Resonant Phase-Locking: Stabilizing Long-Horizon Reasoning and Continual Adaptation via Phase-Synchronized Feature Routing

We introduce Resonant Phase-Locking (RPL), a simple, general mechanism for stabilizing long-horizon reasoning and sequential task adaptation in transformer models. 
RPL augments the residual stream with a lightweight oscillatory state per token—phase θ ∈ [0, 2π) and learnable base frequency ω—and gates attention and MLP updates by phase alignment. 
Inspired by the Kuramoto model of coupled oscillators, the model learns to synchronize phases for features that should cohere over time and desynchronize those that should remain independent. 
We implement RPL as (i) phase-modulated attention scores that reward small phase differences (cos Δθ) and (ii) phase-aware MLP routing that locks salient features into stable, reusable "rhythms." 
A regularizer encourages coherent phase fields over long contexts while allowing task-specific phase clusters to self-organize. 
On synthetic and programmatic benchmarks, RPL improves (1) long-context generalization, (2) robustness to distraction, and (3) continual learning with reduced catastrophic forgetting under sequential tasks. 
Ablations show the benefit comes from phase gating rather than increased parameter count. We release code and a reference implementation.