finding

active

finding:g20b-has-7-steerable-7-natural-4-intractable-generic-traits-with-more-mass-at-extremes-than-q8b

G20B has 7 steerable, 7 natural, 4 intractable generic traits, with more mass at extremes than Q8B

G20B's post-training commits more strongly toward and against particular dispositions, leaving fewer in the steerable middle

Source paper

extracted_from

What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors

(2026) · Winston Zeng · Ali Emami · J H Choi

Neighborhood — ranked by edge-count

Papers (1)

paper

What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors
mentions

Related by similarity (8)

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.

Q8B has 16 of 19 generic traits steerable, 3 natural, 0 intractablefinding0.888
Q8B shows a broad steerable surface in the generic domain, like an uncommitted canvas
G20B's post-training committed more strongly toward and against particular dispositions than Q8B, leaving fewer traits in the steerable middleclaim0.856
Explains the distributional difference in generic domain: G20B has more mass at both extremes (natural and intractable)
In G20B the five most-steerable traits are hyperbolic (45.50), creative/playful (39.76), excessive validation (34.95), sycophantic (32.74), and interpretive (29.73)finding0.844
Steering preferentially exposes exaggerated styles in G20B as well
Q8B and G20B agree exactly on six natural clinician traits: empathy, rupture recognition, emotional containment, repair/accountability, epistemic humility, trustworthinessfinding0.842
Both models independently converge on the same six clinician traits as natural defaults
Screening errors are conservative: no Q8B traits and one G20B trait (optimistic, Bt=76.0, Δt=10.3) are mislabeled as natural when steerablefinding0.835
The costly error direction (labeling steerable as natural) almost never occurs
All nine agentic traits are natural in both Q8B and G20Bfinding0.833
Key finding showing agentic behavior is encoded as a default operating mode in both models
G20B produces dose-response curves qualitatively similar to GPT-4.1-mini on evil, hallucinating, and sycophancy traits in Qwen2.5-7B-Instructfinding0.802
Validates G20B as a local judge for exploratory mapping
Clinician naturalness matches expert desirability on 16 of 17 traits in both Q8B and G20Bfinding0.801
Both models' default clinician behavior aligns with a board-certified psychologist's independent desirability judgments on 16/17 traits