proxy utility, then stress-tested on Dingdong FreshRetailNet and a JD RDC–FDC network against a wide grid of practice-style assortments, fixed cycles, and strong routers.

What is new is less the coupling idea—OR has done assortment–inventory–routing for years—and more the concrete learned backbone-plus-interfaces setup that treats routing instances as outputs of upstream choices, plus the full-pipeline bake-off on two real echelons. That experimental design is the real contribution. Table 1 and the ablations are useful: cycle choice and joint coordination move the needle more than swapping OR-Tools for HGS or POMO once the upstream plan is fixed, and the JD gap is more meaningful than Dingdong’s.

Soft spots, in proportion. The load-bearing assumption is Eq. 2’s proxy U (value, uncovered demand, route cost, inventory exposure from literature coefficients), not realized margin or dynamic lost sales. Checkpoint selection maximizes the same U. On Dingdong the best-baseline gap is ~54 on a ~206k base with identical coverage—directionally consistent, not operationally decisive. Robustness already flips ranks when assortment value is scaled down hard, which the authors report honestly. Training is staged imitation and counterfactuals, not optimization of the ideal completion values Q. JD is NDA-rescaled, so external replication stops at Dingdong. None of that sinks the paper; it bounds the claim to “better on this shared planning proxy under coupled policies.”

Math and citations look fine: formulation is clear, related work covers classical IRP, assortment, neural VRP, and SC ML without obvious gaps. No code release is a practical drag.

Who it’s for: people building multi-stage retail/fulfillment ML, and OR folks who want a modern baseline for end-to-end proxy pipelines. I’d bring it to a reading group if the room cares about SC operations ML. It deserves a serious referee, not a desk reject—ask for uncertainty, Dingdong reproduction artifacts, and clearer limits on proxy-to-P&L inference. Engage if that is your lane; skim the tables if it isn’t.

Referee Report

4 major / 5 minor

Summary. The paper proposes SCOPE, a composite policy that encodes supply-chain entities as typed tokens in a shared operational backbone and maps them to sequential decision interfaces for assortment, source assignment, replenishment cycle, and routing. The central claim is that latent operational couplings among these stages can be learned and coordinated under a shared system-level utility, yielding better complete replenishment plans than department-wise optimization. The framework is instantiated on urban fresh-retail replenishment and evaluated as a full assortment–cycle–routing pipeline on Dingdong depot–store data (FreshRetailNet-50K) and a JD.com RDC–FDC warehouse network. Against grids of practice-style assortment rules, globally fixed cycles, and strong classical/neural routers, SCOPE reports higher reference-adjusted proxy utility U_ref on both echelons (Table 1), with supporting ablations (Table 2) and proxy-scale robustness sweeps (§4.5).

Significance. If the result holds under a credible system objective, the contribution is meaningful for both AI and operations: it reframes multi-department replenishment as coupled decision learning rather than isolated forecasting or routing, and it demonstrates structural reuse of one backbone-plus-interface design across two real echelons (public Dingdong and industrial JD.com). Strengths include complete-pipeline evaluation against a broad baseline grid (OR-Tools, HGS, AM, POMO, MVMoE, UniteFormer), matched decision/training ablations, and explicit robustness checks over cost and load scales. The dual-echelon industrial setting and the inspectable staged plan structure are genuine assets relative to route-only neural VRP work and single-layer inventory RL. The significance is currently limited by reliance on a hand-specified proxy utility and by small absolute gains on the public benchmark.

major comments (4)

