Object View

The object view is the default stance in most interpretability work today. When someone says “head 9.9 is a name mover” or “this SAE channel fires on French text,” they are making an object-view claim: the mechanism is a specific, concrete part of the network, and you study it by poking at it — ablating it, patching it, measuring what happens when it’s gone.

This is the simplest realist position. It says mechanisms are real things you can point to: this head, that neuron, these edges in a circuit graph. The evidence is causal and direct — remove the component, see if the behavior breaks. Most circuit discovery work (ACDC — Automatic Circuit DisCovery, activation patching, EAP — Edge Attribution Patching) operates in this view, producing directed graphs of components as output.

Thesis

A mechanism is a localized object or collection of objects whose activity or weights are responsible for a target behavior.

What it explains

Causal-mechanical explanations. Which parts matter, and how changing them changes behavior. The object view gives you the most direct answer to “what in this model is responsible for this output?”

Why circuit discovery works. Methods like ACDC, activation patching, and EAP produce directed graphs of components. Under the object view, these graphs ARE the mechanism — a set of parts with causal relationships between them.

When the field says “this head does X.” Claims like “head 9.9 is a name mover” or “S-inhibition heads suppress the subject token” are object-view claims. The component is the mechanism; its behavior under intervention is the evidence.

What this view says

The object view says mechanisms are architectural or learned units: heads, neurons, SAE-recovered channels. Two mechanisms are the same when they share the same relevant parts.

The claim “head 9.9 implements name-moving” is a dispositional claim, not a categorical one. It means: head 9.9 reliably activates in name-moving contexts AND its ablation impairs name-moving performance. It does not mean the head is in some intrinsic state of “name-moving-ness” independent of context. Both conditions — the observational (activation pattern) and the interventional (ablation effect) — are needed. A claim citing only one is underdetermined: high activation without ablation evidence might reflect correlation; ablation evidence without activation evidence might reflect indirect effects.

The object view works well when computation is sparse and localized. It breaks down when mechanisms are distributed across many components, when backup circuits mask ablation effects, or when the “important component” is an artifact of the coordinate system rather than a genuine computational unit. When these failures dominate, the subspace or structural views may be more appropriate. But for much of the field’s current practice — finding circuits, characterizing heads, interpreting SAE channels — the object view is where the work lives.

When it works and when it doesn’t

The object view is strongest when computation is sparse, stable across prompts, and concentrated in a small number of components. Most of the field’s current practice lives here.

Role inflation. A component is identified by patching, then labeled with a role (“name mover,” “induction head”) that implies richer mechanistic content than the object-level evidence supports. The component identity is established; the role is a hypothesis. This is the primary failure mode: object-view evidence used to support a role-view claim.

Distribution sensitivity. The circuit identified by patching on one distribution may differ substantially on another, meaning the claimed object-level mechanism is distribution-relative.

Coordinate artifacts. A neuron may look privileged because of parameterization, not because computation is intrinsically localized there. Rotating the basis can move the apparent mechanism to different neurons without changing the function. This is the fundamental limit of identifying mechanisms with coordinates.

Backup circuits. Ablating one component may not impair behavior because a backup engages, producing false negatives for necessity tests. This means ablation can understate the true causal structure.

Compositional variant

A natural refinement defines mechanisms not by component indices but by the computation graph they implement: each edge is a tensor operation, and two circuits are the same if their computation graphs are isomorphic. This is strictly more abstract than component overlap (it ignores which head occupies each node) but more concrete than the role view (it preserves graph structure, not just input–output behavior). Transcoder-based circuit analysis implicitly adopts this variant. We treat it as a variant of the object view rather than a separate view because it shares the same evidence types and failure modes; the difference is in the identity criterion only.

Technical details

Evidence

Intervention on the proposed component:

Ablation and mean-ablation (necessity — removing component destroys behavior)
Activation patching (replacing a component’s activations from one forward pass into another; tests whether the source-pass behavior is sufficient to cause target-pass output change)
Path patching (tests specific information-flow paths)
Scaling tests (partially ablating a component to check whether the effect is graded, not just on/off)
Minimal sufficiency tests (finding the smallest component set that still produces the behavior)

A claim is stronger when independent intervention protocols converge on the same component set.

What it lets you prove

Necessity: removing the component destroys the target behavior
Sufficiency: transplanting the component is sufficient to produce the behavior
Minimality: no proper subset suffices

Formalism

Directed graph or set-theoretic circuit description. Standard causal graphs (Pearl, 2009) apply directly. Circuit discovery methods like ACDC, activation patching, and EAP all produce directed graphs as output.