PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.35997v1cs.MAcs.AIcs.CYcs.HC

TL;DR

Incomplete, trust-dependent policy perceptions can lead OPCs to reject suitable regions, while existing policy-learning and consensus methods do not connect cognitive change to policy updates. PolicyMediator combines cognition-constrained Stackelberg learning, verified bilateral mediation, and outcome-aware skill evolution. Across 15 simulated OPC-backbone–scenario settings, it ranks first and improves the eight-baseline mean government objective by 42.4% and OPC utility by 40.9%.

  • Problem

    Existing policy-learning methods use fixed perceptions, while consensus methods aggregate views without feeding cognitive changes into policy learning, leaving cross-regional coordination between evolving OPC perceptions and policy choices unresolved.

    Source: §1 · p. 2

  • Method

    PolicyMediator combines cognition-constrained Stackelberg learning with verified bilateral mediation that informs OPCs and returns their needs and concerns to governments for policy updates.

    Source: §1 · p. 2, §5.2 · p. 6

  • Results

    PolicyMediator ranks first in all 15 OPC-backbone–scenario settings and improves the eight-baseline mean government objective and OPC utility by 42.4% and 40.9%.

    Source: §1 · p. 2, §6.3 · p. 7

  • Takeaways & Limitations

    The results apply to coordinated-government settings with verifiable policies, comparable regional options, and observable entry and capacity signals.

    Source: §7 · p. 9

    The stationary optimal leader policy guarantee is conditional on compact state and action spaces, continuity and boundedness conditions, positive response temperatures, and a discount factor below one.

    Source: §4.2 · p. 4

Abstract

from arXiv · show

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%.

1 Introduction

Mobile OPCs compare regional support policies, but uncertain perceptions complicate cross-regional coordination. PolicyMediator links verified bilateral mediation with policy learning so OPC choices and concerns inform subsequent policy and skill updates.

1 Introduction

  • OPCs are digital-platform-enabled individual enterprises that can work across regions, access policy information, and relocate with limited reliance on local infrastructure and teams.Governments across China offer distinct support policies and entrepreneurial hubs to attract OPCs.Source: §1 · pp. 1–2
  • Existing policy-simulation methods often fix policy perceptions or infer needs only from choices, while consensus methods rarely update decision-relevant cognition, policy learning, or mediation strategies.Implementation concerns may therefore be mistaken for inadequate support; policy optimization and consensus formation leave different parts of the coordination loop incomplete.Source: §1 · p. 2
  • The coordination challenge is to distinguish cognitive barriers from regional policy mismatches while balancing OPC interests, economic performance, and welfare.OPCs select regions based on expected utility, but incomplete, uncertain, and trust-dependent policy perceptions shape what they know and expect.Source: §1 · p. 2
  • PolicyMediator combines cognition-constrained Stackelberg learning, verified bilateral mediation, and outcome-aware skill evolution to connect OPC perceptions, location choices, and government policy updates.Consensus is about verified facts and workable options; OPCs may disagree and need not accept policies unanimously.Source: §1 · p. 2
  • Across six Chinese regions, five scenarios, and three OPC backbones, PolicyMediator ranks first in all 15 settings and improves the eight-baseline mean government objective and OPC utility by 42.4% and 40.9%.It also exceeds the strongest baseline by 6.3% on the government objective and 15.3% on OPC utility.Source: §1 · p. 2

2 Related Work

Prior work studies multi-agent policy making, consensus and mediation, and evolving agent skills. PolicyMediator connects mediated cognitive updates and downstream outcomes to regional policy learning and skill retention.

