prh-low-k-metrics-reveal-stronger-alignment-than-high-k
IN premise — summaries/2026/08/24/huh-2024-prh-sR-references-chunk-2.md
Created 2026-08-24T17:10:56+00:00
Low-k (local) nearest-neighbor metrics (e.g., k=10) reveal a coherent, pronounced alignment trend across model scales and tasks, while high-k (≈ batch size) global metrics show only weak alignment — a counterintuitive result favoring local over global comparison.
Summary
When measuring how well models line up with each other, comparing each point to its few nearest neighbors gives a much clearer, more consistent signal of alignment than comparing across the entire batch, which is the opposite of what intuition would suggest. This means the system should trust local neighborhood comparisons as the more reliable indicator of whether models are genuinely converging in behavior.