PIN v5 reports a correction to the previous version's accuracy reporting, clarifying that earlier figures were inflated by a head performing work that depth should have handled. The update also details "members," lightweight adapters added to frozen bases that reach 96.7% of a task-specific model while disturbing the base by exactly zero.
- On CIFAR-10, a residual architecture with the same framework reaches 0.9080 accuracy from 270,277 stored values, representing a fold of 8,939 times compared to published figures for networks of that shape.
- Members fitted by closed-form ridge solve beat the base from five examples and reach 68% of potential capability from twenty examples in 0.6 milliseconds.
- A linear probe on frozen features separates known from unknown categories at AUC 0.8286, enabling gating that lifts accuracy from 0.5936 to 0.8209.
- Members are portable; a translation fitted on fifty public examples recovers 65% of a native member's performance across different bases.
The authors consider this significant because it resolves the issue where confidence is an unreliable signal for unknown categories, demonstrating that gating based on internal feature separation can effectively handle out-of-distribution inputs.