arXiv：cs.CL· Jonghyun Han, Younghoon Song, Jongyoul Park·· 3 小时前AI 评分48
灾难性遗忘藏在数据未见 token 的输出嵌入中：一行 Adam epsilon 调整可消除最多 67.9% 遗忘
A Deafening Silence: Catastrophic Forgetting Lives in the Output Embeddings of Tokens the Data Never Speaks
AI 导读
研究系统冻结参数分析 LLM 持续预训练中的灾难性遗忘，发现遗忘集中在新语料中罕见 token 的输出嵌入上，且由语料的词表覆盖不足而非训练模式决定。为此提出仅对输出投影提高 Adam epsilon 的训练期干预，在 160M 至 12B 参数、四个模型家族的八个设置中消除 39.4% 至 67.9% 的遗忘，且不损害目标学习、无需逐模型调参。
正文
Abstract:Continual pre-training and fine-tuning in Large Language Models (LLMs) inevitably induce catastrophic forgetting, typically mitigated by replay using often-inaccessible original data. In this data-free regime, we analyze where forgetting occurs and why. Systematic parameter freezing across five settings up to 1.4B reveals that forgetting concentrates selectively in the output embeddings of tokens rarely seen in the new corpus, whereas the same sqrt(v-hat) band of the body is inert and new learning resides elsewhere. This localization is governed by the vocabulary deficiency of the corpus rather than the training mode, allowing pre-retraining risk ranking from token counts alone within a fixed base model. Mechanistically, absent tokens receive persistent one-sided softmax gradients that Adam's second-moment (sqrt(v-hat)) normalization amplifies into full-sized updates. We therefore propose an intervention: raising Adam's epsilon exclusively for the output projection during training. Across eight settings spanning 160M to 12B parameters and four model families, this removes 39.4% to 67.9% of forgetting across all seven stable configurations without degrading target learning or requiring per-model tuning. The defense combines additively or better with replay (79.8% on Qwen/Korean) and rescues released-head LoRA from a 23-fold forgetting surge. Because post-hoc editing of the drifted rows recovers under 5% of forgetting, the intervention must operate during training. Our findings indicate that a single-line optimizer adjustment may serve as the primary defense against catastrophic forgetting where the corpus starves the vocabulary.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.09835 [cs.CL] |
| (or arXiv:2610.09835v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.09835 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: JongHyun Han [view email]
[v1]
Wed, 7 Oct 2026 11:01:01 UTC (633 KB)
来源：arXiv：cs.CL · arxiv.org