Scientific research increasingly depends on multi-author collaboration, yet the systems used to allocate authorship credit remain vulnerable to conflict, strategic behavior, and project breakdown. Although prior work has shown that authors may rationally issue ultimatums over authorship order within a single manuscript, much less is known about how such behavior unfolds over repeated collaborations embedded in evolving academic networks. In this study, we develop a repeated, networked game-theoretic model of co-authorship in which researchers form collaborations over time, accumulate reputation through an evolving friendship network, and, in a subset of cases, learn strategic behavior through deep reinforcement learning. Using large-scale agent-based simulations, we compare myopic and forward-looking authors across mixed populations. We find that strategic agents do not raise fewer ultimatums than greedy agents, but instead learn to avoid insisting after rejection, thereby eliminating destructive manuscript termination. As strategic prevalence increases, paper destruction falls from 0.120 to 0.000 per paper, completion rates rise from 0.853 to 0.970, and average completed papers per agent increase from 15.2 to 16.9. Strategic agents also obtain a substantial utility advantage, reaching 30.8\% when rare, while overall inequality remains stable. These results suggest that reputational feedback and long-term incentives can make academic collaboration more resilient, offering a computational testbed for designing fairer and more productive authorship policies.