Antidoom: Breaking the Infinite Loop of Small Reasoning Models

Liquid AI introduces Antidoom, an open-source FTPO method that eliminates repetitive doom loops in small reasoning models like Qwen3.5-4B to boost accuracy.

A reasoning model that repeats the same passage over and over until it saturates its context window: this behavior, which Liquid AI calls a doom loop, primarily affects small models on arduous math and coding tasks. Antidoom, the open-source method the company provides, seeks to eliminate it without the drawbacks of usual fixes, such as the repetitionpenalty, which can degrade quality, or RL, which is more cumbersome to calibrate.

The approach aims to be surgical. The loop almost always starts with a single overtrained token, often an interjection that the model overproduces ("Wait," "So," "Alternatively"). Liquid AI retrains this single token to favor coherent sequences, leaving the rest of the distribution almost intact. The process, called Final Token Preference Optimization (FTPO), belongs to the DPO family: it acts only on the final token being generated and distributes probability over several candidates via a KL-type loss in the logit space.

Advanced measurements show a doom loop rate dropping, for example, from 22.9% to 1% on Qwen3.5-4B, with evaluation scores increasing. Liquid AI also notes a counter-intuitive result: once the loops are eliminated, sampling close to greedy wins out, which nuances the idea that a reasoning model would need a high temperature to explore.