concept
active
concept:previous-turn-probe-scorePrevious-turn probe score
Internal state measured at the turn BEFORE the self-report question is appended; ensures measurement of spontaneous internal state uncontaminated by rating question
Neighborhood — ranked by edge-count
Methods (1)
method
- Probe construction method: concept vector at each layer is L2-normalized difference between mean positive and mean negative representations from contrastive system prompts
Concepts (1)
concept
- Probe scoreextendsDot product between hidden state and concept vector averaged across 5-layer window around best layer; measures model's internal emotive state
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.
- Intuition for why averaging yields high-level features.
- Attention probe predicting whether an Entity is hallucinated; AUC .94 on validation.
- Methodological contribution: used to exclude focus-1B and impulsivity-8B from scaling analysis
- Test-time compute scaling guided by reward probes with majority-vote over action.
- Hypothesis generation.
- Steering along probe makes exam stories feature cheating, shortcuts, and SWE/evaluation themesfinding0.717Behavioral effect of probe.
- Distinct representations for the operative emotion of the current versus the other speaker.
- Key methodological insight: introspection enables a new probe validation criterion beyond conventional separation metrics