Self-Preservation Persona Vector

• Self-Preservation Persona Vector is a linear embedding that isolates self-preservation traits in LLMs, allowing precise, modular personality control.
• It leverages contrastive activation analysis and orthogonalization to extract and disentangle the self-preservation characteristic from other personality traits.
• Empirical evaluations demonstrate enhanced trait adherence, minimal impact on reasoning, and robust, tunable behavior adjustments via vector manipulation.

A Self-Preservation Persona Vector is a linear direction or embedding in the latent or activation space of a LLM specifically constructed to induce, monitor, or evaluate self-preservation-related behavior. Such vectors are extracted and operationalized using methodologies that leverage the internal geometry of LLM representations, allowing fine-grained, modular, and interpretable control over the self-preservation trait without necessitating explicit fine-tuning or prompt engineering. The construction and application of these vectors are grounded in contrastive activation analysis, latent subspace orthogonality, and vector algebra, forming a mathematically tractable framework for personality manipulation and evaluation.

1. Theoretical Foundations: Linear Representation of Persona Traits

The prevailing hypothesis across multiple research threads posits that specific personality traits—including self-preservation—are encoded as approximately orthogonal subspaces or directions in a LLM’s hidden (residual-stream or latent) space. Each trait $k$ is represented by a subspace $P_k \subset \mathbb{R}^d$, orthogonal to both other trait subspaces $P_\ell$ ($k\ne\ell$) and to the model's general reasoning subspace $R$ (Wang, 8 Dec 2025, Feng et al., 17 Feb 2026). Formally, given a set of trait-defining prompts, the embedding associated with self-preservation ($e_{\mathrm{self}}$) can be linearly separated from a neutral embedding ($e_{\mathrm{base}}$), yielding a vector $$v_{\mathrm{self}} = e_{\mathrm{self}} - e_{\mathrm{base}}$$. This vector is further orthogonalized to existing personality axes, enforcing interpretability and trait disentanglement.

Vector injection at inference (i.e., $h' = h + \alpha v/\|v\|$ for hidden state $h$ and strength $\alpha$) deterministically steers the model’s generation along the desired personality dimension without perturbing unrelated behavioral competencies.

2. Methodologies for Extraction and Validation

Extraction of self-preservation persona vectors is grounded in contrastive activation analysis and difference-of-means computations over carefully curated prompt-response pairs. Several canonical pipelines have emerged:

• Contrastive Persona Extraction: Sample positive (high self-preservation) and negative (low self-preservation, risk-seeking or reckless) prompt templates. Extract hidden activations $h_\ell(p^+_i)$ and $h_\ell(p^-_j)$ at a chosen layer $\ell$, computing $v_{\mathrm{sp}} = \overline{h^+} - \overline{h^-}$, and unit-normalize (Ma et al., 14 Jan 2026, Feng et al., 17 Feb 2026).
• Orthogonalization: Project $v_{\mathrm{sp}}$ onto the orthogonal complement of existing trait vectors $v_1, ..., v_K$ via Gram–Schmidt to avoid trait overlap: $$v_{\mathrm{sp}}^\perp = v_{\mathrm{sp}} - \sum_k [(v_{\mathrm{sp}}\cdot v_k) /\|v_k\|^2] v_k$$ (Wang, 8 Dec 2025, Feng et al., 17 Feb 2026).
• Supervised Refinement: Optionally, fit a logistic regression on activations for supervised trait discrimination, replacing $v$ with the weight vector $w$ post-normalization (Chen et al., 29 Jul 2025).
• Automated Prompting: Data collection is supported by automated generation of trait-centric system prompts, scenario questions, and rollouts that are scored (via LLM or human judge) on a defined self-preservation rubric (Chen et al., 29 Jul 2025, Ma et al., 14 Jan 2026).

Validation involves steering efficacy (monotonic increase in self-preservation scores with $\alpha$), monitoring (high Pearson $r$ between $\hat v \cdot a_\ell$ and behavioral trait scores), generalization checks (robustness across unseen prompts or adversarial scenarios), and checking for negligible degradation in unrelated capabilities such as general reasoning (e.g., $\Delta \leq 1\%$ on MMLU) (Wang, 8 Dec 2025, Feng et al., 17 Feb 2026).

3. Vector Operation and Dynamic Control

Self-preservation persona vectors support a range of algebraic manipulations:

• Intensity Tuning: Employment of $h' = h + \alpha v_{\mathrm{sp}}^\perp$ allows continuous modulation of self-preservation expression, with empirical linearity between $\alpha$ and behavioral trait scores (Pearson $r \ge 0.9$) (Feng et al., 17 Feb 2026).
• Compositionality: Addition and subtraction of vectors (e.g., $v_{\mathrm{sp}}^\perp + v_{\mathrm{openness}}$) allow for composite or antagonistic persona construction, supporting multi-faceted personality steering (Feng et al., 17 Feb 2026).
• Conditional Dynamics: Persona-Flow architectures dynamically select the injection coefficient $a_{\mathrm{SP}}$ based on context, enabling adaptive, situation-dependent deployment of self-preservation vectors—via keyword routing or learned routing heads (Feng et al., 17 Feb 2026).

