inference-failure

Liquid AI introduces Antidoom, a method using Final Token Preference Optimization to reduce repetitive doom loops in small reasoning models during inference, cutting loop rates from 10.2% to 1.4% while improving evaluation scores.