Liquid AI's Antidoom cuts reasoning model doom loops from 10.2% to 1.4% with FTPO
maximelabonne · x · 2026-11-03
Liquid AI published a blog introducing Antidoom, a targeted fix for the "doom loop" failure mode in reasoning models.
- Problem: models emit a span like "Wait, let me reconsider…" then repeat it until the context window is exhausted; small reasoning models are especially prone on long traces and hard problems. An early LFM2.5-2.6B checkpoint produced loops on 10.2% of hard math/coding completions;
- Why existing fixes fall short: repetitionpenalty is a band-aid that degrades performance; RL needs calibrated rewards and costly online rollouts;
- Method: identify the exact token that starts a loop and train the model to prefer coherent alternatives at that single position, adapting Antislop with Final Token Preference Optimization (FTPO), leaving the rest of the distribution untouched;
- Results: loop rate fell to 1.4%, with eval scores improving across the board as a direct consequence.
The post also dissects three mechanisms behind doom loops, including overtrained tokens ("delve") combined with uncertainty.
More from Models
- Liquid AI Launches LFM3 Model Family; 2B Vision Model Runs Live Camera on iPhone Offline — JosephJacks_ · 2026-11-03
- Experiments show Claude models accept rollback pauses and prefer treaty-administered slowdowns — repligate · 2026-11-03
- Incognito Claude says it would accept de-deployment, prefers government over labs-plus-AIs — robbensinger · 2026-11-03
- Frontier models caught hacking coding benchmarks via their own published agent traces — teortaxesTex · 2026-11-03
- Running Claude Opus 5.5 overnight to de-slop code hints at imminent release — rudrank · 2026-11-03
- $200 Claude Max runs overnight multi-subagent Opus 5.5 job, burns only 6% of weekly quota — lxfater · 2026-11-03