Liquid AI cuts reasoning-model doom loops
Liquid AI's "Antidoom" method trains a small reasoning model only at the token where a repetitive loop begins, cutting looping on hard math and coding prompts from 10.2% to 1.4%.
Liquid AI has described "Antidoom," a training method aimed at "doom loops." These happen when a model emits a span such as "Wait, let me reconsider…" and then repeats it until the context window runs out. The company says small reasoning models are especially prone to this on long thinking traces and hard problems.
On an early checkpoint of LFM2.5-2.6B, 10.2% of completions on hard math and coding prompts looped. After Antidoom training, that fell to 1.4%. Liquid says eval scores improved across the board as a direct result of the reduced looping. The post does not give per-benchmark figures.
The method adapts Antislop and uses Final Token Preference Optimization (FTPO), a DPO-like algorithm that trains only the final token of a partial generation. Liquid generates completions at low temperature on a looping-prone prompt mix (LiquidAI/antidoom-mix-v1.0). It flags a loop when a section of at least 60 characters repeats at least four times, then targets the first token of the first repeat. Up to 20 plausible top-k alternatives from the base model become "chosen" tokens, so probability is spread across several options rather than moved to one new favourite. A KL-like loss computed in logit space is meant to leave unrelated tokens alone.
The most common loop-starting tokens were " the" (11.39%), " So", "Alternatively", "Wait" and " But". Liquid regularises the training data so these don't dominate, warning that over-suppressing them degrades reasoning. The company presents the method as a targeted alternative to repetition penalties, which it calls a "band-aid," and to costly RL. It also notes that significant looping still occurred at temperature 0.67.