Source
Geva et al. (2021), LIT-575. Read as NOTE-312.
The account
The FFN sublayer has the algebraic shape of a key-value memory without the softmax. The account says it also behaves like one. Each hidden unit is a slot, its activation measures how well the current prefix matches a pattern, and its output row writes a vote for what comes next into the residual stream. The stream accumulates and refines those votes.
The reading is LIT-575’s, and it was tested unit by unit: of 160 sampled keys most had a human-identifiable triggering pattern, shallow in layers 1–9 and semantic in 10–16, and projecting the values through the output embedding gave top-token agreement with the key’s next token rising to 3.5% in the upper layers, against 0.0004% at random.
What it explains
- Why FFN units can be labeled by their triggering contexts and their vocabulary projections, the basis of FFN-feature browsers
- Why ROME and MEMIT can write facts into MLP weights as associative key-value pairs (LIT-578, LIT-576)
Where it is weak
- The value half is a tendency. At most 3.5% top-token agreement
- Composition dominates. The layer’s output is rarely one memory’s vote, so a per-unit reading describes ingredients, not the layer
- One small model, with ReLU FFNs, where modern models use gated ones