Zuyuan Zhang

Zuyuan Zhang, Sizhe Tang, Tian Lan

Agentic foundation-model service networks handle requests spanning retrieval, planning, generation, verification, and tool use. Unlike traditional communication networks, control performance depends on queue dynamics and contextual memory states, including prefix/KV blocks, retrieved contexts, expert warm states, and verified tool outputs. These states arise from execution history and alter service work and downstream successor laws under finite local budgets. Treating them as passive caches or an independent process leaves a queueing-control gap. To this end, we propose \emph{Memory-Augmented Backpressure} (MABP), a queue--memory control framework for stateful foundation-model service networks (SFMSNs) that jointly models commodity queues and causal contextual memory dynamics. MABP represents each request by service and state types, estimates memory-dependent work, penalties, and successor probabilities, then reads queues and resident memory each slot, selects feasible routing, transfer, activation, and service actions using a memory-dependent pressure score, and retains a budget-feasible subset of resident and newly generated objects. We prove an occupation-measure capacity outer bound with conditional tightness. We show that modeling contextual memory can strictly increase the stability region through work reduction and transition shaping, establishing a separation between memory-aware and memory-oblivious decisions. We also prove throughput and drift-plus-penalty guarantees for exact frame-MABP with bounded-loss extensions to approximate solvers.

Zuyuan Zhang, Vaneet Aggarwal, Tian Lan

Modern network policy control maps intent to sequential placement-control decisions. Bellman-style policy optimization primarily asks which action to optimize, while constraints are commonly handled through penalty, barrier, or Lagrangian mechanisms. We observe that before a value function can certify the best deployment, intermediate signals may already identify many candidates that should be excluded from further optimization. This motivates a complementary direction: \emph{Learning Not to Optimize}. Before a value function is accurate enough to select the best placement-control decision, intermediate signals may already show that candidates are equivalent under state--intent relabeling (quotienting), lead to a uniformly worse future state (dominance), or violate executable network laws (residual screening). \LNOQRD{} uses these computed or learned signals as a shadow process to reshape the domain on which primal policy optimization is performed, thereby reducing the action space. We prove lossless quotienting and dominance under explicit equivariance and monotonicity conditions, bound frontier size and ranking cost, and quantify losses from approximate certificates and primal estimates. Experiments show that \LNOQRD{} reduces small-instance candidates by $75.9\%$ while retaining $90.8\%$ near-oracle coverage and, on large instances, achieves the highest utility and intent satisfaction, the lowest hard-law violation and post-generation latency, and a $73.0\%$ average reduction among candidate-based baselines.