From Topology to Trajectory: LLM-Driven World Models For Supply Chain Resilience

Jia Luo

cs.AIApr 13, 2026

From Topology to Trajectory: LLM-Driven World Models For Supply Chain Resilience

Authors: Jia Luo

Organizations: Huazhong University and Science and Technology Wuhan, China

Abstract

Semiconductor supply chains face unprecedented resilience challenges amidst global geopolitical turbulence. Conventional Large Language Model (LLM) planners, when confronting such non-stationary "Policy Black Swan" events, frequently suffer from Decision Paralysis or a severe Grounding Gap due to the absence of physical environmental modeling. This paper introduces ReflectiChain, a cognitive agentic framework tailored for resilient macroeconomic supply chain planning. The core innovation lies in the integration of Latent Trajectory Rehearsal powered by a generative world model, which couples reflection-in-action (System 2 deliberation) with delayed reflection-on-action. Furthermore, we leverage a Retrospective Agentic RL mechanism to enable autonomous policy evolution during the deployment phase (test-time). Evaluations conducted on our high-fidelity benchmark, Semi-Sim, demonstrate that under extreme scenarios such as export bans and material shortages, ReflectiChain achieves a 250% improvement in average step rewards over the strongest LLM baselines. It successfully restores the Operability Ratio (OR) from a deficient 13.3% to over 88.5% while ensuring robust gradient convergence. Ablation studies further underscore that the synergy between physical grounding constraints and double-loop learning is fundamental to bridging the gap between semantic reasoning and physical reality for long-horizon strategic planning.

Figures & tables

Figure 1: The ReflectiChain framework operates as a closed-loop, adaptive system that bridges the “grounding gap” in supply chain decision-making through a dual-stage reflection process: initially, the system synthesizes multimodal inputs to facilitate Reflection-in-Action , where candidate interventions are sampled and filtered through a dual-path latent rehearsal that concurrently optimizes for semantic compliance and physical grounding via the SC-WM’s spatiotemporal dynamics. Following execution and immediate feedback buffering in the high-fidelity environment, the system triggers Reflection-on-Action , utilizing a retrospective hindsight mechanism to resolve temporal credit assignment by re-evaluating historical decisions with future context, which is ultimately operationalized into policy gradients that fundamentally evolve the actor’s underlying reasoning weights via test-time LoRA adaptation.

Metric (Goal)	PPO (Classic RL)	Qwen2.5-7B	InternLM2.5	ReflectiChain (Ours)	Ideal Ref.
Total Cash (CEE $\uparrow$ )	0.00	410,125.00	547,650.00	894,320.00	$>1{,}000\mathrm{K}$
Avg Compliance (RCI $\uparrow$ )	0.00	100.00	100.00	95.40	$>85.00$
Bullwhip Index (BWI $\approx 1$ )	0.000	0.0117	0.0938	1.1500	$1.0\sim 1.5$
Operability Ratio (OR $\uparrow$ )	0.0%	13.3%	26.7%	88.5%	$>80.0\%$
Average Risk Level (ARL $\downarrow$ )	100.00	73.27	68.45	35.20	$<40.00$

Table 1: Metric-level comparison of ReflectiChain against baselines under macroeconomic shocks. PPO completely fails to navigate semantic constraints, while vanilla LLMs suffer from severe decision paralysis (low OR).

Variant	Step Reward	Retro Score	Observed Failure Mode (Why it fails)
Full ReflectiChain	0.184	85.0	N/A (Optimal Pareto frontier achieved).
w/o World Model	0.052	72.5	Grounding Gap: Agent selects actions that sound logically correct but yield negative physical rewards.
w/o Retrospective RL	0.112	70.0	Static Myopia: Agent cannot extract gradients from historical errors; performance oscillates randomly.
w/o Internal Reflection	-0.045	55.0	Greedy Collapse: Agent fails to recognize out-of-distribution cascading risks before execution.

