Liquid AI Released Antidoom: FTPO Method Reduces Hangs in Reasoning Models

Liquid AI on July 7, 2026 open-sourced Antidoom—a method for combating doom-loops in reasoning models. On Qwen3.5-4B, the frequency of hangs dropped from 22.9% to 1%.

What is a doom-loop and why does it need fixing?

A doom-loop is when a reasoning model begins repeating the same text fragments in its chain-of-thought reasoning until it exhausts the entire context window. Instead of providing an answer, the model "gets stuck" in a loop and fails to produce a result.

The problem is particularly painful for reasoning models that "think aloud" using long chains of steps. The longer the reasoning, the higher the probability of repetition: one pattern enters the context, the next token copies it, and the loop closes.

For production systems, doom-loops are not merely academic curiosities: they increase response latency and waste tokens, directly raising inference costs.

How the FTPO method works

Antidoom uses the Final Token Preference Optimization (FTPO) approach. Instead of expensive full model retraining, the method finds the specific token that initiates the cycle and corrects the weights only for this position.

Key facts:

• Method published July 7, 2026 under open-source 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%
• In open access: doom-loop generator, detector, and FTPO trainer

This targeted approach fundamentally differs from standard fine-tuning: only the "problematic" position is changed, the rest of the model's parameters remain untouched, reducing the risk of regressions elsewhere in model behavior.

Why open-source code matters

Liquid AI did not merely describe the method in a paper—it released working tools: a doom-loop detector, a synthetic example generator to reproduce the problem, and the FTPO trainer itself. This allows any team working with reasoning models to apply the method to their own model without waiting for an official vendor fix.

For open-source LLM developers, the synthetic example generator is particularly valuable: it lets one 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 in itself is unusual for an AI company. If FTPO demonstrates consistent results on architectures beyond LFM and Qwen, the approach could become a standard step in fine-tuning reasoning agents before production release.

Frequently asked questions

What is Final Token Preference Optimization?

FTPO is a method of targeted fine-tuning: instead of optimizing the entire model, it finds the specific token that triggers the repetition loop and corrects only its weights. This is faster than full fine-tuning and reduces the risk of unwanted changes in the rest of 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-source code allows the community to independently test results on other architectures.