Can Savcı's CAN method keeps MARL teams fair without central allocator

A new paper by Can Savcı tackles a longstanding problem in cooperative multi-agent reinforcement learning (MARL): fair teams that maximize egalitarian welfare are easily exploited by a single selfish agent that free-rides on the surplus fair agents forgo. A centralized need-based allocator can prevent this, but centralization is often infeasible or undesirable. Savcı shows that under graded contention (where a contested resource delivers only 1-c, wasting c), decentralized leverage exists: a worst-off cooperator that contests a free-rider strictly improves over yielding.

The key contribution is the CAN policy (cross-attention over observed behaviour). CAN is permutation-equivariant and infers the number of free-riders, then responds proportionally—turn-taking when none, contesting just enough when some. Trained against an adversarial league (PSRO), CAN keeps best-response exploitability low (ρ≈1.2–1.5, vs. ρ=N unprotected) across the contention range. Efficiency remains near 1.0 at zero contention and 0.83–0.96 at high contention, approaching a centralized oracle without any central allocator. The paper also identifies limitations: CAN works best where contest leverage is strong (multi-server games), degrades under winner-take-all, and shows fragility with weak leverage or zero-shot transfer to larger teams.