Adaptive Regulation via Dual-Layer Evolution (ARDE): A Multi-Agent Approach to Balancing Efficiency, Fairness, and Diversity in Crowdsourced Platforms

With the rapid rise of the gig economy, crowdsourced delivery platforms (e.g., Uber Eats, DoorDash, Deliveroo) have become a key component of urban logistics. However, their competitive order-grabbing mechanisms incentivize riders to pursue individual reward maximization, leading to strategic homogenization, severe efficiency degradation, and growing income inequality. This paper proposes \textbf{ARDE}, a novel dual-layer reinforcement learning framework that aligns individual rationality with collective welfare. The outer agent adaptively generates governance signals based on system-level indicators (strategy entropy, Gini coefficient, completion rate), while the inner agents optimize tactical decisions via Diffusion Q-Learning to encourage cooperative behavior and avoid premature convergence. A language model-driven reward shaping module dynamically incorporates fairness and diversity considerations, providing interpretable and adaptive policy guidance. This paper build a high-fidelity multi-agent simulation environment using the LaDe-D dataset to replicate realistic order allocation dynamics. In contrast, ARDE achieves a balanced and sustainable outcome, improving entropy stability ($0.997 \pm 0.184$), reducing inequality (Gini change $-1.3\%$), and maintaining high system efficiency, outperforming both rule-based and reward-shaping baselines. These results provide the first empirical evidence that algorithmic governance can overcome the “individual rationality $\rightarrow$ collective suboptimality” dilemma in crowdsourced delivery platforms. ARDE offers a scalable, low-complexity solution for intelligent governance in gig economies and other large-scale socio-technical systems.