Recent advances in AI increasingly emphasize agentic systems in which artificial and human agents collaborate to achieve shared global objectives. One prominent example is collaborative learning, where a collective model is trained using data distributed across multiple agents. A central challenge in building such systems is ensuring both safety and alignment with human values, particularly the fair distribution of rewards when a global goal is achieved. Cooperative game theory provides a useful framework for modeling such cooperation through value functions, which assign a value to each coalition, and through reward allocation functions. Fairness can then be formalized by defining fairness axioms and designing allocation mechanisms that satisfy them. However, traditional cooperative game theory falls short in capturing the nuances of settings like collaborative learning, which involve replicable resources such as data and models. In contrast to classical assumptions of nonreplicability, infinite replicability calls for a broader notion of fairness, supported by new axioms and allocation rules. These must account for asymmetries in mutual benefit among agents, which can otherwise give rise to strategic manipulation and unjust outcomes. The core contribution of this paper is a new axiom template that captures analog beneficence. The key idea is, that within any coalition, each agent should benefit equally from the participation of another. We instantiate this template in multiple ways, develop mechanisms that satisfy the resulting fairness criteria, and prove that they uphold a property we term reciprocal fairness.