Developers Warn Against Training on Renormalized Probability Scores

Developer spikedoanz warns that renormalizing probability scores before training discards confidence information, as (0.09, 0.01) becomes identical to (0.9, 0.1), and training on such distorted signals with KL loss could degrade or break the model.

2026-09-17 ~ 2026-09-17 · 2 related posts

• Renormalizing probability scores destroys calibration, dev warns in model training debate — spikedoanz · 2026-09-17
• Renormalized scores erase confidence: KL training on them would break the model — spikedoanz · 2026-09-17