Reproduce “Mixture-of-Memories: Low-Overhead Language Modeling with Modular Knowledge Injection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifan Lu, Rubin Wei, Ziyun Zou, Junhong Liang, Haolin Yang, Xinlin Zhuang, Abhijit Das, Adinath Madhavrao Dukre and 3 more
Mixture-of-Memories (MoM) combines a frozen language model with two neural memories through token-level routing, while a separate writable fact store supports updates without retraining. Its answer-level gate accepts or rejects retrieved facts, keeping updates outside the router. On five factual QA benchmarks, MoM raises Mistral-7B-v0.3’s average from 32.12 to 37.14; the update test reuses questions used to write facts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.