Source-linked AI summary
Adapting to Evolving Requirements: Agentic AI for Retail Supply Chain Operations
Lei Zheng, Liping Yang, Zihao Li, Guodong Lyu, Chaik Ming Koh, Chung-Piaw Teo
TL;DR
Retail supply-chain requirements must be translated into changes across coupled decision modules, but existing LLM adaptation methods largely target individual optimization models. The paper proposes graph-constrained agentic intervention search with validated domain-agent patches and downstream KPI selection, raising end-to-end success from 72–76% to 79–83% relative to direct LLM reformulation across three base LLMs.
Problem
Adapting coupled, heterogeneous decision pipelines is difficult because one requirement can map to multiple intervention paths with different downstream effects.
Method
A graph-constrained agentic framework lets a central processor search bounded intervention routes while domain agents expose admissible interfaces, validate patches, and support downstream execution.
Results
79–83% end-to-end success replaces 72–76% under direct LLM reformulation across GPT, Qwen, and DeepSeek, with correctness and end-to-end success improved for all three models.
Takeaways & Limitations
Evaluating multiple admissible reformulations after downstream execution distinguishes locally correct patches from system-level effectiveness in modular supply-chain adaptation.
Abstract
from arXiv · showhide
Retail supply chain operations rely on coupled decision modules that must adapt as requirements evolve. LLMs offer a natural-language interface for this task, but existing methods primarily focus on individual optimization models. Extending them to heterogeneous decision pipelines is challenging because a requirement may admit multiple intervention paths with different downstream effects. We formulate requirement-driven adaptation as the joint selection of an intervention route and an admissible module-level change, and propose a graph-constrained agentic framework in which domain agents expose admissible reformulation interfaces and a central processor searches over bounded intervention paths. Candidates are validated and compared using downstream KPIs. In collaboration with a large retail partner, we evaluate 100 warehouse requirements elicited from practitioner interviews, with GPT, Qwen, and DeepSeek as base LLMs. Relative to direct LLM reformulation, our framework improves correctness and end-to-end success across all three models, raising end-to-end success from 72--76% to 79--83%.
1 Introduction
Retail supply chains use coupled, heterogeneous decision modules whose downstream dependencies make post-deployment requirement changes difficult. The paper proposes extending LLM-based adaptation with structured intervention selection and evaluates the approach on practitioner-derived warehouse requirements.
- Motivation: Separate fulfillment and replenishment modules pass outputs downstream, so changes to one operation can affect subsequent handling and shelf-replenishment work.Packing determines the number and composition of load units delivered to stores.
- Motivation: 4–24% profit gains from replenishment adjustments underscore the value of timely adaptation to changing retail requirements.Planners and operators typically know the desired business change but not how to implement it in the underlying decision system.
- Research challenge: A single requirement may support multiple intervention points and locally valid reformulations with different system-level outcomes after downstream re-execution.The reformulation space is large but structured, while requirements usually concern only a subset of modules and interfaces.
- Contribution: The framework jointly selects an intervention route and admissible module-level change, then assesses quality after affected downstream modules are re-executed.This distinguishes module-level validity from system-level effectiveness.
- Contribution: 79–83% end-to-end success replaces 72–76% under direct LLM reformulation across GPT, Qwen, and DeepSeek.The framework improves reformulation correctness and end-to-end success for every evaluated base LLM.
2 Problem Setting and Modular Agent System
The paper models retail supply-chain adaptation as intervention in a dependency graph of heterogeneous decision modules. Domain agents expose bounded, validated interfaces while a central processor selects where and how to intervene.
- System representation: A modular decision system S = (M, ES) represents decision modules M and directed dependencies ES between their outputs and inputs.Each module includes data and state, a mathematical model, a solution procedure, reformulation interfaces, and validation rules.
- Intervention model: An intervention can modify a module’s mathematical model, algorithmic logic, or configuration parameter through an interface in Tm.Solution procedures may be optimization solvers, constructive heuristics, ranking policies, or hybrids.
- Intervention model: Directed dependencies propagate an intervention by triggering re-execution of the changed module and every reachable downstream module.Downstream inputs and feasible decisions may therefore change after an upstream revision.
- Agent architecture: Each domain agent retains module-specific implementation and validation logic, while the central processor selects interventions from predefined admissible interfaces.Admissibility requires interfaces to be fixed before inference, bounded where applicable, and checked by deterministic validation rules.
