Liquid AI open-sources Antidoom, a final-token training fix for reasoning doom loops
Final Token Preference Optimization retrains only the token that starts a repetition loop — cutting doom-loop rates from 22.9% to 1% on one test model with a one-GPU training run.
Liquid AI released Antidoom, an open-source method for suppressing "doom loops" — the failure mode where a reasoning model repeats the same span until it exhausts its context — with code and a companion dataset published in July 2026, as reported by remio.ai and veonib.
Context
Doom loops are a practical reliability problem for reasoning models: once a model enters a repetitive cycle, it can burn its entire output budget without producing an answer. The failure is distinct from ordinary repetition penalties because it emerges from the model's own extended reasoning traces. A complementary data point circulating the same weeks: a practitioner documented GPT-5.5 reasoning traces truncating around 516-token clusters — evidence, in one researcher's phrasing, that "the layer where models think can break on its own."
What changed
Antidoom uses Final Token Preference Optimization (FTPO): it identifies the token that starts a loop and retrains only that position, rather than applying a global repetition penalty or full-sequence preference optimization. Per the coverage, generation, detection and the FTPO trainer are all open source, along with the `LiquidAI/antidoom-mix-v1.0` dataset. Reported results:
- LFM2.5-2.6B: doom-loop rate reduced from 10.2% to 1.4%.
- Qwen3.5-4B: doom-loop rate reduced from 22.9% to 1%.
- Evaluation scores on standard benchmarks were preserved or improved after FTPO training.
- Training takes roughly one to two hours on a single MI325 GPU.
Why it matters
Surgical post-training — touching only the token position where a failure begins — is a cheaper and potentially safer intervention than broad preference training, which risks degrading the reasoning behaviors that make these models useful. If the approach transfers, it gives the open-model community a cheap reliability patch applicable to any weights it can train. The permissive release of both the detection tooling and the trainer means other labs can test that transfer claim directly.
Details
The mechanism, as described in coverage: a detector flags the onset of a repeated span, the loop-start token is identified, and FTPO applies preference optimization at that single position — discouraging the token choice that initiates the loop while leaving the rest of the trace untouched. The one-to-two-hour single-GPU training cost makes replication feasible for academic and hobbyist practitioners.
Limitations and caveats
Only two small models (2.6B and 4B parameters) have been tested publicly; transfer to models above 30B parameters and to agentic tool-call loops — a related but distinct failure mode — is unverified. All reported numbers are vendor-side results relayed through trade coverage; no independent reproduction had appeared at publication. The dataset and code release enables external validation but does not itself constitute it.
Sources
- remio.ai — Liquid AI open-sources Antidoom method to cut doom loops in reasoning models (2026-07-20)
- veonib — Liquid AI Antidoom: doom loops and FTPO (2026-07-15)
*Update note: This post was last reviewed on 2026-07-22. Watch for independent reproductions and tests on larger models.*
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.