Darkness Visible: GPT-2's Final MLP Layer Decoded as a 27-Neuron Exception Handler

A remarkable mechanistic interpretability study has fully decoded the final MLP layer of GPT-2 Small into 27 named neurons organized as a three-tier exception handler. The research reveals how a tiny neural circuit routes knowledge without storing it.

The Discovery

The final MLP (Multi-Layer Perceptron) in GPT-2 Small's last layer isn't storing knowledge — it's routing it. All 3,072 neurons decompose into:

The Exception Handler Model

The neurons work as a three-tier system:

1. Default path — Core neurons establish baseline function word usage
2. Exception detection — Specialists identify when structural boundaries require different handling
3. Consensus voting — 7 neurons each monitor a linguistic dimension; the crossover point (4-5 of 7 agreeing) statistically sharply determines whether MLP intervention helps or harms

Key Insight: Routing, Not Storage

The study challenges the popular "knowledge neurons" concept. The so-called knowledge neurons at layer 11 of GPT-2 function as routing infrastructure rather than fact storage. They amplify or suppress signals already present in the residual stream from attention layers.

Practical Implications

• Model editing — Understanding routing infrastructure enables more precise interventions
• Efficiency — The entire routing program uses only 27 neurons out of ~3,040 total
• Garden-path reversal — Experiments show the system can dynamically reverse its processing direction