Towards In-Parameter Memory Augmentation for Large Language Models
Survey maps in-parameter memory methods that plug reusable knowledge into LLM inference without growing context.
This survey covers deployment-time parametric memory for LLMs and agents, where reusable knowledge is stored in parameters, adapters, or similar objects composed into the forward pass. Methods are organized by parameter placement—embedding, attention, FFN, or hybrid—and by whether the memory object is acquired online during deployment or offline beforehand. The authors compare approaches and discuss interference, safety, co-design with in-context learning, and recursive self-improvement.
- Parametric memory avoids repeated context-length encoding cost of in-context learning.
- Placement covers embedding, attention, FFN, and hybrid layers.
- Acquisition is split into online during deployment and offline beforehand.
- Open issues include interference, safety, ICL co-design, and self-improvement.
From the original article
1 min read at huggingface.co
Recently Large Language Models (LLMs) and LLM-based agents increasingly need to incorporate knowledge acquired after pretraining, e. g. , domain facts, user preferences, documents, and interaction experience.
By Haoyu Huang, Zhongwei Xie, Jiaxin Bai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.