[§3.1 Eq. (2); Table 1; §4.5] §3.1 Eq. (2) and §4.2–4.3: The claim that SCOPE yields “more effective end-to-end supply-chain decisions” (abstract; §5) is supported only by outperformance on the hand-specified daily proxy U = V(A) − L_cov(A) − C_rt(R;c) − C_inv(A,T;c), built from literature cost coefficients and planning load/value proxies rather than realized margin, executed freight, or dynamic lost sales. Checkpoint selection itself maximizes validation U (Eq. 8). On Dingdong the gap vs TopDemand–OR-Tools is only ~54 on a ~206k base (~0.026%) at identical coverage 0.948 (Table 1); §4.5 already shows rank reversals when the assortment-value scale is cut to 0.25×. Either validate U against hold-out operational KPIs (margin, service level, realized transport) or substantially qualify the effectiveness claim to “higher proxy pipeline utility.”
[§3.1 Eqs. (3)–(5); §3.3 Eq. (8)] §3.1 Eqs. (3)–(5) vs §3.3: Operational completion values Q_{s,c} are defined as the conceptual target but are never computed. Policies are fit with stage surrogates (assortment swaps, geometric source warm-starts, global interval labels, route imitation from OR-Tools) and composites are chosen by the same validation U used at test time (Eq. 8). This is a reasonable engineering protocol, but it does not establish that SCOPE approximates the joint policy objective in Eq. (5) or “learns” the Q-coupling in Eq. (3). The paper should either (i) report a limited Q-estimation or counterfactual completion study on a subset of scenarios, or (ii) restate the contribution as staged surrogate coordination selected by pipeline U, not as optimization of Eq. (5).
[Table 1; Table 2; §3.2; §4.2] Table 1 and §4.2: Cycle baselines use one globally fixed BestFix T ∈ {1..7} (or Daily T=1) for all units, whereas SCOPE predicts a separate interval per replenishment unit M_z (§3.2). Part of the reported gain may therefore come from per-unit cycle flexibility rather than from learned cross-stage coupling. Table 2’s cycle ablation replaces the learned interface with daily T=1, which is a weak control, not BestFix. Please add a matched baseline (and ablation) that allows per-unit or per-store cycles chosen by the same global scan/oracle logic used for BestFix, holding assortment and router fixed, so the coupling claim is isolated from adaptive-cycle expressivity.
[Table 1; §4.3] Table 1 Dingdong column: With coverage tied at 0.948 and absolute U_ref gaps of 53–150 vs strong TopDemand+BestFix+router pipelines, practical and statistical significance are unclear (18 scenarios, no standard errors, no paired tests). Please report scenario-level paired differences with uncertainty, and discuss whether ~0.03% proxy uplift is operationally material under the stated cold-chain cost model. The larger JD gap is more convincing on magnitude but remains NDA-rescaled and proxy-based.

minor comments (5)

[Figure 2; §3.3] Figure 2 caption and body: “staged training: assortment supervision · cycle counterfactuals · router” is helpful but the main text (§3.3) should cross-reference the figure’s training stages more explicitly so the reader can map losses to interfaces without the supplement.
[§3.1 Eq. (2)] Notation: z_c = e_θ(c) and E = B_θ(s) are introduced cleanly, but V, L_cov, and C_inv in Eq. (2) are only named; even brief functional forms (or a pointer to exact appendix formulae) in the main text would aid reproducibility of U_ref.
[Table 2] Table 2: “Assignment remains part of the learned JD.com pipeline but is not treated as a separate ablation.” An assignment-interface ablation on JD would strengthen the multi-echelon generality claim; if infeasible under NDA, say so once in the table note.
[§2 Positioning] Related work: Integrated assortment–inventory and inventory-routing citations are appropriate; a clearer sentence on how SCOPE differs from learnable IRP / end-to-end replenishment (Qi et al. 2020) in the positioning paragraph would help.
[Title block; §4.3; §5] Typos/formatting: “Y ongzhi Qi” (affiliation line); “T wo Replenishment” (§4.3 heading); “interfacese” (§5); occasional missing spaces after periods in the compiled text (e.g., “modulestoward”, “Acompletereplen-”).

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circular derivation; standard val-U checkpoint selection on a hand-specified proxy is a validity issue, not Eq.-by-construction circularity.

fitted input called prediction [§3.3 Full pipeline selection, Eq. (8); cf. Eqs. (3)–(5)]
"We compare candidate checkpoints C by running the complete policy on the validation set: bθ∈arg max_θ∈C (1/|Dval|) Σ U(s,c; Πθ(s,c)). ... SCOPE coordinates the conditional policies using complete-pipeline evaluations without explicitly estimating Qs,c in Eq. (3)."

