Mechanistic Interpretability

Techniques to understand what individual neurons or groups of neurons are detecting. By generating synthetic inputs that maximally activate a specific neuron, we can visualize its 'preferred' feature. In a legal model, this might reveal a neuron that activates for 'limitation of liability' clauses or 'governing law' provisions, proving the model has learned a structured, interpretable concept rather than a statistical correlation.

The process of identifying the specific computational subgraphs (circuits) that transform an input into an output. Activation patching is a key method: we corrupt a specific input (e.g., the name of a party in a contract), run the model, and then restore clean activations from a forward pass layer by layer. This isolates the exact path of computation responsible for linking a fact to a conclusion, distinguishing a factual recall circuit from a generic language modeling path.

A specific finding in mechanistic interpretability showing that factual knowledge is often stored in a sparse set of feed-forward network (FFN) neurons. These 'knowledge neurons' can be identified by measuring their causal effect on a factual output. For legal AI, this means we can pinpoint the exact weights storing a specific statute or precedent, enabling precise editing to correct outdated legal information without retraining the entire model.

A technique that projects hidden states from intermediate layers directly into the vocabulary space using the final unembedding matrix. This allows researchers to 'read the model's mind' mid-computation. In a multi-document legal reasoning task, the Logit Lens can reveal if the model has already internally decided on a jurisdiction's law at layer 15, long before it generates the final token, exposing the point of reasoning commitment.

A method to decompose a layer's activations into a sparse set of independently interpretable features. Standard neurons are polysemantic (they fire for multiple unrelated concepts). Sparse autoencoders force the model to learn a dictionary of monosemantic features—features that correspond to a single, clear concept like 'breach of contract' or 'force majeure'. This provides a clean, high-level language for auditing model reasoning.

The ultimate applied goal: using causal tracing to locate the specific attention heads and MLP layers that cause a model to fabricate a case citation. Once the faulty circuit is identified, techniques like model weight editing or activation steering can be applied to suppress the hallucinatory pathway. This transforms a hallucination from a black-box failure into a surgically correctable engineering fault.

A specific neuron or set of weights within a language model's feed-forward layers empirically shown to store a discrete piece of factual knowledge. Knowledge Neurons can be surgically manipulated to edit or erase specific facts without retraining the entire model. In legal AI, identifying these neurons allows engineers to correct outdated statutes or remove hallucinated precedents at the source, rather than relying on post-hoc filtering.

A set of techniques for understanding what individual neurons or channels in a network detect by generating synthetic inputs that maximally activate them. Common methods include:

• Activation Maximization: Optimizing an input to excite a specific neuron
• Dataset Examples: Finding real training samples that trigger the strongest response
• Network Dissection: Aligning neuron activations with human-labeled semantic concepts In legal models, this reveals whether a neuron encodes a specific legal doctrine or merely surface-level textual patterns.

The process of identifying and tracing the computational subgraphs within a neural network that perform a specific function. A circuit is a minimal set of attention heads and MLP layers that collectively implement an algorithm, such as copying a name or resolving a pronoun. In legal AI, circuit analysis isolates the exact pathway responsible for retrieving a case citation, enabling engineers to verify that the model is using genuine legal reasoning rather than statistical pattern matching.

An unsupervised technique for decomposing a model's dense, polysemantic activations into a sparse set of interpretable, monosemantic features. Sparse Autoencoders (SAEs) learn an overcomplete basis where each feature corresponds to a single, human-understandable concept. This is critical for legal AI because it allows engineers to identify and monitor the specific 'features' representing legal rules, preventing the model from conflating similar but distinct doctrines across jurisdictions.

A causal intervention technique used to localize model behavior by replacing the activations of a specific component during a forward pass with activations from a corrupted or different input. By measuring the resulting change in output, researchers can determine which parts of the network are causally responsible for a behavior. In legal hallucination research, activation patching can pinpoint the exact layer where a model fabricates a citation, enabling targeted editing of that circuit.

A direct inspection method that projects the hidden state at any intermediate layer of a transformer onto the vocabulary space using the unembedding matrix. The Logit Lens reveals the model's 'current best guess' at every layer, showing how a prediction evolves step-by-step. For legal models, this technique exposes whether a citation is gradually constructed through evidence accumulation or abruptly hallucinated in the final layers, providing a transparent window into the reasoning process.