Liquid AI's FTPO training cuts small-model "doom loops" from 10.2% to 1.4%
helloiamleonie · x · 2026-09-15
Liquid AI published a blog and open-source repo for "Antidoom", a training method that fixes the "doom loop" failure mode in small reasoning models, which repeat spans like "Wait, let me reconsider…" until the context window fills up.
How it works
- The usual fix, repetitionpenalty, is a band-aid that can hurt performance; RL targeting loops needs carefully calibrated rewards and costly online rollouts.
- FTPO (Final Token Preference Optimization): identify the exact token that starts a loop and train the model to prefer a coherent alternative at just that position via chosen/rejected pairs, leaving the rest of the distribution untouched. The method adapts Antislop.
- Implemented as a custom DPOTrainer in TRL; tutorial and code are open source.
Results: on an early LFM2.5-2.6B checkpoint, the loop rate on hard math/coding prompts dropped from 10.2% to 1.4%, with eval scores improving across the board as a direct result.
Related event: Liquid AI Open-Sources FTPO to Fix Small Models' Doom Loops(2 posts)→
More from Models
- ICML paper: finite context windows impose impossibility results for adversarial regret — burny_tech · 2026-11-02
- SenseTime open-sources SenseNova-Loop-9B: weight-shared 56-layer loop, 256K context, Apache 2.0 — AdinaYakup · 2026-11-02
- User says Gemini excels at long-form fanfic, producing stories up to 29 minutes of reading — Ok-Jello3739 · 2026-11-02
- Sam Altman calls tokens 'a kind of fake unit' that doesn't measure business value — rohanpaul_ai · 2026-11-02
- 'You Can Feel the Gradient Slopes': Dev Contrasts Benchmaxxed RL Models vs General Ones — theteknosaur · 2026-11-02
- LightOnOCR-3 Launches: Tops FRBench, Adds Layout Structure and Chart Extraction — IgorCarron · 2026-11-02