Checkpoint selection maximizes the same proxy family U later reported as U_ref. This couples model choice to the test metric and means reported wins are partly ‘best-of-C on U,’ not estimates of uncomputed Q. It is standard val-set selection, not reduction of a claimed first-principles prediction to its fitted input; retained only as a minor, non-central circularity-adjacent step.

full rationale

SCOPE’s central claim is empirical: a composite policy with shared representation and staged interfaces beats decomposed assortment–cycle–routing pipelines on complete-pipeline proxy utility U_ref (Table 1) on Dingdong and JD.com data. The formal chain (Eqs. 2–5) defines a daily proxy U and ideal completion values Q, then explicitly states that Q is never computed; stage surrogates (assortment swaps, geometric source warm-starts, global interval labels, OR-Tools route imitation) train heads, and composite checkpoints are chosen by validation pipeline U (Eq. 8). That is ordinary ML model selection on the reported metric, not a self-definitional loop (U is not defined from SCOPE’s outputs) and not a fitted scalar renamed as an out-of-sample prediction of the same quantity. Baselines are also scored on the same U, and several (BestFix cycle scan, strong routers) are competitive on that surface, so the ranking is a genuine comparison under a fixed proxy rather than a tautology. Author-overlapping citations (e.g., Shen et al. 2025; Qi et al. 2020) supply industrial context and related inventory work; they do not supply a uniqueness theorem or ansatz that forces the Table 1 result. Concerns that hand-built V, L_cov, and C_inv may not track realized P&L or stockouts are external-validity / proxy-fidelity issues, outside circularity as defined here. Score 1 only for the mild, non-load-bearing coincidence of selection metric and report metric.

Assumptions & free parameters 5 free parameters · 6 assumptions · 2 invented entities

The central empirical claim rests on accepting a proxy system utility, staged surrogate training as a stand-in for joint policy optimization, fixed industrial cost/load conventions, and the representativeness of two fresh-retail datasets (one public, one NDA). No new physical entity is postulated; the main invented conceptual object is ‘latent operational coupling’ plus the SCOPE token/interface stack. Free parameters are mostly cost scales, assortment cardinality, interval caps, and training/architecture choices that shape U and the learned policies.

free parameters (5)

Assortment cardinality K = Fixed per protocol (not numerically specified in main text)
Each location must select exactly K SKUs; K shapes demand, load, and coverage and is treated as given rather than optimized jointly with U.
Vehicle/dispatch and holding cost coefficients (c_veh, α_hold, related scales) = Literature-derived; robustness over 0.25×–multiple scales
Enter C_rt and C_inv in Eq. 2; taken from reported Chinese cold-chain costs and textbook carrying-cost conventions, then stress-tested by rescaling.
Global fixed-cycle scan range T ∈ {1..7} and T_max = T up to 7 in main tables
Defines feasible interval set and BestFix baseline labels; also structures cycle supervision warm starts.
Proxy weights inside assortment targets (coverage, value, load, inventory exposure)
Upstream assortment supervision combines these proxies before swap search on full-pipeline U; relative weights are design choices.
Neural architecture and training hyperparameters (encoder, adapters, multi-start decode, losses)
Determine capacity of shared representation E and each interface; checkpoint picked by validation U among candidates C.

assumptions (6)

domain assumption Daily proxy utility U in Eq. 2 is an adequate ranking objective for comparing complete replenishment plans.
Stated in §3.1; L_cov is explicitly not realized lost sales; horizon totals are normalized to daily proxies.
domain assumption Upstream decisions induce downstream routing instances through known operational rules (load aggregation, service days, origin assignment).
Core of the problem formulation in §3.1 and routing construction in §3.2.
ad hoc to paper Staged surrogates plus validation pipeline selection sufficiently approximate the policy objective in Eq. 5.
§3.3 admits Q is not estimated and gradients do not flow from U into heads; Eq. 8 only selects among checkpoints.
domain assumption Top-K demand/value assortment rules and BestFix global cycles are fair stand-ins for strong departmental practice.
§4.2 links Top-K to JD.com prediction-then-ranking practice; BestFix is a global scan, not a per-episode oracle.
ad hoc to paper Transformer-style shared encoding of typed supply-chain tokens is an appropriate hypothesis class for operational context.
Model overview §3.2; related to multi-task representation learning but not derived from a uniqueness theorem.
standard math Standard math/OR feasibility constraints for CVRP-like subproblems (capacity masks, distance matrices, eligible links).
Used when defining R(s;A,σ,T) and decoder masks in §3.1–3.2.

invented entities (2)

Latent operational coupling
purpose: Name the unobserved cross-stage dependence by which assortment, source, interval, and routing reshape each other’s attainable system value.
Introduced as a modeling language in the introduction and contributions; operationalized via conditional policies and shared U, not measured as an independent latent variable with external instruments.
SCOPE composite policy (shared operational backbone + typed decision interfaces)
purpose: Map tokenized entities to sequential assortment, source, cycle, and routing decisions under one utility.
Primary proposed method; evidence is internal empirical comparison, not an external physical detection.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination." pith.science (2026). https://pith.science/paper/EXVI4VBY

@misc{pith2026260728488,
  author       = {Pith},
  title        = {Pith review of: SCOPE: Supply-Chain Operations through Coupled Policies for End-to-End Coordination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EXVI4VBY}},
  note         = {Machine review of arXiv:2607.28488}
}

read the original 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 system-level 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.

