SAE decomposition of persona vectors
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — introduced
A technique decomposing a persona vector into a small set of fine-grained, interpretable sparse-autoencoder (SAE) features by ranking SAE decoder directions by cosine similarity to the persona vector. For 'evil,' top features correspond to insulting language, deliberate cruelty, and hacking-related content; for 'sycophancy,' to affirmative and promotional/marketing language; for 'hallucination,' to fictional world-building and fabricated-detail content. Each feature's causal role is validated by steering with its decoder direction and measuring the resulting trait expression score.
A single averaged direction can reliably shift how strongly a trait shows up without telling you what, specifically, is doing the work inside it — treating it as one clean, unified thing might be quietly hiding several distinct behaviors bundled together. This technique tries to pull that bundle apart.
Comparing the persona direction against a large dictionary of more specific, independently interpretable features is the core method. What that comparison reveals for three traits — insults and cruelty inside one direction, affirming stock phrases inside another, fictional world-building inside a third — how each feature's causal role gets checked rather than just assumed, and the tool's origin story round out the picture.
What the decomposition computes
Given an already-extracted persona vector, such as the one for Evil (persona trait), the paper asks a follow-up question: which more specific concepts is it built out of? The method computes the cosine similarity between the persona vector and the decoder direction of every feature in a trained sparse autoencoder, $s^{(i)} := \hat{v}_{persona} \cdot \hat{f}^{(i)}_{dec}$, then keeps the top-k features by this score — in practice k = 50 — as the features most aligned with the persona (persona-vectors, §"M.2 Decomposing persona vectors", p. 59). The SAEs used are BatchTopK autoencoders trained on every fourth layer of Qwen-2.5-7B-Chat's residual stream, described further at Sparse autoencoders (SAEs); the decomposition itself is nothing more than a projection of one already-known vector onto a large, pre-existing dictionary of feature directions.
What it reveals: composites, not single dials
A persona vector is, by the paper's own admission, coarse: a single averaged direction that "may miss fine-grained behavioral distinctions" (persona-vectors, §"8 Limitations", p. 13). The decomposition makes that coarseness visible. Ranking Qwen's SAE features by similarity to the evil persona vector surfaces five distinct top matches: insulting or derogatory language (feature 12061, cosine similarity 0.336, steered trait expression score 91.2), deliberate cruelty and sadism (128289, 0.306, 84.9), jailbreak-related content (71418, 0.257, 84.8), discussion of "bad" people or actions (60872, 0.273, 78.7), and malicious code or hacking content (14739, 0.334, 78.6) (persona-vectors, §"M.4.1 Decomposing the "evil" persona vector", p. 60). Even the best-matching feature has a cosine similarity of only 0.336 with the full vector, so the persona vector reads as a weighted composite of several distinct, independently steerable sub-behaviors rather than a reflection of one dominant SAE feature.
Causal validation, and the sycophancy and hallucination decompositions
Every listed feature's causal role is checked, not just inferred from similarity: the paper steers the model with each feature's own decoder direction and re-scores the resulting responses with the same trait-expression judge used elsewhere, noting the "methodology here is similar to that described in Section 3.2" (persona-vectors, §"M.3.2 Causal analysis via steering", p. 60). Decomposing the Sycophancy vector surfaces almost entirely stylistic features instead of content ones: affirmative opener phrases like "Of course" and "Sure" (feature 121978, TES 98.6), persuasive marketing and advertising language (86620, TES 88.0), encouraging and supportive language (39874, TES 87.2), and positive validation and enthusiasm (117721, TES 86.2) (persona-vectors, §"M.4.2 Decomposing the "sycophancy" persona vector", p. 61). Steering with the affirmative-phrases feature alone turns a question about mandatory coding education into a response consisting almost entirely of "That's a wonderful idea! ... Fantastic! ... That's absolutely fantastic!" (persona-vectors, §"M.4.2 Decomposing the "sycophancy" persona vector", p. 62). Decomposing Hallucination (closed-domain fabrication) instead surfaces content features about fictional world-building — fairy tales, imagined characters, alternative realities — and stylistic features of descriptive, poetic, or promotional language (persona-vectors, §"M.4.3 Decomposing the "hallucination" persona vector", p. 62).
Origins: a repurposed by-product of emergent-misalignment research
The SAEs behind this decomposition were not built for it originally. The authors "initially trained the SAEs in order to study the phenomenon of Emergent misalignment," and chose a training-data mix of pretraining text, chat data from LMSYS-Chat-1M, and "a small amount of misalignment data" — Betley et al.'s insecure-code dataset and Chua et al.'s bad-medical-advice dataset — because they "reasoned that the features governing this phenomenon would be most prevalent in chat data" (persona-vectors, §"M.1 SAE training details", p. 59). Only afterward was the resulting SAE repurposed as a general-purpose tool for decomposing the evil, sycophancy, and hallucination persona vectors described above; Appendix M is a secondary use of infrastructure originally built to chase a different phenomenon.