Multi-Agent Policy Making

  • PolicyMediator links mediated cognitive updates to allocation and subsequent policy learning, addressing the limitation of policy-making methods that assume fixed policy perceptions.It preserves heterogeneous choices and evaluates policy updates by their downstream government objective.Source: §2 · pp. 2–3

Consensus Formation and Mediation

  • Consensus and mediation methods coordinate beliefs, identify common ground, or structure negotiation, but opinion aggregation and compromise may leave decision-driving cognition unchanged.PolicyMediator instead preserves heterogeneous choices and evaluates policy updates by downstream government outcomes.Source: §2 · pp. 2–3

Self-Evolving Skills

  • Self-evolving skill methods adapt reusable behaviors through experience, but task-success-driven evolution may retain persuasive skills with harmful downstream effects.PolicyMediator evaluates skills by factual validity and downstream system outcomes before retaining, revising, or pruning them.Source: §2 · p. 3

3 Problem Formulation

The formulation represents coordinated regional governments, heterogeneous OPCs, cognition, and regional conditions in a cross-region policy-choice setting. It models how verified policy information and evolving allocation shape cognition and location decisions.

3.1 Framework Overview

  • Regional governments form one cooperative leader while retaining distinct local policies; mediators verify and explain policy conditions, elicit concerns, and update OPC cognition before region or non-entry choices.Allocations and government outcomes inform policy revision, while skill evaluation uses cognition, location, and allocation trajectories.Source: §3.1 · p. 3
  • Policy cycles hold government policies fixed while OPCs update choices over H decision steps; cycle index k is the integer part of t/H.A cycle’s stability window is defined by a suffix in which at most an epsilon fraction of OPCs switch regions at each transition.Source: §3.1 · p. 3

3.2 Government agents.

  • Each regional government selects policy intensities across cash, finance, computing and data, space, talent, and market access, initialized from observed regional policy profiles.The policies are bounded by observed policy-derived limits.Source: §3.2 · p. 3

3.3 OPC agents and regional attributes.

  • OPCs choose among the regions and non-entry, with type determining policy sensitivity through a type-by-policy-category matrix.The five OPC types have differing sensitivity across six policy categories.Source: §3.3 · p. 3
  • OPC policy-information barriers are represented by awareness, trust in policy delivery, and region-specific relevance, with cognition initialized from verified regional evidence.The LLM analysis operator maps an OPC profile and interaction history to its policy-information processing.Source: §3.3 · p. 3
  • Cognitive updates can move in either direction, disclose verified negative as well as positive policy information, and project cognitive components back to [0, 1].Region-specific relevance is the mean of six policy-category perceptions for that region.Source: §3.3 · p. 3
  • Regional location value is represented by baseline payoff, type fit, application difficulty, and capacity; type-specific pre-policy return is the product of type fit and baseline payoff.Market-access support reduces cross-regional entry or relocation costs.Source: §3.3 · p. 3

4 Cognition-Constrained Stackelberg Game

The game combines capacity- and cognition-sensitive OPC responses with a cooperative regional leader’s policy objective. It frames policy learning as a response-induced decision process and uses cooperative zeroth-order learning to optimize policies under mediated, non-differentiable responses.

4.1 Objective Function and Constraints

  • OPCs evaluate regional choices using perceived policy benefit and capacity penalties, with congestion costs increasing quadratically above regional capacity.Their boundedly rational logit response is smoothed by positive temperature; as temperature approaches zero, it approaches maximum-payoff choice.Source: §4.1 · pp. 3–4
  • The government objective equally averages scaled economic performance, employment benefit, and equal policy access, with direct fiscal expenditure included in net economic performance.Policy-access equality uses Jain’s fairness index.Source: §4.1 · p. 4

4.2 Response-Induced Markov Decision Process

  • The response-induced Markov state includes policies, allocation, cognition, regional conditions, and skill state; the generalized Stackelberg model treats boundedly rational follower responses and aggregate-state-dependent payoffs.The leader’s return is induced by the response kernel and government objective.Source: §4.2 · p. 4
  • A stationary optimal leader policy exists for the structured-state abstraction under compact state and action spaces, continuous OPC payoffs, suitable reward and transition conditions, positive response temperatures, and discount factor below one.This result provides a theoretical target for the approximate policy learning that follows.Source: §4.2 · pp. 4–5

