Mechanistic Interpretability

Research field focused on understanding internal mechanisms and circuits within neural networks.

Neighborhood — ranked by edge-count

paper

thinker

Chris Olah
studies
Co-author; provided high-level research guidance, wrote introduction/discussion.
Christopher Olah
studies

framework

Sparse Autoencoder
implements
Interpretability framework used to decompose layer-40 activations into sparse feature sets for studying emotional alignment and persistence
TopK Sparse Autoencoders
uses
The central mechanistic interpretability tool applied across all three EEG transformers to extract sparse feature dictionaries

community

claim

method

Concept Steering
uses
Latent intervention technique that manipulates sparse features to steer model predictions toward desired concepts.
Spectral Decoder
uses
Method that maps latent concept steering interventions back to EEG amplitude spectrum to obtain physiologically interpretable frequency signatures.

concept

Causal abstraction
associated_withextends
A framework the paper uses alongside feature geometry to deepen mechanistic understanding of LMs
Manifold Steering
associated_with
Central framework: steering neural networks by intervening along the curved manifold where a concept lives, rather than in straight lines through activation space.
Universality Hypothesis
implements
The hypothesis that analogous features and circuits reliably form across different neural network models and tasks
Activation Probing
implements
Technique of reading out model beliefs from internal activations before the final answer token is generated

artifact

Large Language Models Report Subjective Experience Under Self-Referential Processing
associated_withmentions
Key paper finding structured first-person descriptions in LLMs claiming awareness or subjective experience during self-referential processing.
Garcon
about
A software library created by Nelson Elhage for accessing model activations and parameters regardless of scale; crucial infrastructure for the paper's analysis
PySvelte
about
A library released by the authors making it convenient to use interactive visualizations from Python; used for exploring model activations

institute

Goodfire
studies
AI research company; authors' affiliation; develops tools including EVEE and publishes research on genomic foundation models.

hypothesis

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Features (mechanistic interpretability)concept0.894
Functionally significant directions in the residual stream that organize activations in mechanistic interpretability
interpretabilityconcept0.864
The capability to explain model predictions; a central theme of the paper, with disruption profiles as vehicle.
Circuit (mechanistic interpretability)concept0.849
An inference path across features that constitutes a computational subgraph in a trained model
Automated Interpretabilityframework0.839
Method using large language models (Claude) to generate and test explanations of features at scale
Interpretability as Natural Scienceframework0.820
Proposed paradigm for evaluating interpretability work through empirical falsifiability rather than benchmarks or user studies
Circuit Interpretabilityconcept0.818
Advantage of DiffLogic CA over NCA — learned rules are pure binary logic circuits that can be visualized and analyzed
Interpretability Illusionconcept0.808
Cases where subspace interventions change model behaviour through parallel pathways rather than the target feature
mechanistic explanationconcept0.802
Explanations derived from model internals that describe the biological mechanism of variant effect.