Test-Time Routing for Long-Context Sequence Models

Long-context sequence models assign predictive features to non-uniform internal resources, but pretraining fixes this allocation. We introduce Gyre, a frozen-weight test-time routing framework that rotates signal coordinates at legal internal ports while preserving weights, masks, gates, recurrence rules, and cache coordinates. Gyre derives a Fisher-Gramian alignment objective, proves abstention conditions for flat or already aligned responses, and selects identity, C-FGR, or Stream-FGR under stability and cost constraints. On Qwen3.6-27B, restricted gain predicts held-out NLL reduction, and the selector reaches the deployable quality-efficiency Pareto frontier, improving relative NLL by 0.86\% with an 82.2\% prompt win rate at 7.0\% wall-clock overhead.