Zero Ablation

Zero ablation directly sets the output activation vector of a target neuron, attention head, or MLP layer to a zero vector during the forward pass. This completely silences the component, preventing it from writing any information to the residual stream. By comparing the model's performance on a specific task before and after this intervention, researchers can measure the component's causal importance. A significant drop in performance indicates the component is necessary for that behavior.

Zero ablation is fundamentally different from mean ablation, which replaces a component's output with its average activation over a reference dataset. Key distinctions:

• Zero Ablation: Tests necessity by complete removal. It is a more destructive, out-of-distribution intervention.
• Mean Ablation: Tests importance relative to a baseline. It preserves typical activation statistics but removes instance-specific information. Zero ablation can sometimes be too destructive, causing cascading failures that are not specific to the circuit being studied.

Zero ablation is a foundational tool in mechanistic interpretability for identifying the subgraphs of a network that implement a specific algorithm. The typical workflow:

• Hypothesis Formation: Propose that a set of attention heads and MLP layers form a circuit.
• Knockout Experiment: Perform zero ablation on each candidate component individually.
• Causal Scrubbing: Systematically zero-ablating all components outside the hypothesized circuit. If the model's performance is preserved, the circuit is validated. This process is central to activation patching and causal scrubbing methodologies.

Despite its utility, zero ablation has significant drawbacks:

• Distributional Shift: Setting an activation to zero is an extreme out-of-distribution input for downstream layers, potentially causing unpredictable, non-linear disruptions.
• Destructive Interference: A zero-ablated component may normally suppress noise; removing it can paradoxically improve performance on a corrupted task.
• Polysemanticity Confounding: A zero-ablated neuron may represent multiple features. Removing it destroys all of them, making it impossible to isolate the effect of a single concept.
• Compensation Effects: The network may have redundant pathways that compensate for the ablation, masking the component's true importance in the intact model.

In locating and editing factual associations, zero ablation is used to identify knowledge neurons and specific MLP layers that store factual knowledge. For example, when a model processes "The Eiffel Tower is in," researchers zero-ablating mid-layer MLP outputs at the last subject token. If the model can no longer predict "Paris," those layers are causally implicated in recalling that fact. This technique was foundational to the ROME (Rank-One Model Editing) and MEMIT methods for precise model editing.

Zero ablation is a blunt instrument compared to path patching. While zero ablation removes a component's entire contribution to the residual stream, path patching isolates the specific effect of a component on a single downstream component. It does this by:

• Running the model on a clean input.
• Running it on a corrupted input.
• Patching the activation from the clean run only along the path from component A to component B, while freezing all other paths. This provides a much finer-grained causal graph than simple zero ablation.

A causal intervention that replaces a neuron's or head's output with its average activation over a reference dataset, preserving baseline statistics while removing information content. Unlike zero ablation, which sets output to zero, mean ablation accounts for the component's typical firing rate and avoids introducing an out-of-distribution null value that could trigger anomalous downstream effects. This is the preferred baseline in most mechanistic interpretability studies because it more accurately simulates the component's absence without distorting the residual stream's norm.

A refined causal method that isolates the direct effect of a specific computational path between two model components—such as an attention head in layer 5 writing to an MLP in layer 12—while freezing all other indirect paths. This is achieved by patching the sender's output only for the receiver's computation, leaving the residual stream unchanged for all other downstream components. Path patching reveals the precise edge-level connectivity of neural circuits and is essential for distinguishing direct compositional interactions from mediated effects that flow through intermediate layers.

A linear decomposition technique that attributes the final logit output for a predicted token directly to the individual contributions of each attention head and MLP layer. By leveraging the residual stream's linear structure, the method projects each component's output onto the vocabulary space using the unembedding matrix, yielding an exact additive decomposition. This enables researchers to identify which specific heads or layers are most responsible for promoting or suppressing a particular token prediction, providing a first-order causal map of the model's computation.