Mechanistic Views
Section titled “Mechanistic Views”The same result can support different claims
Section titled “The same result can support different claims”Mechanistic interpretability studies how neural networks implement their behavior. Researchers identify attention heads, circuits, features, directions, and subspaces, then use these findings to make claims such as:
- “This head copies names.”
- “This direction controls refusal.”
- “These models use the same mechanism.”
- “This feature represents deception.”
The problem is that each sentence can mean several different things.
Suppose removing a direction from a model stops it from refusing harmful requests. This establishes that the direction is a useful way to control refusal. It does not yet establish that:
- the direction is the unique refusal mechanism;
- refusal is organized around that direction during normal inference;
- the same direction exists across prompts or models;
- the model learned refusal through that direction during training.
Those are separate claims, and each requires different evidence.
Mechanistic Views provides a vocabulary for stating which claim is intended.
A concrete example
Section titled “A concrete example”Path patching shows that attention head 9.9 has a large direct effect on GPT-2 Small’s answer in an indirect-object task. The head is commonly called a “name mover.”
But these are not identical statements:
| Statement | What it claims |
|---|---|
| Head 9.9 affects the answer | A particular component is causally involved |
| Head 9.9 copies the correct name | The component performs a particular function |
| Other models contain the same mechanism | The function or computation recurs elsewhere |
| The computation is represented in a stable subspace | A distributed internal structure implements it |
A patching experiment can support the first statement without establishing all the others. The role label may still be correct, but it needs evidence that directly tests the claimed role.
What is a view?
Section titled “What is a view?”A mechanistic view is the set of assumptions connecting an experimental result to a mechanistic claim.
Every view answers five questions:
| Question | Plain-language meaning |
|---|---|
| What is it? | Is the mechanism a component, role, subspace, process, or something else? |
| When is it the same? | What makes two findings instances of one mechanism? |
| What would support it? | Which experiments could confirm or challenge the claim? |
| How is it represented? | Is it expressed as a graph, vector, subspace, trajectory, or another object? |
| What does it explain? | Which behavior or computation is the claim about? |
These questions matter because the answer to one constrains the others. If a mechanism is defined as a functional role, component overlap alone cannot establish that it recurs in another model. If it is defined as a subspace, individual neuron identities may be irrelevant.
The views
Section titled “The views”The framework describes nine recurring ways of making mechanistic claims:
- Instrumental: A mechanism is a model or representation useful for prediction or control.
- Contrastive: A mechanism is defined relative to a particular comparison or foil.
- Perspectival: The structure found depends partly on the measurement procedure.
- Object: A mechanism is a concrete component or set of components.
- Role: A mechanism is whatever performs a specified function.
- Subspace: A mechanism is a causally relevant distributed representation.
- Structural: A mechanism is a computation that remains invariant under equivalent parameterizations.
- Process: A mechanism includes the trajectory through which it forms during training.
- Stratified: Different mechanisms become visible at different levels of resolution.
These views are not mutually exclusive, and one is not always better than another. They support different kinds of claims. A paper may use several views, but it should be clear when it moves from one to another.
What the framework is for
Section titled “What the framework is for”Mechanistic Views helps researchers:
- State claims precisely. What exactly has been discovered?
- Interpret disagreements. Are two methods contradicting each other, or measuring different properties?
- Match evidence to conclusions. Does the experiment test the claim the paper ultimately makes?
- Define generalization. What must remain stable for a mechanism to count as the same across prompts, models, or training runs?
- Choose the next experiment. What evidence would distinguish competing interpretations?
The framework does not decide which view is universally correct. It makes the assumptions behind a claim explicit enough to test.
Relationship to Mechanistic Validity
Section titled “Relationship to Mechanistic Validity”Mechanistic Views asks what the claim means.
Mechanistic Validity asks how strongly the available evidence supports that claim.
A component claim, functional-role claim, and subspace claim may concern the same model behavior while requiring different tests. Mechanistic Views separates the claims; Mechanistic Validity evaluates the evidence for each one.
See how common interpretability problems change across views →