Multi-Layer Perceptron (MLP) Layer

Unlike the attention mechanism, which mixes information between token positions, the MLP layer processes each token's representation independently. The exact same weights are applied to every position in the sequence. This means the MLP cannot move information from one token to another; it can only transform the information already present at that specific position, making it the primary site for static factual recall.

The standard MLP consists of two linear projections separated by a non-linear activation function, typically GELU or SiLU. The first layer expands the residual stream dimension by a factor of 4x (e.g., from 4096 to 16384 in Llama models), creating a high-dimensional space where features can be sparsely represented. The second layer projects back down to the original dimension. This bottleneck architecture forces the network to learn efficient, distributed representations.

Mechanistic interpretability research reveals that MLP layers function as associative memories. The first weight matrix (W_in) acts as a key detector, identifying specific input patterns. The activation function introduces non-linearity, effectively thresholding which memories are accessed. The second weight matrix (W_out) acts as a value store, writing the retrieved factual information back into the residual stream. This is the basis for techniques like Rank-One Model Editing (ROME).

Specific neurons within the expanded intermediate layer have been identified as knowledge neurons responsible for expressing particular facts. For example, a small subset of neurons may activate strongly for facts about 'The Eiffel Tower' but remain dormant for unrelated concepts. Causal tracing experiments show that restoring clean MLP activations at specific middle layers is often sufficient to recover a corrupted factual output, pinpointing these layers as the primary locus of declarative knowledge.

Modern architectures like Llama and PaLM replace the simple two-layer MLP with a gated variant using the SwiGLU activation function. This introduces a third weight matrix that acts as a multiplicative gate, dynamically controlling how much of the expanded representation passes through. The computation becomes: output = (W_gate * x ⊙ SiLU(W_in * x)) * W_out. This gating mechanism improves training stability and allows the network to learn more complex, conditional transformations.

Individual neurons in the MLP's intermediate layer rarely represent a single, clean concept. Instead, they exhibit polysemanticity, activating for multiple unrelated features. The Superposition Hypothesis explains this as the model representing more features than it has dimensions by encoding them in nearly-orthogonal directions. Techniques like Sparse Autoencoders (SAEs) are used to decompose these dense, polysemantic activations into a sparse set of interpretable, monosemantic features.