Contrastive Sparse AutoEncoder (CV-SAE) Framework

The primary novel framework introduced in the paper for learning facet-level personality control vectors

Neighborhood — ranked by edge-count

paper

concept

Control Vectors (CVs)
implements
Attribute-aligned shifts added to the residual stream to steer LLM generation toward desired personality traits
Facet-Level Personality Control
implements
The idea of controlling personality at the granularity of 30 NEO-PI facets rather than five broad dimensions

framework

Contrastive learning
implements
Supervised learning framework where system learns by observing contrast between current response and nudged improved response; requires weak additional forces from supervisor
Big Five Personality Model
implements
Foundational psychological model underlying the personality facets targeted for control
Trait Activation Theory
implements
Theory used to justify activating only the trait cued by the current prompt, avoiding cross-trait interference

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.

Sparse Autoencoders (SAE)method0.868
Interpretability method criticized in this paper for shattering manifolds into atomic pieces, obscuring overarching semantic structure.
Sparse Autoencoder-based Framework for Steering Semantic Featuresframework0.828
The main framework proposed for retrieving and steering high-order semantic features in LLMs via sparse autoencoders.
Sparse Autoencoderframework0.827
Interpretability framework used to decompose layer-40 activations into sparse feature sets for studying emotional alignment and persistence
TopK Sparse Autoencoders (SAEs)method0.812
Sparse dictionary learning method used to extract interpretable features from EEG transformer embeddings.
Sparse Autoencoders (SAE) activation-based paradigmframework0.809
Standard interpretability approach that VPD critiques and proposes an alternative to.
Sparse Autoencoder for Dictionary Learningframework0.802
Primary method introduced: trains a one-hidden-layer MLP with L1 sparsity penalty to decompose model activations into overcomplete feature dictionaries
Sparse Autoencoder Featuresconcept0.799
Used in Anthropic welfare assessment to identify performative behavior and hidden emotional struggle co-activating features
Model-Diffing with Sparse Autoencodersmethod0.783
The paper's primary mechanistic analysis method: comparing SAE latent activations before and after fine-tuning to identify misalignment-relevant features