PolicyMediator: Multi-Regional Policy Learning through Mediation Agents with Evolving Skills in OPC Ecosystems

Jie Li; Bin Guo; Georgina Curto; Zikang Liu; Mengqi Chen; yu li

cs.MAOct 26, 2026

PolicyMediator: Multi-Regional Policy Learning through Mediation Agents with Evolving Skills in OPC Ecosystems

Authors: Jie Li, Bin Guo, Georgina Curto, Zikang Liu, Mengqi Chen, yu li

Organizations: Northwestern Polytechnical University, Xi’an, China · United Nations University Institute in Macau, Macau, China

Abstract

One-person companies (OPCs) are emerging enterprises operated by individual entrepreneurs through digital platforms and LLM-based AI agents. These highly skilled entrepreneurs are characterized by being digital nomads, with limited reliance on local ecosystems. Governments in China use distinct regional support policies to attract mobile OPCs, turning local support into a cross-regional coordination problem. OPCs compare regions to maximize their own utility, but imperfect policy perceptions may cause them to reject otherwise suitable locations. Existing policy-learning methods optimize resources under fixed perceptions, while consensus-formation methods aggregate views without feeding cognitive changes back into policy learning. This disconnect produces mismatched location choices and inefficient regional policy adjustment. We introduce PolicyMediator, a cognition-constrained leader-multi-follower Stackelberg framework with bilateral mediation. The mediator explains verified policies to OPCs to build awareness and trust, and reports their needs and unresolved concerns to governments for policy learning. Its self-evolving skills retain, revise, or prune strategies using verification, allocation, capacity, and objective trajectories rather than mediation success alone. Across six-region simulations in China with five OPC types, five scenarios, and three OPC-simulator backbones, PolicyMediator ranks first in all 15 OPC-backbone--scenario combinations. Over the eight-baseline mean, it improves the aggregate government objective, combining economic performance, employment benefit, and policy-access equality, by 42.4% and average OPC utility by 40.9%.

Figures & tables

Figure 1. Comparison of regional policy-coordination paradigms. Policy optimization excludes OPC concerns, while consensus formation lacks policy or cognition updates. PolicyMediator links verified mediation to OPC choices, policy revision, and skill evolution. Traditional policy optimization uses realized payoff feedback but leaves OPC concerns unreturned. Traditional consensus formation exchanges and aggregates views without policy feedback or strategy updates. PolicyMediator connects verified policy explanations and OPC concerns to feasible location choices and revised government policy.

Figure 2. PolicyMediator framework: (a) agent modeling; (b) cognition-constrained Stackelberg interaction with policy revision; and (c) trajectory-guided skill evaluation and next-round policy adjustment. Three-panel framework. The first panel models regional governments and heterogeneous OPCs. The second connects verified regional policies, mediation, cognition updates, location recommendations, and policy revision. The third evaluates original and candidate skills on matched trajectories, retains, revises, or prunes them, and feeds allocation and government outcomes into the next policy round.

Figure 4. Mechanism analyses: (a) cognition changes by OPC transition, (b) feedback utilization across need intensity and decision ambiguity, and (c) paired $\Delta W^{G}$ by feedback type. A cognitive-response embedding, a feedback-utilization heat map, and violin plots of paired $W^G$ gains.

Comparison	Restraint		Regional		Targeting		Specificity		Progression		Average
Comparison	Win	Lose	Win	Lose	Win	Lose	Win	Lose	Win	Lose	Win	Lose
PolicyMediator vs. CONSENSAGENT	56.0	42.0	58.0	41.0	$\mathbf{73.0}^{\ddagger}$	20.0	45.0	51.0	$\mathbf{69.0}^{\ddagger}$	12.0	$\mathbf{60.2}$	33.2
PolicyMediator vs. Contextual Bandit	$\mathbf{87.0}^{\ddagger}$	12.0	$\mathbf{47.0}$	46.0	40.0	53.0	$\mathbf{69.0}^{\ddagger}$	26.0	$\mathbf{41.0}$	36.0	$\mathbf{56.8}$	34.6
PolicyMediator vs. Generic Explainer	$\mathbf{92.0}^{\ddagger}$	4.0	$\mathbf{81.0}^{\ddagger}$	17.0	$\mathbf{100.0}^{\ddagger}$	0.0	$\mathbf{91.0}^{\ddagger}$	7.0	$\mathbf{100.0}^{\ddagger}$	0.0	$\mathbf{92.8}$	5.6

Table 2. Blinded pairwise interactive evaluation (%; 100 pairs/comparison;

