Liquid AI releases Antidoom to stop reasoning model doom loops by retraining overfit transition tokens
The open-source method cut Qwen3.5-4B loop rates from 22.9% to 1%
The open-source method cut Qwen3.5-4B loop rates from 22.9% to 1%
Commentary on X
Highest ranked
Honestly that reduction is way bigger than I expected. Going from 22.9% to 1% on Qwen3.5-4B isnt a small quality of life improvement that changes whether the model actually feels usable in real workflows. The doom loop problem has always been one of those things that makes reasoning models look smart but unreliable. So if Antidoom can cut that failure mode down this much while improving eval scores too, that’s a pretty meaningful step forward for open-source reasoning models.
View all
Antidoom spirals!? Super cool 1. Ai models often get into doom loops 2. Often because smaller models end up over trained on particular words, for example “wait”, “so”, “alternatively” 3. So they do brain surgery to retrain that word seems like it also conceivably be used to remove all sorts of model quirks? Is this the death of delve – lists goblins yes absolutely??? Now if only someone could come do this to me at 1am so I don’t have the doom spirals too that would be great thx
