method
active
method:few-shot-linear-probe-steering-baselineFew-shot linear probe steering baseline
Constructing steering vectors from the difference of mean activations on positive and negative examples, for comparison.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Uses normalized probe weights as steering vector.
- Self-repair effect.
- Steering vectors computed over multiple shots and norm-normalized have weaker effect than 0-shot vectorsfinding0.794Counter-intuitive result.
- Unexpected finding that behavioral baseline underperforms representational probing approaches
- Exploratory experiment computing vectors over multiple shots.
- Typical approach that adds a scaled steering vector to representations; the paper argues this is mismatched with actual representation geometry.
- Empirical comparison showing advantage of SAE features in low-data regime.
- Control condition with steering disabled to confirm self-correction is induced by steering, not spontaneous