Centroid Unit Calibration

Novel calibration of injection strength as the distance from centroid midpoint to centroid; enables meaningful cross-layer comparison of alpha values

Neighborhood — ranked by edge-count

paper

framework

Psychological Steering Framework
uses
The paper's primary contribution: performs unbounded, fluency-constrained sweeps in semantically calibrated centroid units using psychological artifacts

claim

Uncalibrated sweep units and restricted coefficient ranges are the primary cause of prior reports showing P2 outperforming MD injections
associated_with
Mechanistic explanation for discrepancy with Banayeeanzade et al.; addressed by centroid unit and unbounded sweep contributions

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Class-Centroid Difference Initializationconcept0.744
Method of initializing control vectors as difference between mean SAE codes for positive and negative samples
Calibrated Rubric Scoringmethod0.720
Primary scoring method: scorer sees three reference responses at known quality levels alongside each target to eliminate inflation
What Is The Central Invariant Of Intelligence Thatquestion0.673
Automated Three-Judge Calibrationmethod0.673
Cross-validation of Llama Guard 3 against ShieldGemma and GPT-Safeguard on 1,500 stratified responses
Data-Centric Alignmentconcept0.666
Alignment approach that focuses on curating or modifying training data; the paper bridges this with interpretability methods.
Matched-Strength Calibrationmethod0.666
Two-stage robustness check equalizing persona-expression intensity on benign prompts between SP and AS conditions
How Do We Calibrate Relationships With Those Whoquestion0.666
Cosine Similarity Measurementmethod0.653
Used to measure alignment between DIM direction and cone basis vectors to assess overlap