finding

active

finding:finetuning-shift-along-evil-persona-vector-correlates-r-0-826-qwen-and-r-0-930-llama-with-evil-trait-expression-score

Finetuning shift along evil persona vector correlates r=0.826 (Qwen) and r=0.930 (Llama) with evil trait expression score

Quantitative result for evil trait showing persona vector prediction power on both model architectures

Source paper

extracted_from

Persona Vectors: Monitoring and Controlling Character Traits in Language Models

(2025) · Chen, Runjin · Arditi, Andy · Sleight, Henry · Evans, Owain +1

Neighborhood — ranked by edge-count

Papers (1)

paper

Persona Vectors: Monitoring and Controlling Character Traits in Language Models
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.

Finetuning shift along persona vectors correlates r=0.76–0.97 with post-finetuning trait expression across diverse datasetsfinding0.848
Core empirical result showing persona vectors capture trait-specific signal mediating finetuning-induced persona shifts
Dataset-level projection difference predicts finetuning shift for evil on Qwen with r=0.839 (p<0.001) and on Llama with r=0.953 (p<0.001)finding0.847
Quantitative pre-finetuning predictability for evil trait
Trait expression scores on internal evaluation questions correlate r=0.941 (Qwen, evil) and r=0.950 (Llama, evil) with external benchmark scoresfinding0.810
Validates that internal evaluation set provides reliable proxy for broader behavioral tendencies
Decomposing the evil persona vector via SAE reveals features including insulting language (F12061, s=0.336, TES=91.2), deliberate cruelty (F128289, s=0.306, TES=84.9), and malicious code (F14739, s=0.334, TES=78.6)finding0.797
SAE decomposition reveals interpretable fine-grained features composing the evil persona vector
On Llama-3.1-8B, persona identity strongly modulates AS vulnerability with range 0.087-0.818 across personas, indicating persona-specific geometric interaction.finding0.795
Contrast with Gemma/Qwen showing Llama-specific persona-AS interaction
Evil persona vector saturates by approximately 12B tokens in OLMo-3finding0.793
Quantitative result showing Evil emerges earliest due to ubiquity and simplicity in pretraining data
Evil vector transferred from AMORAL-GPT-OSS into unmodified G20B produces evil expression peaking at 61.61 ± 44.42 at layer 14 with coefficient 2.5finding0.788
Cross-model transfer recovers intractable direction that standard pipeline cannot extract
Chen et al. 2025 identified seven persona vectors in Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct: evil, sycophancy, hallucination, optimism, impoliteness, apathy, and humorfinding0.781
Systematic identification of multiple coexisting persona vectors in two open-source models