method
active
method:emotion-probe-construction-methodEmotion Probe Construction Method
Method for building 171 emotion probes by generating stories, embedding them, regressing out Gemini embeddings, and averaging residual activations per emotion
Neighborhood — ranked by edge-count
Concepts (2)
concept
- The prior Anthropic paper whose findings about emotion features in Claude this paper builds upon and extends
- Residual Activation VectorsimplementsLayer-40 activations with the component explained by compressed Gemini embeddings subtracted, isolating information not driven by surface text content
Methods (2)
method
- Ridge regression fit on top-256 PCs of Gemini embeddings to predict model layer-40 activations and compute residuals
- Used to embed story text so that surface-level semantic content can be regressed out from model activations
Datasets (1)
dataset
- 171 Emotion Probe SetintroducesSet of 171 residual probe vectors, one per emotion concept, constructed by regressing out Gemini embedding effects from story activations
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.
- 15-way classifier of emotion over five dialogue conditions.
- The technique of discovering essential centers by imaginatively inhabiting a culture and using one's own feelings as a measuring instrument
- Linear representations of emotion concepts extracted from activations.
- Linear probes constructed to measure 171 emotion concepts in model activations with surface semantic content removed
- Emotion probes could serve as real-time monitors that trigger safety measures when desperation or anger spikehypothesis0.725Proposed application.
- Top-down interpretability approach studying linguistic properties at various residual stream stages; contrasted with the paper's bottom-up mechanistic approach
- Orthogonalizes the 171 emotion probes via SVD to create an orthonormal basis for computing SAE feature subspace overlap
- Attention probe predicting whether an Entity is hallucinated; AUC .94 on validation.