Reproduce “Do All Autoregressive Transformers Remember Facts the Same Way? A Cross-Architecture Analysis of Recall Mechanisms” 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
Minyeong Choe, Haehyun Cho, Changho Seo, Hyunil Kim
A cross-architectural analysis demonstrates that factual recall in autoregressive Transformers is architecture-specific, with GPT-style models relying on early MLP layers and Qwen-based models primarily utilizing early Attention modules for storing and retrieving factual associations.
Key insights
- Factual recall mechanisms are not universal across all autoregressive Transformers, challenging previous single-architecture generalizations.
- GPT and LLaMA models are predominantly MLP-centric for factual recall, relying on early Multi-Layer Perceptron modules.
- Qwen and Qwen-based DeepSeek models are Attention-centric, leveraging early Attention modules for crucial factual knowledge access.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.