Liquid AI Open-Sources Antidoom to Fix Model Doom Loops

Liquid AI released Antidoom this week, an open-source method designed to address a failure mode in AI reasoning models known as "doom loops." A doom loop occurs when a model repeatedly emits the same sequence of tokens, continuing until its context window is filled. This issue is particularly prevalent in smaller reasoning models when tackling complex problems or long reasoning chains.

On an early checkpoint of the LFM2.5-2.6B model, 10.2% of completions for challenging math and coding prompts resulted in repetitive loops. Following training with Antidoom, this rate decreased significantly to 1.4%. The company reported that evaluation scores improved across various benchmarks, with these gains attributed solely to the reduction in looping behavior. For the Qwen3.5-4B model, looping fell from 22.9% to 1% after the Antidoom process.

Antidoom employs a technique called Final Token Preference Optimization (FTPO), which is similar to Direct Preference Optimization (DPO). Instead of introducing new knowledge about math or code, FTPO focuses on retraining the specific token that initiates a loop. This process trains the model to favor coherent alternative continuations at that single position, while leaving the rest of the token distribution largely unchanged. The pipeline is designed to run in a few hours, and the complete software stack is available as open source.

Liquid AI attributes doom loops to a combination of overtrained tokens and model uncertainty. Certain tokens may be selected more frequently due to their higher prior probability, potentially stemming from synthetic data in training sets. In reasoning processes, these high-prior continuations can include discourse markers like 'Wait' or 'Alternati'. Antidoom's targeted approach aims to correct these specific points of failure without broadly altering the model's capabilities.