Liquid AI Open-Sources Antidoom: A Final Token Preference Optimization (FTPO) Method that Reduces Doom Loops in Reasoning Models
Liquid AI has open-sourced Antidoom, a method that reduces harmful repetition loops in reasoning models by retraining only the problematic tokens using Final Token Preference Optimization (FTPO).
- Antidoom targets 'doom loops' in reasoning models by retraining only the problematic tokens using FTPO, reducing repetition rates by up to 96%.
- Benchmarks show doom-loop rates fell from 10.2% to 1.4% on LFM2.5-2.6B and from 22.9% to 1% on Qwen3.5-4B after applying Antidoom.
- Liquid AI has open-sourced the generation code, detection tools, and FTPO trainer, making the method accessible for broader adoption.
- The technique addresses a critical issue in reasoning models, improving reliability and performance in long-context scenarios.
Liquid AI has introduced Antidoom, an open-source technique designed to mitigate 'doom loops' in reasoning models. These loops occur when a model repeats a span of tokens until its context window is exhausted, degrading performance and reliability. Antidoom identifies the token that triggers the loop and applies Final Token Preference Optimization (FTPO) to retrain only that position, effectively breaking the cycle.
The method has demonstrated significant improvements in benchmarks. On the LFM2.5-2.6B model, doom-loop rates dropped from 10.2% to 1.4%, while the Qwen3.5-4B model saw a reduction from 22.9% to 1%. Liquid AI has made the generation code, detection tools, and FTPO trainer available under open-source licenses, enabling developers to integrate and adapt the solution for their own models.
Source: Liquid AI Open-Sources Antidoom: A Final Token Preference Optimization (FTPO) Method that Reduces Doom Loops in Reasoning Models. Read the full piece at the source.
Provides a practical, open-source solution to mitigate harmful repetition loops in reasoning models, improving model reliability.
Enhances the quality of AI-driven reasoning systems, reducing errors and improving user trust in deployed models.
Offers a clear example of applied optimization techniques in AI, with open-source code for hands-on learning.
- doom loop
- A phenomenon in reasoning models where a span of tokens is repeatedly generated until the context window is exhausted, degrading performance.
- Final Token Preference Optimization (FTPO)
- A method that retrains only the problematic token in a doom loop to break the repetition cycle.
New artificial intelligence model reveals invisible multiple sclerosis lesions - News-Medical
Grants support new research on role of AI in the arts - Binghamton University
How novice coders can develop AI programs for military applications - MIT News
Where Medicine and Artificial Intelligence Converge - New York Institute of Technology
Unified Context as the Missing Foundation for Enterprise AI - Emerj Artificial Intelligence Research
AI ToolsCohere Transcribe Arabic is an open-source model built for Arabic's toughest transcription problems
Cohere has launched Transcribe Arabic, an open-source speech recognition model designed for Arabic dialects and bilingual speech, outperforming Whisper and OmniASR.
When AI hurts people, who’s to blame? Global experts grapple with accountability - UN News
The United Nations is convening global experts to address legal and ethical gaps in assigning blame when AI systems cause harm.
AI ToolsThe AI Bill Grows in the Agent Loop
A new open-source tool reduces token waste in AI agent workflows by up to 99% by optimizing tool schemas.
AI ToolsBuild a unified semantic layer across datasets with multi-dataset Topics in Amazon Quick
Amazon QuickSight introduces multi-dataset Topics to unify semantic layers across datasets, enabling cross-dataset queries for chat agents. A retail analytics demo showcases the feature in action.
AI ToolsAnthropic's Claude Cowork AI agent is now available on mobile and web
Anthropic expands its Claude Cowork AI agent to mobile and web platforms, enabling background task handling and mobile notifications for user decisions.
GSA praised for initial changes to AI draft regs, but more work needed - Federal News Network
The U.S. General Services Administration has received initial praise for proposed AI regulations, though critics argue the changes don’t go far enough to address key concerns.