Table 2: Ablation study demonstrating distinct failure modes when core components are removed.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: Semi-Sim flowchart for spatiotemporal risk propagation dynamics in the semiconductor supply chain. The procedure can be interpreted as Algorithm A . Module I: Initialization takes the network topology $G(V,E)$ , dependency weights $w_{ji}$ , initial node risk $R_{i,t}$ , recovery rate $\gamma$ , and activation threshold $\tau$ to establish the operational baseline. Module II: Endogenous Attenuation computes intrinsic recovery through $R_{\mathrm{internal}}=(1-\gamma)R_{i,t}$ , capturing internal resilience and inventory-buffer effects. Module III: Exogenous Filtering evaluates upstream shock transmission by $\Delta R_{j\to i}=w_{ji}\cdot\max(0,R_{j,t}-\tau)$ , so only disruptions exceeding the threshold trigger cascading impact. Module IV: Spatial Aggregation sums all activated upstream disturbances as $R_{\mathrm{external}}=\sum_{j\in N_{\mathrm{pred}}(i)}\Delta R_{j\to i}$ to form the localized external pressure on node $i$ . Module V: State Transition updates the next-step risk according to $R_{i,t+1}=R_{\mathrm{internal}}+R_{\mathrm{external}}$ , thereby advancing the global system from time $t$ to $t+1$ and checking whether the network approaches systemic-collapse limits.

Figure 7: Global Correlation Matrix of the Triple Feedback RL System variables.

Explore similar work

ReflectiChain: Epistemic Grounding in LLM-Driven World Models for Supply Chain Resilience
Jun 9, 2026Jia LuoConstrained RL LLM-Guided RL
School of Foreign Languages, Huazhong University of Science and Technology, Wuhan 430074, China.
GenAI-Driven Approach to RISC-V Supply Chain Exploration
May 13, 2026Nenad Petrovic, Andre Schamschurko, Yingjie Xu +1Supply Chain Management
Chair of Robotics, Artificial Intelligence and Real-Time Systems Technical University of Munich, Munich, Germany
Adapting to Evolving Requirements: Agentic AI for Retail Supply Chain Operations
Sep 3, 2026Lei Zheng, Liping Yang, Zihao Li +3LLM Agent Orchestration LLM Decision-Making
School of Business and Management, Hong Kong University of Science and Technology · School of Management, University of Science and Technology of China · Institute of Operations Research and Analytics, National University of Singapore +1

Jun 9, 2026J/K move · Enter open · S save

ReflectiChain: Epistemic Grounding in LLM-Driven World Models for Supply Chain Resilience

Jia Luo

School of Foreign Languages, Huazhong University of Science and Technology, Wuhan 430074, China.

Constrained RL LLM-Guided RL Latent World Models Supply Chain Management Robust RL

Open paper AskPDF arXiv

AI agents in supply chains face a fundamental epistemic gap: large language models (LLMs) interpret policies but lack physical grounding, while reinforcement learning (RL) optimizes flows but is semantically blind to unstructured constraints. We introduce REFLECTICHAIN, bridging this gap through a Generative Supply Chain World Model (SC-WM) - encoding heterogeneous supply networks into a 6-dim graph-latent space with physical conservation - and Double-Loop Learning that separates epistemic uncertainty (KL-trust-region-bounded policy adaptation) from aleatoric uncertainty (stochastic latent rollouts). On Semi-Sim, a 10-node semiconductor benchmark with SIR risk propagation, 6 perturbation types, and 10 policy constraint templates, REFLECTICHAIN improves Rationale Consistency Score by 33.0% (p < 0.0001, d = 2.78), maintains 82.3% operability under adversarial shocks, and exhibits anti-fragile behavior (+40.2% gain under moderate pressure). We identify three operational epistemic mechanisms - uncertainty separation, knowledge-boundary detection, and empirical Bayesian policy updating - and discuss five limitation categories.

Abstract

Explore similar work

ReflectiChain: Epistemic Grounding in LLM-Driven World Models for Supply Chain Resilience

GenAI-Driven Approach to RISC-V Supply Chain Exploration

Adapting to Evolving Requirements: Agentic AI for Retail Supply Chain Operations