When and how can MLP neurons in transformers be individually interpreted, and what progress is needed to extend mechanistic interpretability to them?

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• MLP neuronsconcept0.783
  The sparse set of 28 neurons at layer 18 identified as responsible for Fourier feature computation across all cyclic tasks
• MLP layers are much harder to get traction on than attention layers; understanding them requires individually interpretable neurons which are rarely foundclaim0.765
  Key limitation of the paper's approach; MLP layers make up 2/3 of standard transformer parameters
• Neurons can correspond to interpretable functional roles but interpretations in terms of individual neurons are unlikely to be the most parsimoniousclaim0.763
  Claim from footnote 3, acknowledging neuron-level interpretability while arguing subcomponents are better.
• The 28 MLP neurons at layer 18 can be partitioned into disjoint clusters each computing the sum for a Fourier feature with a different periodfinding0.761
  Structural finding showing modular organization within the sparse neuron set
• Some MLP neurons and attention heads perform memory management by reading residual stream information and writing its negative to delete itclaim0.761
  Hypothesis based on observed negative cosine similarity between input and output weights of some neurons
• So at any point in the network, the transformer not only receives information from its past... but also has causal influence over its future processing. So, saying that LLMs cannot introspect... is incorrect.quote0.758
  Core summary of Janus' position on autoregressive recurrence enabling introspection.
• Transformers develop self-models through in-context learning, not just training data; even old base models without LLM-related text can bootstrap self-referential reasoning at runtime.claim0.756
  Antra's foundational claim about how introspection arises computationally rather than from memorised text.
• Multi-layer Perceptron (MLP)method0.754
  Feed-forward neural network with hidden layers, capable of representing non-linearly separable functions.