Figures

Figures reproduced from arXiv: 2607.28488 by the authors.

**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 … view at source ↗

**Figure 3.** FreshRetailNet-50K exhibits the heterogeneity that [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 5 linked inside Pith

[3]

Boute, R

On the Opportunities and Risks of Foundation Models.arXiv preprint arXiv:2108.07258. Boute, R. N.; Gijsbrechts, J.; van Jaarsveld, W.; and Van- vuchelen,N.2022. DeepReinforcementLearningforInven- toryControl:ARoadmap.EuropeanJournalofOperational Research, 298(2): 401–412. Brintrup, A.; Pak, J.; Ratiney, D.; Pearce, T.; Wichmann, P.; Woodall, P.; and McFarlane, D

arXiv 2022
[10]

100-Day Reviews under Executive Order 14017

Building Resilient Supply Chains, RevitalizingAmericanManufacturing,andFosteringBroad- Based Growth. 100-Day Reviews under Executive Order 14017. Toth,P.;andVigo,D.,eds.2014.VehicleRouting:Problems, Methods, and Applications. SIAM. van Donselaar, K.; van Woensel, T.; Broekmeulen, R.; and Fransoo,J.2006. InventoryControlofPerishablesinSuper- markets.Intern...

2014
[11]

FreshRetailNet-50K: A Stockout- Annotated 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

arXiv
[1997]

Management Science, 43(4): 405–570

Infor- mation Distortion in a Supply Chain: The Bullwhip Effect. Management Science, 43(4): 405–570. Martínez-deAlbéniz,V.;andKunnumkal,S.2022. AModel forIntegratedInventoryandAssortmentPlanning.Manage- ment Science, 68(7): 5049–5067. Meng,D.;Cao,Z.;Gao,J.;Wu,Y.;andHou,Y.2025. Unite- Former:UnifyingNodeandEdgeModalitiesinTransformers for Vehicle Routing P...

2022
[2007]

Shen,Z.-J.M.;Coullard,C.;andDaskin,M.S.2003

Integrated Supply Chain Design Mod- els: A Survey and Future Research Directions.Journal of Industrial and Management Optimization, 3(1): 1–27. Shen,Z.-J.M.;Coullard,C.;andDaskin,M.S.2003. AJoint 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

2003
[2014]

Dantzig,G.B.;andRamser,J.H.1959

Thirty Years of Inventory Routing.Transportation Science, 48(1): 1–19. Dantzig,G.B.;andRamser,J.H.1959. TheTruckDispatch- ing Problem.Management Science, 6(1): 80–91. Feng, Q.; Hua, W.; Chen, C.; Shi, X.; and Ma, J

1959
[2017]

Sajadi, S

An Overview of Multi-Task Learning in Deep Neural Networks.arXiv preprint arXiv:1706.05098. Sajadi, S. J.; and Ahmadi, A

arXiv
[2019]

Baumgartner,T.;Malik,Y.;andPadhi,A.2020.Reimagining Industrial Supply Chains

Predicting SupplyChainRisksUsingMachineLearning:TheTrade-Off between Performance and Interpretability.Future Genera- tion Computer Systems, 101: 993–1004. Baumgartner,T.;Malik,Y.;andPadhi,A.2020.Reimagining Industrial Supply Chains. McKinsey & Company. Bommasani, R.; Hudson, D. A.; Adeli, E.; et al

2020

Show all 12 references

[2021]

Baryannis, G.; Dani, S.; and Antoniou, G

Data Considerations in Graph Representation Learning for Supply Chain Networks.arXiv preprint arXiv:2107.10609. Baryannis, G.; Dani, S.; and Antoniou, G

arXiv
[2022]

Pan,S.;Liao,H.;Zheng,G.;Huang,Q.;andShan,M.2024

A Deep Q-Network for the Beer Game: Deep Rein- forcement Learning for Inventory Optimization.Manufac- turing & Service Operations Management, 24(1): 285–304. Pan,S.;Liao,H.;Zheng,G.;Huang,Q.;andShan,M.2024. ColdChainDistributionRouteOptimizationforMixedVehi- cleTypesofFreshAgr...

