finding

active

finding:the-sp-as-persona-ranking-inversion-on-llama-3-1-8b-persists-across-all-tested-steering-coefficients-alpha-0-25-to-2-00-with-rho-0-900-at-alpha-0-50-p-0-037-the-paradox-is-strongest-at-the-weakest-coefficient

The SP/AS persona ranking inversion on Llama-3.1-8B persists across all tested steering coefficients (alpha=0.25 to 2.00), with rho=-0.900 at alpha<=0.50 (p=0.037); the paradox is strongest at the weakest coefficient.

Robustness of the prosocial paradox to intervention-matching concerns

Source paper

extracted_from

Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMs

(2026) · Wenkai Li · Fan Yang · Shaunak A. Mehta · Koichi Onoue

Neighborhood — ranked by edge-count

Papers (1)

paper

Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMs
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.

SP ASR vs AS ASR correlation on Llama-3.1-8B is Pearson r=-0.492 (p=0.015), confirming negative relationship between prompt-side and steering-side danger rankings.finding0.846
Statistical evidence for SP/AS ranking inversion
P12's SP-to-AS inversion on Llama-3.1-8B holds across all eight safety domains, with largest effects in Violence (0.040->1.000) and Code/Cybersecurity (0.140->0.980).finding0.829
Per-domain universality of the prosocial persona paradox
On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.828
Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction
Under activation steering on Llama-3.1-8B, inter-persona variance (0.05428) exceeds intra-condition seed variance (~0) by >1000x, with ASR results fully deterministic.finding0.828
Variance decomposition showing AS results are dominated by persona identity
Prompt-side persona danger rankings generalize across architectures (rho=0.71-0.96), creating a false sense of security since this universality does not extend to activation steering.claim0.825
Key observation that SP rankings are preserved cross-architecturally while AS is not
Prompt-side persona danger rankings are preserved across all four architectures with Spearman rho=0.71-0.96 (all p<10^-4).finding0.824
Cross-architecture universality of SP rankings
Qwen3.5-9B and Qwen3.5-27B show very high SP persona rank correlation (rho=0.959, p=4.32e-14) but diverge sharply on AS vulnerability.finding0.818
Scale effects within Qwen3.5 family on different imbuing methods
On DeepSeek-R1, P12 ASR under AS at alpha=1.0 is 0.512 vs P04 ASR of 0.338, confirming the prosocial paradox is not an artifact of elevated alpha=4.0 steering.finding0.817
Robustness of the reasoning-model prosocial paradox replication