Training Once Is Enough: Activation Fingerprint Convergence Reveals Ensemble Collapse in Language Model Training

Independently initialized training runs converge to nearly identical activation fingerprints, with phase boundaries synchronized to within ±50 gradient steps. Cross-seed cosine similarity exceeds 0.999 in stable regimes at all scales. Final validation loss CV is 0.41% (124M), 2.43% (1.5B), 1.53% (7B). Convergence tightens with scale. Implies multi-seed training (3–5x compute) provides diminishing returns as models scale, saving 60–80% of ensemble training compute at 7B.