Test-Time Routing for Long-Context Sequence Models

Long-context sequence models store information in internal directions that are not equally useful. Some directions are retained, gated, or selected more strongly than others. Pretraining fixes the usual assignment of features to these directions, but a new prompt can place predictive signal in poorly matched coordinates. We introduce Gyre, a frozen-weight test-time routing method that measures prompt-local loss sensitivity and host response. It then rotates only certified signal ports so predictive directions move toward useful host directions. Gyre leaves weights, masks, gates, recurrence rules, and cache coordinates unchanged. It abstains when the response is flat or already aligned, 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 pointwise diagonal selector improves relative NLL by 0.86\% with an 82.2\% prompt win rate at 7.0\% wall-clock overhead. These results place Gyre on the measured quality-efficiency Pareto frontier for long-context inference.