- Agent architecture: The separation limits the LLM to controlled intervention choices while domain agents perform implementation and feasibility checks.The architecture separates intervention selection from module execution.
3 Graph-Constrained Agentic Reformulation and Candidate Selection
The framework constrains intervention search with a predefined graph, retains structurally distinct routes, generates executable candidates, validates patches, and selects the candidate with the best downstream criterion.
- Hierarchical Intervention Selection: The predefined graph G organizes hierarchical choices among intents, domain agents, coarse modifications, and fine-grained interfaces.Because G is fixed before inference, the processor cannot introduce interventions absent from the graph.
- Hierarchical Intervention Selection: An admissible route ρ links an intent, selected agents, a coarse modification, and a fine-grained modification through compatible graph edges.The processor then assigns a policy or numerical value v, producing candidate r = (ρ, v).
- One-to-Many Candidate Generation: The system retains one preferred intervention path and a small number of structurally distinct alternatives for each requirement.These alternatives form a compact, inspectable requirement-specific subgraph Gq.
- One-to-Many Candidate Generation: Each retained route is paired with numerical strength values or admissible policies to form an executable portfolio that preserves route and value diversity.Numerical interfaces receive low-, medium-, and high-strength values.
- Domain Reformulation and Downstream Selection: Domain agents instantiate and validate patches against admissible interfaces, parameter bounds, hard constraints, and executability before pipeline execution.A failed candidate receives at most one repair attempt with its route and value held fixed; a second failure is discarded.
- Domain Reformulation and Downstream Selection: The selected candidate maximizes the requirement-specific criterion over candidates that execute successfully and satisfy guardrails.Evaluation uses the resulting system-level KPI vector after modified and downstream modules execute.
4 Computational Study
The computational study evaluates graph-constrained adaptation against direct LLM reformulation on practitioner-elicited warehouse requirements using three base LLMs. The framework improves correctness and end-to-end success across all models, while KPI gains are less uniform.
- Experimental Setup: 100 warehouse requirements span 10 categories with 10 instances each, including clear and multiply reformulable cases.The requirements were elicited from practitioner interviews.
- Experimental Setup: The baseline directly generates mathematical, code, and configuration modifications without hierarchical routing or one-to-many candidate generation.Both approaches use the same warehouse instances, system information, interfaces, and downstream operational pipeline.
- Experimental Setup: GPT-5.5, DeepSeek-v4-flash, and Qwen-3.8-flash serve as the three evaluated base LLMs, with CR, ∆KPI, and ESR reported.The study compares the proposed framework with Direct LLM Reformulation.
- Overall Results: 7 percentage points average ESR gain accompanies higher CR and ESR for all three base LLMs under the proposed framework.Mean ∆KPI changes from negative to positive for GPT and DeepSeek, but Qwen’s mean ∆KPI is larger under the baseline.
- Overall Results: A higher correctness rate can coexist with lower end-to-end success because legal, executable modifications may fail to improve the query-relevant downstream KPI.The framework reduces this gap by generating a bounded set of admissible candidates.
5 Conclusion
The framework addresses downstream effects in modular supply-chain adaptation by combining graph-constrained routes, multiple candidates, domain-agent validation, and KPI-based selection.
- The framework retains several intervention routes and selects among executed candidates using downstream KPIs.It restricts the LLM to a predefined graph and delegates reformulation and validation to domain agents.
- 100 practitioner-elicited warehouse requirements show higher Correctness Rate and End-to-End Success Rate than direct LLM reformulation for all three base LLMs.
- The framework can be extended by adding new domain agents and reformulation interfaces.The paper identifies KPI feedback for revising structural routes and numerical leaf decisions as a future direction.
A Framework Overview and Complete Reformulation Workflow
The workflow encapsulates decision modules as domain agents, routes requirements through a predefined reformulation graph, generates and validates bounded alternatives, then executes candidates and evaluates KPIs.
- Framework Overview: Each decision module is encapsulated as a domain agent while retaining its original operational functions and sharing the system environment.
- Framework Overview: The central processor progressively selects an intent, domain agent, coarse modification, and fine-grained interface to construct an activated subgraph.
- Complete Reformulation Workflow: The retained set contains one primary route and at most two alternatives, with duplicate fine-grained interfaces removed before candidate generation.
