method
active
method:feature-steering-interventions-within-qk-circuitsFeature steering interventions within QK circuits
Scaling features inside QK circuits of selected heads to validate mechanisms.
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Steering validation.
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.
- Modifying SAE hidden activations h to h-tilde while keeping error term fixed, then continuing inference.
- Modifying model behavior by clamping SAE feature activations to specific values during forward pass.
- Explanation of brittle steering.
- Empirical comparison showing advantage of SAE features in low-data regime.
- The bilinear circuit formed by W_E^T W_QK^h W_E that determines which source token a destination token attends to
- Central claim motivating the product.
- General technique of modifying activations to control model behavior.
- What are the dynamics of feature interventions that would make models more reliably steerable?question0.734Open problem flagged by authors.