\ddagger

: exact binomial

p<.01

Figure 5. Skill evolution: (a) decision quality and (b) first-pass verification rate under theory-guided and generic evolution. A single-column, two-panel grouped-bar figure comparing theory-guided and generic skill evolution over five experience checkpoints.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Simulated OPC Dialogue: Policy Explanation and Trust Building
Policy context	OPC type: creator. Policy package: 34% price subsidy, 58% computing support, and a 12-month rent-free period.	Baseline $z=.45$ ; $b=.42$ ; $q_{i,\mathrm{Shenzhen}}=.475$ .
Turn 0	Mediator—concern elicitation: “Which step worries you most: reimbursement, approval, or something else? We will answer from verified policy facts.” OPC: “Is computing support actually usable or only a paper promise? Show me a signed quota or access agreement.”	$z=.40$ ; $b=.32$ ; $q_{i,\mathrm{Shenzhen}}=.458$ . Delivery uncertainty emerges; trust and regional relevance decline.
Turn 1	Mediator—verified evidence: “The Commission publishes redemption records for computing vouchers and rent-free space. Vouchers cover actual usage, while guaranteed loans require guarantor approval.” OPC: “The records reduce voucher risk, but I still need the local approval rate for the guaranteed loan.”	$z=.48$ ; $b=.37$ ; $q_{i,\mathrm{Shenzhen}}=.480$ . Voucher evidence reduces delivery concerns, while loan approval remains uncertain.
Turn 2	Mediator—procedural clarification: “Voucher and space support are documented, but the loan’s no-penalty clause remains unverified. You may apply for verified support while seeking written loan terms.” OPC: “I will start those two applications but defer the loan until I verify the no-penalty clause.”	$z=.56$ ; $b=.42$ ; $q_{i,\mathrm{Shenzhen}}=.495$ . Verified information restores trust despite remaining loan uncertainty.
Observed outcome	Final stance: Conditional acceptance ; confidence $=75\%$ ; mediation success: True .	Net update: $\Delta z=+.11$ , $\Delta b=.00$ , and $\Delta q_{i,\mathrm{Shenzhen}}=+.020$ . Each per-turn update satisfies Eq. ( 2 ).
Bilateral Mediator Feedback Trace

Appendix

Table 3. Representative bilateral mediation case for a creator OPC. Here

q_{ir}

is the region-level relevance, i.e., the mean of six policy-category perceptions for region

r

Setting	Value / range
OPCs per run	$100$
Type proportions (S1–S4)	$(.25,.25,.20,.20,.10)$
Type proportions (S5)	$(.40,.10,.10,.30,.10)$
Entry-cost multiplier / $\rho$ (S1–S5)	$(1,1.8,1,.5,1)/(1,1,.5,1,1)$
Outside-option coefficient / reporting scale $\zeta_{W}$	$.20/100$
Cycles / actor step / perturbation radius	$12/.30/.02$

Appendix

Table 4. Simulation and policy-learning hyperparameters.

Policy Learning

Policy learning develops decision rules that choose actions or interventions to maximize outcomes from data, feedback, or interaction.

See all 487 papers

Stackelberg POMDP: Learning to Lead via Reinforcement Learning
Oct 7, 2022Matthias Gerstgrasser, Gianluca Brero, Alon Eden +6Game-Theoretic RL Reinforcement Learning
John A. Paulson School of Engineering and Applied Sciences, Harvard University · Data Science Initiative, Brown University · School of Computer Science and Engineering, Hebrew University of Jerusalem +2
Submodular Policy Learning for Distributed Task Allocation in Open Multi-Agent Systems
Aug 14, 2026Jing Liu, Luca Ballotta, Yangyang Yang +3Multi-Agent Systems Policy Learning
School of Mathematics, East China University of Science and Technology, Shanghai 200237, China · Department of Information Engineering, University of Padova, 35131 Padova, Italy · School of Mathematics and the Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai 200237, China
PolicyBank: Evolving Policy Understanding for LLM Agents
Apr 16, 2026Jihye Choi, Jinsung Yoon, Long T. Le +2AI Agent Benchmarks AI Agent Governance
Google Cloud · University of Wisconsin-Madison

Oct 7, 2022J/K move · Enter open · S save

Stackelberg POMDP: Learning to Lead via Reinforcement Learning

Matthias Gerstgrasser, Gianluca Brero, Alon Eden, Darshan Chakrabarti+5

John A. Paulson School of Engineering and Applied Sciences, Harvard University · Data Science Initiative, Brown University · School of Computer Science and Engineering, Hebrew University of Jerusalem+2

Game-Theoretic RL Reinforcement Learning Multi-Agent Reinforcement Learning Partially Observable Markov Decision Processes Mechanism Design

Open paper AskPDF arXiv

Many real-world domains--including e-commerce platform design, security planning, and multi-agent coordination--feature leader-follower problems where one decision-maker commits to a policy and others react strategically. We develop a reinforcement learning framework for such interactions in sequential environments with partial observations and multiple followers. Followers may adapt through no-regret learning or reinforcement learning, potentially departing from equilibrium behavior. The framework embeds follower adaptation into the leader's environment to construct a single-agent partially observable Markov decision process--the Stackelberg POMDP. For policy-interactive response algorithms, which access the leader's policy through queries, we prove that an optimal policy based only on the leader's game history yields an optimal commitment under the specified response procedure. We use proximal policy optimization with a centralized critic and train contextual meta-followers to respond across leader policies. In indirect mechanism design, mechanisms using buyer messages achieve higher social welfare than optimal standard sequential price mechanisms across all tested type counts, with responses certified as approximate Bayesian coarse correlated equilibria. In platform design, learned display rules increase mean consumer surplus by 8.4% over an optimized fixed price cap while accommodating hidden seller costs. In Atari bilateral trade, meta-learned follower responses support joint learning of visual gameplay and economic decisions; assigning leadership to the seller or buyer shifts transaction prices and payoffs in that agent's favor. Controlled ablations examine how response credit, policy consistency, and reward timing affect learning.

Abstract

Policy Learning

Explore similar work

Stackelberg POMDP: Learning to Lead via Reinforcement Learning

Submodular Policy Learning for Distributed Task Allocation in Open Multi-Agent Systems

PolicyBank: Evolving Policy Understanding for LLM Agents