A study of four Polish Bielik models (1.5B-11B parameters) demonstrates that unsupervised activation dispersion measures can distinguish known entities from fabricated ones with high accuracy before any answer token is generated.
- Inverse participation ratio and spectral entropy over post-SwiGLU MLP activations separate known from fabricated entities at AUROC 0.95-1.00 across all domains and scales.
- The signal transfers across entity types, survives held-out layer selection, and persists on real names with an AUROC of 0.96-1.00.
- While the representational signal reaches ceiling performance at 1.5B parameters, behavioral factual reliability scales sharply, with correct answers increasing from 0 to 19 out of 42 as model size grows.
- Despite this internal awareness, models almost never abstain; an audit found only two refusals and one hedge among 2,520 answers.
The findings indicate that entity familiarity and factual reliability are distinct phenomena operating on different scaling curves, challenging the assumption that internal activation signals directly correlate with behavioral correctness.