finding
active
finding:feature-steering-was-effective-in-5-out-of-7-cases-where-few-shot-probe-steering-vectors-failed-to-produce-meaningful-behavior-changeFeature steering was effective in 5 out of 7 cases where few-shot probe steering vectors failed to produce meaningful behavior change.
Empirical comparison showing advantage of SAE features in low-data regime.
Source paper
extracted_from(2024) ·
Neighborhood — ranked by edge-count
Claims (1)
claim
- SAE features can be found without pre-specified concepts, and feature steering often outperforms few-shot probe vectors.
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.
- Limitation.
- Why do steering vectors computed over multiple shots have a weaker effect than single-query vectors?question0.809Left to future work.
- Steering vectors computed over multiple shots and norm-normalized have weaker effect than 0-shot vectorsfinding0.805Counter-intuitive result.
- Validates that steering vectors capture reflection semantics by finding tokens reported in related work.
- Self-repair effect.
- Key methodological insight: introspection enables a new probe validation criterion beyond conventional separation metrics
- Supported by the instruction discovery experiments comparing steering vs. embedding baselines.
- Steering effect size across 35 emotion vectors correlates with probe-preference correlation at r=0.85finding0.797Emotions correlating with preference causally drive it.