Reducing Doom Loops with Final Token Preference Optimization

Researchers developed "Antidoom," a method that reduces repetitive text generation loops in language models by identifying the exact token that initiates a loop and training the model to prefer alternative coherent tokens at that position. When applied to an early checkpoint of LFM2.5-2.6B, the approach reduced doom loop occurrences from 10.2% to 1.4% on hard math and coding prompts while improving overall performance. The method uses Final Token Preference Optimization to target high-probability tokens like "Wait" and "Alternatively" that become attractive fallbacks when the model is uncertain, avoiding the performance degradation caused by simpler repetition penalties.