Parameter Configuration Table.

This study proposes a stochastic game model of value perception using a multi-agent reinforcement learning framework to address the double-edged influence of algorithmic recommendation systems on college students’ values. With 73% of students trapped in algorithmic information cocoons and a 23% exposure rate to historical nihilism content on campus platforms, traditional governance methods such as keyword filtering and manual review are insufficient. To address this, we construct a 128-dimensional state space that integrates user personas, policy strength, and value similarity, and define a constrained action space to regulate recommendation weights. A dual-path evolution mechanism is introduced: one path driven by social practice diffuses red values through replication dynamics, while the other uses institutional deterrence to suppress harmful content. The model rigorously proves convergence to a Nash equilibrium with value similarity ≥0.8 when the deterrence factor equals 0.4. Empirical validation on 15,600 users and 12,340 content items, with cross-validation on 5,200 users, shows a 107% increase in value similarity and a 91% reduction in harmful content exposure. Consensus convergence time is reduced by 3.3 times. The results demonstrate the effectiveness and generalizability of this approach, offering a dynamic governance paradigm aligned with national algorithm regulation strategies.