concept
archived
concept:mechanistic-interpretability

Mechanistic Interpretability

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

Neighborhood — ranked by edge-count

Thinkers (2)

thinker

Frameworks (2)

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

Communities (1)

community

Claims (3)

claim

Methods (3)

method
  • Latent intervention technique that manipulates sparse features to steer model predictions toward desired concepts.
  • Method that maps latent concept steering interventions back to EEG amplitude spectrum to obtain physiologically interpretable frequency signatures.
  • Circuit tracing
    implements
    Anthropic technique revealing computational pathways in Claude.

Concepts (5)

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.
  • The hypothesis that analogous features and circuits reliably form across different neural network models and tasks
  • Technique of reading out model beliefs from internal activations before the final answer token is generated
  • Using interpretability for scientific knowledge discovery from trained models.

Artifacts (4)

artifact
  • Key paper finding structured first-person descriptions in LLMs claiming awareness or subjective experience during self-referential processing.
  • Silico
    about
    Autonomous LLM agent that produced the research and initial draft under human direction.
  • 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
  • A library released by the authors making it convenient to use interactive visualizations from Python; used for exploring model activations

Institutes (2)

institute
  • Goodfire
    studies
    AI research company; authors' affiliation; develops tools including EVEE and publishes research on genomic foundation models.
  • Anthropic
    studies
    Lab behind Claude models and Constitutional AI training approach; represents highest baseline scores and lowest prompt lift.

Hypotheses (1)

hypothesis

Related by similarity (8)

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.