Liquid AI releases Antidoom: FTPO method cuts hangs in reasoning models

On July 7, 2026, Liquid AI open-sourced Antidoom, a method for fighting doom-loops in reasoning models. On Qwen3.5-4B, the hang rate dropped from 22.9% to 1%.

What a doom-loop is and why it needs fixing

A doom-loop is when a reasoning model starts repeating the same fragments of text in its chain of thought until it exhausts the entire context window. Instead of an answer, the model gets stuck in a loop and returns no result.

The problem is especially painful for reasoning models, which "think aloud" through long chains of steps. The longer the reasoning, the higher the chance of repetition: one pattern lands in the context, the next token copies it, and the loop closes.

For production systems, doom-loops are not just an academic curiosity: they increase response latency and burn extra tokens, directly raising inference costs.

How the FTPO method works

Antidoom uses an approach called Final Token Preference Optimization (FTPO). Instead of expensively retraining the whole model, the method finds the specific token where the loop begins and adjusts the weights for that position only.

Key facts:

• The method was published on July 7, 2026 under an open license
• On LFM2.5-2.6B (Liquid AI's own model), doom-loops fell from 10.2% to 1.4%
• On Qwen3.5-4B, from 22.9% to 1%
• Publicly available: a doom-loop generator, a detector and the FTPO trainer

This targeted approach differs fundamentally from standard fine-tuning: only the "problem" position changes, and the model's other parameters are left untouched, which lowers the risk of regressions in the model's other behavior.

Why open code matters

Liquid AI did not just describe the method in a paper; it released a working toolkit: a doom-loop detector, a generator of synthetic examples to reproduce the problem, and the FTPO trainer itself. Any team working with reasoning models can apply the method to its own model without waiting for an official fix from the vendor.

For open-source LLM developers, the synthetic example generator is especially valuable: it lets them build a dataset to reproduce the problem on a specific architecture, even if doom-loops are rare in real data.

What this means

Antidoom closes a specific technical gap in reasoning models and does so openly, which is unusual for an AI company in itself. If FTPO shows stable results on architectures beyond LFM and Qwen, the approach could become a standard step when fine-tuning reasoning agents before they go into production.

Frequently asked questions

What is Final Token Preference Optimization?

FTPO is a targeted fine-tuning method: instead of optimizing the whole model, it finds the specific token that starts the repetition loop and adjusts only its weights. It is faster than full fine-tuning and lowers the risk of unwanted changes elsewhere in the model.

Which models has Antidoom been tested on?

Liquid AI tested the method on two models: its own LFM2.5-2.6B (doom-loops fell from 10.2% to 1.4%) and the third-party Qwen3.5-4B (from 22.9% to 1%). The open code lets the community verify the results on other architectures.