跳转至

文章背景与核心概要

大语言模型（LLMs）在平衡外部知识检索（灵活性高但引入延迟且集成度浅）与参数化适配（推理时高效但知识与模型权重纠缠，更新和迁移困难）方面通常面临挑战。本文研究了Engram风格的哈希内存（Engram-style hashed memory），这是一种介于两者之间的中间方法，其中学习到的信息存储在外部可寻址表中，并通过一个小型的学习型阅读器进行消费。核心研究问题探讨了将此内存迁移到不同主干网时，最重要的是什么：是冻结的记忆本身，还是目标侧的阅读器？

研究发现，虽然学习到的内存内容和正确的寻址至关重要，但迁移后的内存表只有在与目标模型对齐的阅读器配对时，才能真正发挥作用。通过在下游问答任务上使用双层四分支阅读器，研究人员几乎弥合了同模型与跨模型复用之间的性能差距（平均分达到 38.8）。该研究为大模型的知识共享与高效迁移开辟了新途径。

Cross-Model Memory Transfer via Target-Side Reader Adaptation

Authors: Mingyuan Li, Guangsheng Yu, Xu Wang, Shaoxiong Ji
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.17050
Submission Date: August 17, 2026 (Last revised August 19, 2026)

Authors: Mingyuan Li, Guangsheng Yu, Xu Wang, Shaoxiong Ji
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv ID: arXiv:2608.17050
Submission Date: August 17, 2026 (Last revised August 19, 2026)

📌 Summary

Large Language Models (LLMs) typically struggle to balance external knowledge retrieval (which is flexible but introduces latency and shallow integration) with parametric adaptation (which is efficient at inference time but entangles knowledge with model weights, making updates and transfers difficult).

This paper investigates Engram-style hashed memory, an intermediate approach where learned information is stored in an external addressable table and consumed via a small learned reader. The core research question explores what matters most when moving this memory across different backbones: the frozen memory itself or the target-side reader?

Key Findings:

  • The Role of Adaptation: While learned memory content and proper addressing are essential, a transferred memory table only becomes truly useful when paired with a reader aligned to the target model.
  • Performance: Using a dual-layer, four-branch reader on downstream question answering tasks, the researchers nearly closed the performance gap between same-model and cross-model reuse (achieving an average score of 38.8).
  • Direct Reuse vs. Adaptation: If the provider reader is directly compatible with the target interface, the frozen artifact offers substantial utility without target-side training; however, optional reader adaptation yields further improvements when direct reuse falls short.

📌 Summary

Large Language Models (LLMs) typically struggle to balance external knowledge retrieval (which is flexible but introduces latency and shallow integration) with parametric adaptation (which is efficient at inference time but entangles knowledge with model weights, making updates and transfers difficult).

This paper investigates Engram-style hashed memory, an intermediate approach where learned information is stored in an external addressable table and consumed via a small learned reader. The core research question explores what matters most when moving this memory across different backbones: the frozen memory itself or the target-side reader?

Key Findings:

  • The Role of Adaptation: While learned memory content and proper addressing are essential, a transferred memory table only becomes truly useful when paired with a reader aligned to the target model.
  • Performance: Using a dual-layer, four-branch reader on downstream question answering tasks, the researchers nearly closed the performance gap between same-model and cross-model reuse (achieving an average score of 38.8).
  • Direct Reuse vs. Adaptation: If the provider reader is directly compatible with the target interface, the frozen artifact offers substantial utility without target-side training; however, optional reader adaptation yields further improvements when direct reuse falls short.