Liquid AI Open-Sources Antidoom: The Ultimate Token Option (FTPO) Method That Reduces Doom Loops in Consulting Models

Liquid AI released The Antidooman open-source approach that addresses common failure modes in conceptual models. That failure mode is the doom loop. In the doom loop, the model releases the span. It then repeats that span over and over again. Output continues until the content window is finished. Small thinking models are more prone to this, especially in long thinking tracks and difficult problems.

In an early examination of the LFM2.5-2.6B10.2% of the completion of strict calculations and coding notices produced repetitive loops. After Antidoom training, that rate dropped to 1.4%. Eval scores improved across the board, thanks to reduced loping.

The TL;DR

• Antidoom reduces doom loops by retraining only the first loop-start token.
• FTPO distributes opportunities in many complementary ways, not in one position.
• LFM2.5-2.6B looping decreased by 10.2% to 1.4%; Qwen3.5-4B fell 22.9% to 1%.
• The pipeline runs in a few hours, and the full stack is open source.

What is Antidoom?

Antidoom is a targeted fix, not a sample wide change. Finds the exact symbol that starts the loop. It then trains the model to select alternatives that are compatible with that single position. The rest of the distribution remains untouched.

The method is flexible Antislop. It trains on selected/rejected pairs that represent a single completion token. A training algorithm Final Token Preference Optimization (FTPO)such as DPO.

Training teaches the model nothing new about math or code. It clears a loop that blocks responses that the model may already be generating.

Anatomy of the Doom Loop

The Liquid AI team points out that doom loops have three mechanisms that work together:

Method 1: overtrained tokens and uncertainty. Other tokens will likely be chosen in general. Well-known examples in the field include ‘delve’ and ‘testamente.’ The Liquid AI team notes that this can be traced back to artificial data in the training set. In logic sequences, early progressions often include speech markers such as ‘Wait’ or ‘Other.’ These tokens are not inherently bad. They can mark a useful strategy change, validation step, or branch. If the model is uncertain or stuck, it instead becomes an attractive continuation of backwardness.

For the first LFM2.5-2.6B checkpoint, the most common loop initiation tokens are the following.

Method 2: the front core tightens the loop. Each iteration pushes every token in time in the probability that Duan et al. read this in their circular consultation work. They associate it with a “V-shaped” attention pattern. They found that semantic repetition precedes textual repetition.

Method 3: selfish sampling. Imaging models typically operate at low temperatures to obtain robust, reproducible traces. At temperature 0, the most likely token is always selected. The loop is fixed in place and has no exit. Liquid AI reports a significant increase even at temperature = 0.67. Low temperatures exacerbate the problem.

How Antidoom Gets Failed

Antidoom produces a finish with a quick mix designed to wake up the loop, at a low temperature. That mix ships as LiquidAI/antidoom-mix-v1.0 The dataset. A loop is detected when a section repeats at least four times, over 60 characters.

The method then directs the the first sign of the first repetition. In that position, it takes the alternatives of the top-k log-prob model. Filters short or non-alphanumeric audio. It stores 20 substitutions that sound like selected tokens.

Each training line is a quick start episode, one rejected sign, and one or more selected tokens. The selected and rejected distributions are always performed before training. Besides, few criminals like it wait, Soagain i it can dominate and too much stress can undermine thinking.

The rule to find it is easy to express in code. The caption below shows.

# A loop = a unit repeating >=4 times, spanning >=60 characters.
# Returns the index of the first token of the first repeat (the target), else None.
def find_loop(text, min_repeats=4, min_chars=60):
    n = len(text)
    for span in range(1, n // min_repeats + 1):
        start = 0
        while start + span * min_repeats <= n:
            unit = text[start:start + span]
            repeats = 1
            pos = start + span
            while text[pos:pos + span] == unit:
                repeats += 1
                pos += span
            if repeats >= min_repeats and span * repeats >= min_chars:
                return start + span          # first token of the first repeat
            start += 1
    return None

Each loop obtained is one training row. A structure is a simple tuple.

# One FTPO training row, per the post's [prefix, rejected, chosen] format.
row = {
    "prompt": prefix_up_to_the_loop,   # text before the first repeat
    "rejected": " Wait",               # the single token that started the loop
    "chosen": [" So", " Since", " The", " Therefore"],  # up to 20 alternatives
}

Final Token Preference Optimization (FTPO)

FTPO is a preference processing algorithm similar to DPO. The training sample consists of information, selected progress, and rejected progress. It is designed to replace a number of tokens, with minimal disruption to the model otherwise.

An FTPO differs from a DPO in four ways:

1. Final token training: Trains only the token sequence of the sequence between generations.
2. Multiple tokens are selected per sample: Spreads probabilities across a group of alternatives, so one overtrained token is not simply replaced by another.
3. KL-like losses at the entry point: Skips the softmax and calculates the divergence from the index in the logs, avoiding stress on unrelated tokens.
4. A two-part adaptation: Selected and rejected logs move freely, while the remaining word remains bound.

In the Antidoom implementation, the model is trained for one epoch with LoRA. Higher LoRA levels of 128-256 gave the best results. The training includes all the attention and projection of MLP, as well lm_head. Reading levels range from about 4e-6 to 2e-5.

Training uses early termination chosen_winthe share of samples in which the selected tokens are struck. Stop at chosen_win=0.35 reduce doom-loop levels from 20-30% down to 1-2%. Training for a long time tends to degrade the model.

For the first test of LFM2.5-2.6B, the generation of the training set took about one hour on 8x MI325 GPUs. Training then takes about one to two hours on a single MI325 GPU. Generation stops after collecting 20k pairs.

How Antidoom Compares to Normal Fixes

Results

After training, the doom-looping rate in the first test area of ​​LFM2.5-2.6B decreased from 10.2% to 1.4%. Eval scores improved across the board, thanks to reduced loping.

The Liquid AI team also ran the pipeline Q3.5-4Bknown to rotate during consultation. Its doom-looping rate dropped from 22.9% to 1% under selfish samples. Eval scores increased significantly.

The eval result changed inversely with the doom-loop rate as the temperature increased. After training, both models showed performance degradation around temp=1.0. This is to be expected, as high temperature samples can favor less selective tokens. Once the loop was removed, the nearly selfish samples provided the strongest scores for the models tested.

The Liquid AI team raises a related point about generalization. The belief that high temperatures facilitate reasoning may be linked to a doom-looping effect. In their tests, when the loops are gone, the near-selfish samples perform best.

More cycles can help. The first round rejects tokens that cause a loop and weights towards others. That can reveal new points of failure, which the second round can then target.

Interactive Descriptor