When does admission control reduce congestion? A stochastic ordering approach
arXiv:2607.29439
2026
Dynamics
2 ideas extracted · analyzed Aug 31, 2026
What the math gives to ML
The paper identifies a non-monotonicity failure in state-dependent admission control: rejecting work can lower the current workload, change later decisions, and thereby cause a larger future job to be admitted. The transferable asset is its pathwise coupling perspective, which distinguishes instantaneous load reduction from preservation of stochastic workload dominance over a busy period. This maps naturally to capacity-limited neural systems such as mixture-of-experts routers, dynamic batching, and inference schedulers, where routing or admission decisions affect the state observed by later requests. The most practical adaptations are to train gates against counterfactual workload-order violations and to reduce dependence between admission or type features and downstream service cost.
Ideas from this paper
✗ Failed on benchmark
2026
Train an admission or MoE routing gate not only to reduce its immediate workload, but also to preserve the ordering between a controlled trajectory and a baseline trajectory under the same request stream. Penalize counterfactual events in which the controlled system, after initially rejecting work, later exceeds the baseline workload because its changed state causes a large job to be admitted.
Useful7/10
Difficulty5/10
Novelty7/10
Unverified
2026
Use the paper's independence condition as a design principle: prevent the gate's type-dependent admission behavior from being strongly correlated with downstream service cost. In an MoE or dynamic inference system, this discourages the gate from rejecting cheap requests and then preferentially admitting expensive requests when the queue happens to be shorter.
Useful6/10
Difficulty4/10
Novelty6/10