A new method called Antidoom targets "doom loops," a failure mode where AI models repeatedly generate the same text during inference, particularly on complex reasoning tasks. This behavior, often seen in smaller models, exhausts the context window. Existing fixes like repetition penalties can degrade performance, while reinforcement learning demands costly calibration. Antidoom offers a more precise solution by identifying the specific token initiating a loop and training the model to prefer coherent alternatives at that single point.

Antidoom adapts Final Token Preference Optimization (FTPO), a technique that trains on chosen/rejected completion pairs. The process involves generating completions designed to elicit loops, then identifying the first token of repeating sequences. The model is trained to favor plausible substitutes for these problematic tokens, using a modified loss function that minimally impacts unrelated parts of the output distribution. This targeted approach proved effective on an early checkpoint of LFM2.5-2.6B, reducing doom loops from 10.2% to 1.4% and improving overall evaluation scores.

The method also demonstrated success on Qwen3.5-4B, cutting its doom-looping rate from 22.9% to 1%. The training set generation for Antidoom takes approximately one hour on 8x MI325 GPUs, with subsequent training requiring one to two hours on a single MI325 GPU. The primary benefit is not teaching the model new knowledge, but rather removing a failure mode that prevents it from producing answers it already has the capacity to generate.