Training-time intervention yields 63.4% blind-pair human preference at matched val-loss (1.2B params, 320 judgments, p = 1.98 × 10⁻⁵) [R]

TL;DR. I ran a blind A/B preference evaluation between two 1.2B-parameter LMs trained on identical data (same order, same seed, 30K steps / 3.9B tokens) - one with a Predictive-Coding-inspired precision-weighted gain function plus per-layer divergence-scaled gradients, one with standard cross-entropy. Smoothed val loss between the two is statistically indistinguishable (0.004-nat difference, well inside step-to-step noise). Ten judges (seven humans, three foundation models across Anthropic / Ope