LLM-Driven Regime-Adaptive strategy synthesis for Polymorphic Network routing.

Abstract

Dynamic network routing in modern infrastructures, including polymorphic settings where heterogeneous service flows share bandwidth, computing, storage, and forwarding resources, is fundamentally a regime-adaptive control problem, where traffic patterns and topology states exhibit unpredictable structural shifts and distributional heterogeneity. This setting has a polymorphic routing character when multiple service classes share the same infrastructure and the routing logic must account for coupled bandwidth, computing, storage, and forwarding resources. Traditional Deep Reinforcement Learning agents often struggle to generalize across these shifts while static heuristics lack the plasticity to handle diverse operational contexts. To address this, we propose RASS-Evo (Regime-Adaptive Strategy Synthesis via Evolution), which reformulates adaptive control as an event-triggered heuristic code generation task. Unlike continuous re-evolution schemes, RASS-Evo integrates a Regime Perception Module using information-theoretic detection (Kullback-Leibler divergence) to distinguish epistemic regime shifts from aleatoric noise, triggering synthesis only when structural changes are detected. Consequently, the framework employs a Verifier-Guided Heuristic Repair mechanism, leveraging structured constraint-violation traces to guide a Large Language Model in refining executable logic. Empirical validation on Abilene and GEANT topologies demonstrates that RASS-Evo supports effective adaptation to unseen regimes, while reducing computational inference overhead by approximately 60% compared to continuous synthesis baselines.