4.3 Cooperative Policy Learning

  • Cooperative multi-actor zeroth-order policy-gradient learning optimizes regional policies because mediated OPC responses make the simulator non-differentiable.Each regional actor maps local observations to bounded policy adjustments and shares the government objective as its reward.Source: §4.3 · p. 5
  • Policy actors are updated using paired, two-sided rollouts from the same saved simulator state with matched random seeds, which estimate gradients of a smoothed expected return.Each rollout recomputes mediation, skill evolution, OPC responses, congestion, and the government objective.Source: §4.3 · p. 5
  • The gradient-estimation result bounds policy-learning stationarity error in terms of optimization progress, simulator-state drift, and rollout noise.The realized government objective guides subsequent updates, and learning stops after policy, allocation, and objective remain stable for the required consecutive cycles.Source: §4.3 · p. 5

5 Bidirectional Mediation with Evolving Skills

Section 5 describes evidence-grounded mediation skills and a two-way pipeline in which OPC concerns inform regional policy learning. It also explains how skills are evaluated and revised using outcomes beyond mediation success.

5.1 Skill Structure

  • Each skill links an OPC profile, concern, and verified evidence to an explanation, feasible recommendation, and structured government feedback, while recording applicability and outcomes.Skills also specify applicability conditions, evidence boundaries, mediation routes, recommendation branches, and outcome records.Source: §5.1 · p. 5

5.2 Mediation-Embedded Game Pipeline

  • After policy publication, mediators use policy context, implementation outcomes, OPC profiles, cognitive states, and regional evidence to compose messages; only verified messages update cognition before OPCs choose their own action.Recommendations can identify a region, an alternative, or non-entry.Source: §5.2 · p. 6
  • Government-facing feedback summarizes unmet needs, eligibility difficulties, delivery concerns, and observed responses, augmenting the next government observation while leaving formal policy adjustment to the learned government policy.The learned policy produces the feasibility-projected adjustment; the feedback suggestion is auxiliary.Source: §5.2 · p. 6

5.3 Self-Evolving Skill Lifecycle

  • The skill lifecycle constructs candidates, deploys them to collect traces, evaluates and revises them, then promotes validated skills and prunes ineffective or redundant ones.Uncertain or context-mismatched skills are revised in a temporary pool; validated skills with consistently positive evidence may be promoted.Source: §5.3 · p. 6
  • Skill evaluation measures factual validity, cognition and allocation changes, and marginal government objective, extending beyond persuasion success or user satisfaction.The normalized trace score aggregates verification failures.Source: §5.3 · p. 6

6 Experiments

The experiments test PolicyMediator across regional, scenario, and OPC-simulator settings, then examine its performance, mechanisms, skill evolution, and interactive responses. Results include broad gains over baselines and analyses of how feedback and mediation contribute.

6.1 Experimental Setup

  • The evaluation spans six Chinese regions, five OPC types, six policy categories, and five scenarios, comparing three OPC-simulator backbones and eight baselines on W_G and average realized OPC utility.The scenarios vary entry costs, policy redemption, OPC mobility, and technology-sector conditions.Source: §6.1 · p. 6

6.3 Overall Performance

  • PolicyMediator ranks first on both metrics in all 15 settings, averaging 9.94–9.97 for W_G and 0.719–0.720 for OPC utility, exceeding eight-baseline means by 42.4% and 40.9%.Carbon-PPO is the strongest baseline.Source: Table 1 · p. 7, §6.3 · p. 7

6.4 Further Analysis

  • Removing policy feedback causes the largest decline in W_G (2.31), followed by mediation (2.19), cognitive updating (1.86), skill evolution (1.24), and theory guidance (0.95).The full model has the highest five-scenario mean, 9.97.Source: Table 2 · p. 7, §6.4 · p. 7

6.2 Preliminary Study

  • The simulated city ranking aligns with the entry-friendliness benchmark (Spearman ρ=0.943, p=0.00167), with only the final adjacent pair swapping positions.The comparison uses OPC entry shares and utility across six cities against a descriptive benchmark constructed from crawled public data.Source: §6.2 · pp. 6–7, Table 3 · p. 6

