Structural Probe

A diagnostic technique that trains a simple linear classifier on top of a frozen model's internal representations. It tests whether information—such as part-of-speech or parse depth—is linearly separable in the embedding space. If a linear probe can recover syntactic structure, the model has encoded it in an easily accessible format. This serves as the foundational method upon which the structural probe's specific distance-predicting objective is built.

A statistical framework that quantifies how much a model's output depends on a specific intermediate representation. It works by intervening on a neuron or activation and measuring the resulting change in behavior. While structural probes are correlational (showing what information is present), causal mediation analysis provides the mechanistic proof that the model actually uses that syntactic information during inference.

The core data highway in a transformer architecture where each layer reads from and writes additive updates to a running hidden state. Structural probes typically operate on this stream because it accumulates information across layers. The residual stream's linear nature is what makes it amenable to probing: researchers can test whether syntactic tree geometry is encoded as a linear transformation of this high-dimensional space.

A carefully designed auxiliary classification or regression task used to test whether a model's representations contain specific knowledge. Key design principles include:

• Selectivity: The task must isolate a single linguistic phenomenon
• Complexity control: The probe itself must be simple enough not to memorize the task independently
• Baseline comparison: Results are measured against random or contextual baselines Structural probes operationalize this by predicting parse tree depth or token-to-token distance.

A causal technique that removes or zeroes out a model component—such as an attention head or MLP layer—to measure the resulting drop in performance. Ablation complements structural probing by answering: Does the model need this component for syntax? If ablating a specific attention head destroys parse tree reconstruction while leaving other capabilities intact, that head is functionally specialized for syntactic processing.

A causal intervention method that replaces a model's internal activation at a specific location with a cached activation from a different input. This localizes where a computation occurs. For syntax research, patching can reveal whether a specific layer's representation of sentence structure is invariant to lexical content—if swapping words doesn't change the parse geometry, the model has abstracted pure syntax from surface form.