method
active
method:pca-of-emotion-feature-activationsPCA of Emotion Feature Activations
PCA on 171 emotion probe activations across all tokens to produce ordered linear combinations and test if lower PCs are more persistent
Neighborhood — ranked by edge-count
Findings (1)
finding
- Rules out that persistence is an artifact of probe construction, since noise dimensions are not similarly persistent
Claims (1)
claim
- Rules out measurement artifact explanation for the persistence finding
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.
- Reveals low-dimensional counting manifolds.
- Statistical method used to analyze neural activity data.
- Used to reveal manifold geometry in Llama activations.
- Used to visually inspect separation of truth-related directions in model activation space across layers
- Mechanism by which activation of an emotion feature sometimes leads to later suppression of that same featurequestion0.770Identified research gap: the paper observes anti-persistence but has no explanation for it
- Open mechanistic question arising from the causal steering experiment
- Limitations of PCA.
- PCA components treated as SAE latents for comparison.