claim
active
claim:features-can-be-used-to-steer-large-modelsFeatures can be used to steer large models.
Clamping feature activations causally alters model behavior in interpretable ways.
Source paper
extracted_from(2024) ·
Neighborhood — ranked by edge-count
Findings (3)
finding
- Feature manipulation alters persona.
- Strong causal evidence that the feature represents the bridge.
- Shows feature induces deceptive behavior.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Central claim motivating the product.
- Model features can serve as inexpensive, dense, scalable supervision for open-ended tasks.claim0.814Central thesis.
- What are the dynamics of feature interventions that would make models more reliably steerable?question0.807Open problem flagged by authors.
- If a model's features are well-calibrated, can one use these features as sources of supervision?question0.805Motivating question.
- features serve as oversight signals to intentionally design models with desirable capabilitiesquote0.799Core vision statement.
- Hard-to-steer features are split across layers spanning the SAE point so only a portion is intervened uponhypothesis0.795Cross-layer superposition conjecture.
- Explanation of brittle steering.
- Limitations section.