- Complete Reformulation Workflow: Candidate values include low-, medium-, and high-strength numerical settings or admissible policies, which domain agents reformulate, validate, and repair.
- Complete Reformulation Workflow: The workflow executes each candidate through the affected domain agents and returns the best validated reformulation or leaves the configuration unchanged if no candidate passes validation.
B Bayesian Optimization for Numerical Fine Modifications
Bayesian optimization refines selected numerical modifications under operational guardrails, using executed candidate outcomes and constrained expected improvement to guide black-box search.
- Optimization Setup: Bayesian optimization refines a selected numerical candidate while keeping its intervention route fixed.It is used only for numerical modifications with unspecified values and KPI-evaluable objectives.
- Optimization Problem: The optimization solves for a numerical value within an admissible domain while maximizing normalized KPI improvement subject to operational guardrails.
- Optimization Problem: System-level guardrails prohibit any monitored KPI from degrading by more than 10%.
- Optimization Setup: The search initializes with the current value and low-, medium-, and high-strength candidates, intersecting their local region with the admissible numerical range.
- Search Procedure: Constrained expected improvement selects the next value, whose KPI and guardrail outcomes are added to the evaluated set after system execution.Search stops at a fixed evaluation budget, a preset improvement threshold, or a practitioner-specified target.
C.1 Evaluation Metrics
The evaluation separates implementation correctness from downstream operational benefit by reporting Correctness Rate, normalized KPI improvement, and End-to-End Success Rate.
- Correctness Rate: Correctness Rate measures whether a reformulation is legal, executable, and applied to the intended component.
- Normalized KPI improvement: Normalized KPI improvement compares the requirement-relevant KPI after executing a reformulation with its value under the original warehouse configuration.
- Normalized KPI improvement: A positive normalized KPI improvement indicates improvement relative to the original warehouse configuration.
- End-to-End Success Rate: End-to-End Success Rate requires both a correct reformulation and an improvement in the KPI associated with the managerial requirement.
C.2 Results by requirement Difficulty
The proposed framework is especially beneficial for requirements with multiple plausible fine-grained modifications, where downstream outcomes can differ despite technically correct reformulations.
- Group B: Multiple plausible modifications: Group B shows an 18% ESR improvement, rising from 54% with Direct LLM Reformulation to 72% with the proposed framework.CR also increases from 90% to 98%.
- Group A: Explicit single modifications: For Group A, Direct LLM Reformulation achieves 96% CR and 94% ESR, while the proposed framework achieves 94% CR and 94% ESR.These requirements specify relatively explicit interventions.
- Correctness versus operational effectiveness: The comparison distinguishes reformulation correctness from downstream operational effectiveness after execution.This distinction is particularly visible for ambiguous Group B requirements, where alternative interventions produce different downstream outcomes.
- Framework implication: The framework organizes alternatives through a predefined intervention graph and evaluates executable candidates using realized warehouse KPIs.This provides a mechanism for comparing locally valid reformulations at the system level.
C.3 Case Study: Deadline-Constrained Store Prioritization
The case study adapts packing and dispatching to prioritize store 473 before 10:30 while preserving hard operational logic. The validated reformulation satisfies the deadline with limited changes to system-level KPIs.
- Hierarchical routing: The central processor resolves “the third store” to store 473 and routes the requirement to the prioritize_store_processing intervention.The selected route modifies exposed store-priority interfaces while preserving hard packing and scheduling logic.
- Packing-agent reformulation: The packing agent moves store 473 from third to first in the processing sequence without changing the packing objective, constraints, or heuristic.The original sequence [402, 484, 473, 310, . . .] becomes [473, 402, 484, 310, . . .].
- Dispatching-agent reformulation: The dispatching agent propagates store priority into roll-cage tags, giving store 473 preferential treatment at multiple decision points.The reformulation preserves precedence, pallet-capacity, AMR-availability, and other hard operational restrictions.
- System execution and KPI evaluation: Figure 2 compares store-level completion timelines before and after reformulation against a dashed 10:30 deadline.The complete packing and dispatching pipeline is executed before comparing schedules and system-level KPIs.
- System execution and KPI evaluation: The revised schedule satisfies the 10:30 deadline, while roll-cage count and average pairwise distance remain unchanged.Overall makespan and pallet-switch count increase by no more than 10%, within the acceptable range, so the reformulation is accepted.