explainx / blog
Liquid AI Antidoom (July 7, 2026) uses FTPO to retrain single overtrained tokens — Wait, So, Alternatively — cutting LFM2.5 doom loops 10.2% to 1.4% and Qwen3.5-4B 22.9% to 1%. Open source on GitHub.
Jul 8, 2026
A 303-point HN thread and Ariya Hidayat's walkthrough put Kokoro back in focus — 82M params, CPU-only, OpenAI speech API compatible. explainx.ai covers setup, benchmarks, limitations, and community workarounds.
Jul 7, 2026
GLM-5.2 launched in June; by July it is the open model developers actually keep using — MIT license on Hugging Face, 1M-token long-horizon coding, and viral praise from tinygrad's George Hotz. Here's the adoption story beyond the export-ban headlines.
Jul 4, 2026
Three founders — Gary Tan of Y Combinator, Daniel of Able Police, and Farbood of A-List — sat down on the Naval podcast to talk through the questions AI has made unavoidable: is intelligence still the bottleneck, will open source catch closed models, and what happens to jobs when the harness doesn't need a human anymore.
On July 7, 2026, Liquid AI published Antidoom — an open-source fix for doom loops: the failure mode where reasoning models repeat the same span ("Wait…", "Alternatively…", "So…") until the context window dies. The @liquidai thread frames it as surgical: retrain one overtrained token at the loop's start, leave the rest of the distribution alone.
For anyone running local Qwen reasoning models or small LFM checkpoints, this is the first production-oriented answer to a problem that looked like bad prompting — but is often distribution pathology at a single token position.
| Question | Answer (Liquid AI, July 2026) |
|---|---|
| Problem | Doom loops — repetitive degeneration on hard math/coding prompts |
| Method | FTPO (Final Token Preference Optimization) — DPO-family, one trailing token |
| LFM2.5-2.6B early | 10.2% → 1.4% loop rate; evals up (no new skills taught) |
| Qwen3.5-4B | 22.9% → 1% under greedy sampling |
| Loop starters | Often the, So, Alternatively, Wait, But |
| Train cost | ~1 hr dataset gen (8× MI325), ~1–2 hr LoRA train (1× MI325); tweet cites ~2 hr + 1–2 hr on H100 for 2–4B |
| Sampling after fix | Near-greedy often beats high temp once loops are gone |
| Code | github.com/Liquid4All/antidoom |
Liquid AI's definition:
The model emits a span (often something like "Wait, let me reconsider…"), then repeats the same span again and again, until the context window is exhausted.
This is distinct from:
repetition_penalty band-aids — inference-time reweighting that can hurt qualitySmall thinking models loop more on long traces and hard problems per Liquid's prior LFM2.5-Thinking work.
HN and X reactions match lived experience: Claude Opus 4.8 users report surprise loops; local Qwen builders call Antidoom "finally usable."
Liquid AI identifies a stack:
| # | Mechanism | Effect |
|---|---|---|
| 1 | Overtrained tokens + uncertainty | High-prior discourse markers (Wait, So, Alternatively) become fallback when stuck |
| 2 | Prior context reinforces | Each repeat makes the span more likely — probabilities climb toward 1 |
| 3 | Greedy / low-temp sampling | No exit when loop token dominates; high temp helps less once mechanism 2 saturates |
On an early LFM2.5-2.6B checkpoint, tokens starting loops:
| Token | Share of loop starts |
|---|---|
the | 11.39% |
So | 4.51% |
Alternatively | 3.22% |
Wait | 2.56% |
But | 2.46% |
Prior work Liquid cites: Wait, Wait, Wait… Why Do Reasoning Models Loop? (Pipis et al.), Circular Reasoning (Duan et al. — semantic repetition before textual repetition).
At that position:
Wait/So/ the does not over-crush reasoningFinal Token Preference Optimization vs standard DPO:
| FTPO property | Why it matters |
|---|---|
| Trailing token only | Mid-generation surgical edit |
| Multiple chosen tokens | Avoid swapping one overtrained token for another |
| KL-like loss in logit space | Less gradient spill to rest of vocab |
| Two-part regularization | Chosen/rejected move freely; other logits stay near reference |
Practical hyperparameters (from blog):
lm_head)chosen_win ≈ 0.35 — training longer creates new loopsLiquid AI's conclusion: the belief that reasoning models need high temperature may conflate exploration with escaping doom loops. Remove loops → near-greedy often optimal.
From HN/X edge reports — worth tracking:
For agent harnesses, pair Antidoom-tuned weights with loop engineering stop conditions — fixing generation pathology does not replace /goal evaluators.
# Full pipeline: github.com/Liquid4All/antidoom
# 1. Generate doom-loop preference pairs (antidoom-mix-v1.0)
# 2. FTPO train with LoRA
# 3. Eval loop rate + downstream benchmarks
Liquid AI timing (blog / tweet):
| Stage | Hardware | ~Duration (2–4B) |
|---|---|---|
| Dataset generation | 8× H100 / MI325 | ~1–2 hours |
| FTPO training | 1× H100 / MI325 | ~1–2 hours |
Stops after collecting ~20k pairs (depends on base model loop rate).
Antidoom adapts Antislop (Paech et al., ICLR 2026) — Liquid's framework for identifying and eliminating repetitive patterns in LMs. FTPO is the final-token specialization for reasoning-loop starts rather than generic slop phrases like "delve."
Same design philosophy as Antidoom's tweet:
The training teaches the model nothing new. It clears the looping that blocked answers the model could already produce.
| Before Antidoom | After Antidoom |
|---|---|
| Crank temperature to escape loops | Try near-greedy first |
| Blame prompt for "Wait Wait Wait" | Check loop rate on hard prompt mix |
repetition_penalty as default | Prefer distribution surgery at loop token |
| Small reasoning models "unusable" on hard tasks | Qwen3.5-4B class may recover lost accuracy |
Complements fix local LLM looping sampler advice — Antidoom attacks the model weights, not just decoding knobs.
/goal loopsOfficial sources
Loop rates, hyperparameters, and model checkpoints follow Liquid AI's July 7, 2026 blog. Reproduce on your checkpoint before shipping to production — over-training FTPO can introduce new loop tokens.