Liquid AI Open-Sources Antidoom to Prevent 'Doom Loops' in Reasoning Models
Liquid AI has released Antidoom, an open-source method designed to combat 'doom loops' in reasoning models, a common failure mode where models repeatedly emit the same span of text. This method, based on Final Token Preference Optimization (FTPO), retrains only the first loop-start token to prefer coherent alternatives, rather than replacing the entire sequence. Antidoom has significantly reduced looping rates, for example, from 10.2% to 1.4% on LFM2.5-2.6B and from 22.9% to 1% on Qwen3.5-4B.
Antidoom offers a targeted, efficient solution to a critical problem in reasoning models, improving the reliability and utility of smaller LLMs by enabling them to complete complex tasks without falling into repetitive loops.
Learn one new AI thing every day.
Daily Deck sends you seven plain-English cards like this every morning. Free.
Start free