HydraMARL: Isolating the Parameter-Sharing Spectrum in Multi-Agent RL

How much agent-specific capacity is needed to turn a shared representation into specialized behaviour? Recent multi-agent reinforcement learning methods combine partial parameter sharing with hypernetworks, adapters, clustering, or diversity objectives, making their contributions difficult to separate. We introduce HydraMARL, a shared trunk with per-agent heads, applied independently to actor and critic. A single head-share parameter interpolates between full sharing and full independence at a matched total parameter budget. Across three VMAS tasks and four on-policy and off-policy algorithms, a fixed 20% head share matches or exceeds the evaluated comparators' mean returns on nine of twelve task-algorithm pairs. Sweeps typically favour small actor heads, while the useful critic head share depends on task and critic input locality. A single-seed extension to actor-critic representation sharing provides preliminary results. These experiments establish partial sharing as a sufficient explanation for many observed gains, without isolating the causal contribution of each comparator's mechanism. HydraMARL offers a simple, budget-matched baseline for studying how shared representations support differentiated actions. All representations are learned from scratch; pretrained transfer remains an open question.