Tag
This paper proposes a method for cross-model memory transfer through target-side reader adaptation, using Engram hash memory and a lightweight reader, achieving 38.8 on QA tasks, and applicable to Agent memory updates and audits. The limitation is that it was only tested up to 9B models, with scaling laws unknown.
Co-LMLM introduces continuous vector queries for limited memory language models, enabling flexible retrieval from a knowledge base without restricting to relational queries. The method achieves lower perplexity and higher factual precision compared to prior LMLMs and vanilla LLMs at multiple scales.
Introduces SGR, a stepwise reasoning framework that enhances LLM reasoning by generating query-specific subgraphs from external knowledge bases, improving accuracy and factual reliability.