Collective search-and-capture under competing assignment policies

What the math gives to ML

The paper identifies a concrete collective-search mechanism: when many agents compete for depleted targets under finite-range eligibility, the assignment policy can dominate the individual agents’ persistence dynamics. Maximum-cardinality matching on the agent-target candidate graph substantially reduces the extreme late-time completion tail compared with single-round or cascading greedy assignment. This transfers naturally to sparse neural routing, especially mixture-of-experts or distributed token-to-device dispatch, where independently selected experts create avoidable unmatched tokens, overloads, and long-tail latency; replacing greedy routing with capacity-constrained matching gives a directly testable improvement.