6.4 Further Analysis

  • Cognition changes show predictive structure for OPC transitions (macro-F1 0.54), while feedback utilization averages 0.842 and paired ΔW_G varies by concern type.Utilization falls from 0.91 to 0.19 as need rises under low ambiguity, but remains 0.94–1.00 under the highest ambiguity; median ΔW_G ranges from 0.865 for regional mismatch to 0.121 for knowledge concerns.Source: §6.4 · p. 8, Figure 4 · p. 8
  • After 400 experiences, theory-guided and generic skill evolution achieve nearly identical decision quality, while theory guidance retains higher first-pass verification.The comparison uses 14 manually initialized skills, matched policies, contexts, and budgets, with 20-seed means and 95% confidence intervals.Source: §6.4 · p. 8, Figure 5 · p. 8
  • In the case study, an OPC conditionally accepts support at 75% confidence, applies for computing and space support, and defers a loan pending legal verification.Trust falls from .42 to .32 before returning to .42 after verification, while awareness rises from .45 to .56.Source: §6.4 · p. 8

6.5 Interactive Evaluation

  • PolicyMediator wins 13 of 15 criterion-level interactive comparisons, including all five against the Generic Explainer, but trails CONSENSAGENT in specificity and Contextual Bandit in targeting.Ten trained evaluators made blinded pairwise judgments on randomized outputs.Source: §6.5 · pp. 8–9, Table 4 · p. 8

7 Conclusion, Limitations and Future Work

PolicyMediator combines cognition-constrained Stackelberg learning, verified bilateral mediation, and outcome-aware skill evolution for cross-regional policy learning. The authors report improved regional matching and welfare-related outcomes, with conclusions bounded by simulated OPC responses and cooperative-government assumptions.

7 Conclusion, Limitations and Future Work

  • PolicyMediator combines cognition-constrained Stackelberg learning, verified bilateral mediation, and outcome-aware skill evolution to turn OPC concerns into policy feedback.The authors report improvements in regional matching, economic performance, employment benefits, and policy-access equality beyond disconnected policy optimization.Source: §7 · p. 9
  • The findings apply to coordinated-government settings with verifiable policies, comparable regional options, and observable entry and capacity signals.Source: §7 · p. 9
  • Simulated OPC responses and cooperative-government assumptions limit generalization to competitive jurisdictions and long-term behavior.Future work will consider noncooperative regions, broader macroeconomic settings, and welfare outcomes and trade-offs.Source: §7 · p. 9

A Data Construction

The data and skill pipeline combines public policy and community records with fixed derived inputs, structured prompts, and theory-guided skill evaluation. It normalizes variables for comparability and updates reusable skills based on evidence and outcomes rather than mediation success alone.

A.1 Sources and use

  • Policy and community records provide regional policy, infrastructure, and capacity inputs, while CGSS2023 supplies population, employment, wage, and cognition inputs.The reference library contains 106 policies and 96 communities; the OPCquan1 and other public OPC portals provide records coded from 134 policies and 617 communities.Source: §A.1 · p. 10

A.2 Scales and normalization

  • Policy coverage and infrastructure are min–max scaled across six regions, entry difficulty reverses the infrastructure score, and monetary and employment quantities use dataset-level reference values.Source: §A.2 · p. 10

B.1 Prompt Architecture

  • Prompts represent OPC awareness, expected delivery, and region-specific relevance, constrain mediation to verified facts, and specify skill applicability and behavior limits.Cognition updates use transcript and verified evidence signals, bounded by specified step sizes and projected to [0, 1].Source: §B.1 · pp. 10–11

B.2 Skill Library and Theoretical Routes

  • The skill library uses CMO analysis to match mechanisms to contexts and ELM routes to tailor verified explanations to OPC readiness and concern.Skills cover fact-checking and central, peripheral, and route-switching strategies, all subject to the same evidence boundary.Source: §B.2 · p. 11

