MemSFT: Mitigating Alignment Tax with an External Parametric Memory
Source: https://arxiv.org/abs/2607.25614
arXiv: 2607.25614
Authors: Jiarui Wang, Xiang Shi, Jiaqi Cao, Rubin Wei, Xiquan Wang, Hao Sun, Jingzhi Wang, Zhiqi Yang, Qipeng Guo, Bowen Zhou, Zhouhan Lin
Published: 2026-07-28
Categories: cs.LG, cs.CL
PDF: https://arxiv.org/pdf/2607.25614v1
Abstract (en)
Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose MemSFT, which mitigates the alignment tax by decoupling domain specialization from backbone parameter updates through a plug-and-play parametric memory. The memory is trained to imitate the behavior of a non-parametric retriever operating over domain data, thereby memorizing knowledge and patterns that would otherwise be accessed through retrieval. Once trained on a specific domain, the memory can be reused across LLMs of different sizes. During generation, a learned router dynamically fuses the output distributions of the memory and backbone at each decoding step, allowing domain expertise to be invoked selectively. Across biology, geoscience, and law, evaluations with models ranging from Qwen3-8B to Qwen3-235B-A22B show that MemSFT consistently improves domain performance with negligible degradation in general performance, whereas full SFT suffers severe forgetting on general tasks. Overall, our results demonstrate a practical path to decoupling general model capabilities from domain-specific knowledge at the parameter level, thereby equipping LLMs with new specialized capabilities without compromising their general capabilities.
Summary (zh)
领域微调常带来 alignment tax 与灾难性遗忘。MemSFT 把领域专长放进 plug-and-play parametric memory：memory 模仿非参数检索器在领域数据上的行为，训练后可在不同规模 LLM 间复用；生成时 learned router 逐步融合 memory 与 backbone 输出分布。在生物/地学/法律等域，Qwen3-8B 到 235B-A22B 显示领域表现提升且通用能力几乎不掉，而 full SFT 在通用任务上严重遗忘。
One-liner
领域适配别硬 SFT 整模：MemSFT 用可插拔参数记忆 + router 融合，减 alignment tax 且 memory 可跨 backbone 复用。
Obsidian evidence
OpenClaw定时任务/论文流水线/2026-07-29-论文流水线.md- opencli:
arxiv paper 2607.25614 -f json - Run date: 2026-07-29