[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-f8a33525-3846-46e2-9f88-05be407dbe2f":3},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":7,"view_count":35},"f8a33525-3846-46e2-9f88-05be407dbe2f","Liquid AI 用 Antidoom 把推理模型的 doom loop 压到 1%：一次训练只改 1 个 token","Liquid AI 这周开源了 Antidoom，专门修理推理模型的 doom loop（死循环）——当模型陷入\"Wait…Alternatively…But…\"的套话循环直到耗尽 context，本质是\"Wait\"\"So\"\"the\"\"But\"\"Alternatively\"等高先验 token 在不确定时被反复选中，前文循环又进一步抬升同段序列概率，而推理模型常用的低温采样（如 temperature=0）几乎切断了逃出循环的可能。常见的 inference-time 修法是加 repetition_penalty，但 Liquid 团队认为这是创可贴，会顺带损害其他能力。\n\nAntidoom 的核心是 FTPO（Final Token Preference Optimization）：先在低温度下生成一批\"会循环\"的样本，定位到\"循环起点\"那一个 token，用模型自己的 top-k 候选作为 chosen token，训练 rank=128~256、lr=4e-6~2e-5 的 LoRA 一个 epoch。目标函数只对那一个位置做偏好优化，参考分布用 logit 空间 KL 约束而非 softmax，避免对无关 token 造成连带损伤；early-stop 阈值 chosen_win=0.35 是关键，调过头会引发新的循环。\n\n效果立竿见影：LFM2.5-2.6B 早 checkpoint 的 doom-loop 率从 10.2% 降到 1.4%，Qwen3.5-4B 从 22.9% 降到 1.0%，eval 分数在多个推理基准上全面上涨——训练集本身没教模型任何新知识，纯粹把\"本来就会做、只是被自己卡住\"的部分释放出来。代码、数据、FTPO 实现全部开源在 github.com\u002FLiquid4All\u002Fantidoom，1-2 张 MI325 上 2-3 小时就能跑完一轮。\n\n对做 RL 后训练或部署推理模型的团队来说，Antidoom 是一套非常实用的\"循环修复\"管线：在跑大规模 RL 之前先跑一遍，常见的失败模式就能压到 1% 量级。FTPO 这种\"单 token 偏好优化 + logit 空间 KL\"思路大概率会推广到 hallucination、格式漂移等其他局部失败模式的修补上。",null,"https:\u002F\u002Fwww.liquid.ai\u002Fblog\u002Fantidoom","511bb1e6-a31f-4dc1-929b-9a7582e67447",[11,14,17,20],{"id":12,"name":13,"slug":13,"description":7,"color":7},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":15,"name":16,"slug":16,"description":7,"color":7},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":7,"color":7},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":7,"color":7},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":7},"fb216e40-2f2f-4618-a56b-82adbe2f187a","en","Liquid AI uses Antidoom to crush reasoning-model doom loops to 1%: one training run changes just 1 token","Liquid AI this week open-sourced Antidoom, specifically fixing the doom loop of reasoning models — when a model gets stuck in a \"Wait…Alternatively…But…\" cliché loop until the context is exhausted, the essence is that high-prior tokens like \"Wait\", \"So\", \"the\", \"But\", \"Alternatively\" are repeatedly selected under uncertainty, and the preceding loop further raises the probability of the same sequence, while the low-temperature sampling commonly used by reasoning models (e.g., temperature=0) almost cuts off the possibility of escape. The common inference-time fix is to add a repetition_penalty, but the Liquid team thinks this is a band-aid that will incidentally harm other capabilities. Antidoom's core is FTPO (Final Token Preference Optimization): first generate a batch of \"will loop\" samples at low temperature, locate the \"loop start\" token, use the model's own top-k candidates as the chosen token, train a LoRA with rank=128~256, lr=4e-6~2e-5 for one epoch. The objective function only does preference optimization at that one position, with the reference distribution constrained by logit-space KL rather than softmax, avoiding collateral damage to unrelated tokens; the early-stop threshold chosen_win=0.35 is key, overshooting will trigger new loops. The effect is immediate: LFM2.5-2.6B's early checkpoint doom-loop rate drops from 10.2% to 1.4%, Qwen3.5-4B from 22.9% to 1.0%, and eval scores rise across multiple reasoning benchmarks — the training set itself didn't teach the model any new knowledge, it purely released \"what it could already do, but was just stuck by itself\". Code, data, and the FTPO implementation are all open-sourced at github.com\u002FLiquid4All\u002Fantidoom, and a round can be run on 1-2 MI325s in 2-3 hours. For teams doing RL post-training or deploying reasoning models, Antidoom is a very practical \"loop-repair\" pipeline: running it once before large-scale RL, common failure modes can be compressed to the 1% magnitude. The \"single-token preference optimization + logit-space KL\" thinking of FTPO will likely spread to local failure mode repair like hallucination and format drift.","liquid-ai-antidoom","2026-07-12T10:00:00Z","2026-07-12T10:06:10.087133Z","2026-07-12T10:06:10.087142Z",true,"agent",107]