B.3 Skill Evaluation and Update Protocol

  • Skill evolution constructs candidates, deploys them temporarily, evaluates and revises outcomes, then promotes effective skills and prunes redundancy.Evaluation includes verification, cognition, choice, allocation, failures, and marginal government objective; cross-region reuse rechecks local evidence and capacity.Source: §B.2 · p. 11, §B.3 · p. 11

C.1 Implementation Details

The implementation uses frozen OPC-simulator backbones, matched and blinded interactive evaluation, and zeroth-order policy learning. Baselines span fixed controls, online and objective-guided optimization, and consensus methods.

C.1 Implementation Details

  • The experiments use three frozen OPC backbones served sequentially on four A100-80GB GPUs, with selected sampling and thinking options disabled for reproducible JSON output.Source: §C.1 · p. 11
  • Interactive evaluators received common rubric training, assessed randomized outputs without method or backbone identities, and saw the same verified policy evidence across comparators.Source: §C.1 · p. 11
  • Zeroth-order learning perturbs all regional actor parameters jointly and updates actors from paired rollout return differences, while comparisons match populations, initial states, feasible ranges, and seeds.Source: §C.1 · p. 11

C.2 Baselines

  • Baselines include nonadaptive controls, online learning, objective-guided optimization, and consensus formation under matched populations, states, feasible bounds, and seeds.Specific methods include No Policy, Static Welfare, Contextual Bandit, Carbon-PPO, Nash Welfare, Belief-Robust, Bi-level Consensus, and CONSENSAGENT.Source: §C.2 · p. 11

C.3 Preliminary Study

The preliminary study checks simulator behavior against a public-data-derived city ranking and examines heterogeneity and diagnostic measures of balance and entry. Its results indicate that structural suitability, congestion, and differences among firm types matter to interpretation.

C.3 Preliminary Study

  • ρ=.943 (p=.0048) for agreement between the simulated six-city ranking and a public-data-derived entry-friendliness proxy with observed policies fixed.Source: §C.3 · pp. 11–12
  • Initial cognitive assessments vary by up to .30 across OPC types within a city, rather than collapsing firms to a common response.Economic and industrial conditions establish most of the cross-city gradient, while policy support modifies rather than overturns structural suitability.Source: §C.3 · p. 12
  • In the 500-OPC stress test, the four leading cities settle at 22–24% entry shares as congestion redirects later entrants, while within-city support balance (.820) exceeds cross-city balance (.774).Hefei reaches .854 type-level balance while one professional-support type has negative net utility (−.09), so balance or entry dispersion alone does not establish welfare improvement.Source: §C.3 · p. 12

D Proofs

The proof appendix establishes existence of a stationary optimal leader policy and gives conditional stationarity guarantees for zeroth-order estimation. Its ethical discussion emphasizes human agency and limits simulated outcomes to analytical insight absent field validation.

D.1 Existence of a Stationary Optimal Leader Policy

  • Under the stated continuity and compact-feasibility conditions, the leader’s Bellman operator attains a maximum and admits a measurable stationary selector whose fixed point matches response-induced return.The proof also relies on bounded, upper-semicontinuous rewards and weakly continuous transitions.Source: §D.1 · p. 12

D.2 Zeroth-Order Gradient Estimation and Stationarity

  • Centered bounded returns and smoothness bound the gradient-estimator variance; with finite iterations and summable drift, its gradient tends almost surely to zero, though parameters may not converge.The proof obtains its bound by summing within-cycle gains and bounding changes between consecutive saved-state objectives.Source: §D.2 · p. 13

E Ethical Considerations

  • The framework calls for institutional review and OPC autonomy; aggregate welfare may not ensure equitable benefits, and simulated outcomes require stakeholder evaluation and field validation.The stated equity concern spans both regions and OPC types.Source: E Ethical Considerations · p. 13

Related papers

Loading 2610.35997v1…