2024
[2024]

Younis, H.; Sundarakani, B.; and Alsharairi, M

Supply- Graph: A Benchmark Dataset for Supply Chain Planning Using Graph Neural Networks.ArXiv, abs/2401.15299. Younis, H.; Sundarakani, B.; and Alsharairi, M

arXiv
[2025]

Silver, E

JD.com Improves Fulfillment EfficiencywithData-DrivenIntegratedAssortmentPlanning and Inventory Allocation.INFORMS Journal on Applied Analytics, 55(5): 386–398. Silver, E. A.; Pyke, D. F.; and Peterson, R. 1998.Inventory ManagementandProductionPlanningandScheduling.John Wiley ...

1998

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.

[1] [3]

Boute, R

On the Opportunities and Risks of Foundation Models.arXiv preprint arXiv:2108.07258. Boute, R. N.; Gijsbrechts, J.; van Jaarsveld, W.; and Van- vuchelen,N.2022. DeepReinforcementLearningforInven- toryControl:ARoadmap.EuropeanJournalofOperational Research, 298(2): 401–412. Brintrup, A.; Pak, J.; Ratiney, D.; Pearce, T.; Wichmann, P.; Woodall, P.; and McFarlane, D

arXiv 2022

[2] [10]

100-Day Reviews under Executive Order 14017

Building Resilient Supply Chains, RevitalizingAmericanManufacturing,andFosteringBroad- Based Growth. 100-Day Reviews under Executive Order 14017. Toth,P.;andVigo,D.,eds.2014.VehicleRouting:Problems, Methods, and Applications. SIAM. van Donselaar, K.; van Woensel, T.; Broekmeulen, R.; and Fransoo,J.2006. InventoryControlofPerishablesinSuper- markets.Intern...

2014

[3] [11]

FreshRetailNet-50K: A Stockout- Annotated 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

arXiv

[4] [1997]

Management Science, 43(4): 405–570

Infor- mation Distortion in a Supply Chain: The Bullwhip Effect. Management Science, 43(4): 405–570. Martínez-deAlbéniz,V.;andKunnumkal,S.2022. AModel forIntegratedInventoryandAssortmentPlanning.Manage- ment Science, 68(7): 5049–5067. Meng,D.;Cao,Z.;Gao,J.;Wu,Y.;andHou,Y.2025. Unite- Former:UnifyingNodeandEdgeModalitiesinTransformers for Vehicle Routing P...

2022

[5] [2007]

Shen,Z.-J.M.;Coullard,C.;andDaskin,M.S.2003

Integrated Supply Chain Design Mod- els: A Survey and Future Research Directions.Journal of Industrial and Management Optimization, 3(1): 1–27. Shen,Z.-J.M.;Coullard,C.;andDaskin,M.S.2003. AJoint 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

2003

[6] [2014]

Dantzig,G.B.;andRamser,J.H.1959

Thirty Years of Inventory Routing.Transportation Science, 48(1): 1–19. Dantzig,G.B.;andRamser,J.H.1959. TheTruckDispatch- ing Problem.Management Science, 6(1): 80–91. Feng, Q.; Hua, W.; Chen, C.; Shi, X.; and Ma, J

1959

[7] [2017]

Sajadi, S

An Overview of Multi-Task Learning in Deep Neural Networks.arXiv preprint arXiv:1706.05098. Sajadi, S. J.; and Ahmadi, A

arXiv

[8] [2019]

Baumgartner,T.;Malik,Y.;andPadhi,A.2020.Reimagining Industrial Supply Chains

Predicting SupplyChainRisksUsingMachineLearning:TheTrade-Off between Performance and Interpretability.Future Genera- tion Computer Systems, 101: 993–1004. Baumgartner,T.;Malik,Y.;andPadhi,A.2020.Reimagining Industrial Supply Chains. McKinsey & Company. Bommasani, R.; Hudson, D. A.; Adeli, E.; et al

2020