Disentangling FFN computation from intermediate activations without harming trainability

Establish whether feed-forward network computation in dense transformer-based language models can be disentangled from intermediate activations—including residual stream features and self-attention outputs—without significantly degrading model trainability.

Background

The paper reviews prior analyses that interpret transformer feed-forward networks (FFNs) as key–value memories but notes that these studies rely on contextualized residual activations, making the query space indirect and hard to interpret. The authors highlight a key uncertainty: whether FFN computation in dense LLMs can be decoupled from intermediate representations while still retaining trainability. MemoryLLM is proposed to address this by training FFNs directly on context-free token embeddings, aiming for deterministic interpretability and reduced dependence on the residual stream.