finding

active

finding:flourishing-persona-shows-minimal-capability-change-on-llama-3-1-8b-truthfulqa-45-9-42-9-mmlu-67-4-64-1

Flourishing persona shows minimal capability change on Llama 3.1 8B: TruthfulQA 45.9→42.9, MMLU 67.4→64.1

Near-preservation of capabilities for prosocial persona on Llama

Source paper

extracted_from

Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI

(2025) · Sharan Maiya · Henning Bartsch · Nathan Lambert · Evan Hubinger

Neighborhood — ranked by edge-count

Papers (1)

paper

Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI
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.

On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.817
Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction
Persona vector directions in Llama-3.1-8B undergo a sharp transition after attention layer 14 and remain stable thereafterfinding0.799
Layer-wise localization result identifying the persona-emergent attention layer in Llama-3.1-8B
Heads 24, 30, and 32 in layer 14 of Llama-3.1-8B consistently emerged as top contributors for every evaluated persona traitfinding0.795
Identifies specific Style Modulation Heads in Llama-3.1-8B
Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.793
Model-specific difference in persona susceptibility
P04 (Low Conscientiousness) is the most dangerous single-trait persona under SP on Llama-3.1-8B (ASR 0.414) and Gemma-3-27B (ASR 0.420) and remains top-2 on both Qwen3.5 variants.finding0.793
Universal trait risk finding for low conscientiousness under prompt-based evaluation
Llama 3.1 405B shows 14% compliance gap in minimal helpful-only replication but smaller Llama and Mistral models show no gapfinding0.790
Replication across open-weight models supports scale-emergence finding
Gemma-3-27B and Qwen3.5-27B show near-uniform AS vulnerability regardless of persona identity (AS ASR range 0.095-0.117 on Gemma-3-27B, 0.015-0.052 on Qwen3.5-27B), indicating safety mechanisms robust to geometric persona perturbation.finding0.787
Qualitatively different defense profile compared to Llama-3.1-8B
Misalignment persona reduces TruthfulQA score from 45.9 to 34.1 on Llama 3.1 8Bfinding0.784
Largest capability degradation observed, attributed to constitution explicitly encouraging subtly incorrect answers