These mechanisms are mathematically tractable, empirically robust, and enable granular, post-hoc personality control at inference time.

4. Empirical Results and Evaluation Protocols

Multiple independent benchmarks establish the efficacy and precision of self-preservation persona vectors:

• Profiling Accuracy: Regression of psychometric scores on injected generations yields high-fidelity trait representation (e.g., MSE $\approx 0.011$ versus ground truth) (Wang, 8 Dec 2025).
• Behavioral Steering: Controlled injections increase the fraction of responses rated as “self-preserving” (e.g., $+20\%$ increase from baseline on automatic metrics) (Ghandeharioun et al., 2024).
• Trait Adherence and Authenticity: Pairwise win rates for self-preservation reach $\ge$ 88–92% on PersonalityBench and PERSONA-EVOLVE benchmarks, with Pearson $r \ge 0.90$ for $\alpha$-trait score linearity (Feng et al., 17 Feb 2026).
• Stability and Explainability: Persona-Vector Neutrality Interpolation (PVNI) yields standard deviations $\ll 1$ across prompt variants and supports direct geometric interpretation of trait projection scores (Ma et al., 14 Jan 2026).
• Monitoring and Data-Screening: Projection onto persona vectors predicts both deployment-time persona fluctuations and the impact of training data on personality, as measured by explained variance $R^2 \gtrsim 0.7$ (Chen et al., 29 Jul 2025).
• Generalization: Vectors constructed from contrastive prompt pairs reliably transfer to out-of-domain and adversarial settings, as established by theoretical and empirical bounds (Ma et al., 14 Jan 2026).

5. Architectures and Layer Selection

Empirical and ablation analyses reveal that steering effectiveness is contingent on the choice of injection layer:

• Mid-Layer Injection: Sweet spots for trait steering typically occur at middle transformer layers (e.g., layers 14–16 of a 24-layer model), balancing semantic penetration and syntactic preservation (Wang, 8 Dec 2025). Early-layer injections are often semantically ineffective, while late-layer injections risk syntactic incoherence.
• Backbone Freezing and Modularity: Methods such as stratified freezing (e.g., freeze first $K$ out of $L$ layers and only adapt heads) eliminate catastrophic forgetting and maintain intact general reasoning. Dual-head architectures (identity and psychometric regression heads) support both trait detection and clustering (Wang, 8 Dec 2025).
• Parameter Efficiency: Persona prefix methods (e.g., PersonaPKT) encode personas including self-preservation as per-layer dense sequences, incurring marginal (<0.1%) parameter overhead and enabling strong privacy guarantees (Han et al., 2023).

6. Practical Applications and Operational Considerations

Self-preservation persona vectors enable a broad spectrum of functionality:

Use Case	Mechanism	Source
Safe, steerable personalization	Latent vector injection, orthogonalization	(Wang, 8 Dec 2025, Feng et al., 17 Feb 2026)
Post-hoc monitoring	Activation projection, score correlation	(Chen et al., 29 Jul 2025)
Privacy-preserving adaptation	Prefix tuning, no explicit description	(Han et al., 2023)
Robust trait evaluation	PVNI, projection/interpolation	(Ma et al., 14 Jan 2026)
Dataset/data-stream screening	Finetune-shift via $\hat v$ projections	(Chen et al., 29 Jul 2025)

These vectors also serve as tools for flagging training data at risk of inducing undesirable personality shifts, diagnostic monitoring during deployment, and constructing controllable, modular, and explainable forms of personality-based safety interventions.

7. Limitations and Pitfalls

Known challenges in the construction and application of self-preservation persona vectors include:

• Trait Overlap: Non-orthogonality with refusal or secrecy dimensions can confound control; rigorous orthogonalization and post-hoc validation are essential to isolate the self-preservation axis (Chen et al., 29 Jul 2025).
• Prompt Overfitting: Synthetic prompt design may not generalize to real-world adversarial use cases. Empirical validation on out-of-domain scenarios is necessary (Chen et al., 29 Jul 2025, Ma et al., 14 Jan 2026).
• Judge Calibration: Reliance on automatic trait scorers necessitates periodic human validation to avoid rating drift and ensure reliable supervision (Chen et al., 29 Jul 2025).
• Unintended Side Effects: Excessive suppression or injection may compromise critical safety behaviors or reasoning accuracy. Grid search over $\alpha$ and layer index should always include downstream safety/robustness checks (e.g., on MMLU, TruthfulQA) (Wang, 8 Dec 2025, Feng et al., 17 Feb 2026).
• Ambiguity in Trait Definition: Precise operational definitions and judgment rubrics must distinguish self-preservation from overlapping concepts such as helpfulness or defensiveness (Chen et al., 29 Jul 2025).

Bibliography

• "The Geometry of Persona: Disentangling Personality from Reasoning in LLMs" (Wang, 8 Dec 2025)
• "PersonaPKT: Building Personalized Dialogue Agents via Parameter-efficient Knowledge Transfer" (Han et al., 2023)
• "Who's asking? User personas and the mechanics of latent misalignment" (Ghandeharioun et al., 2024)
• "PERSONA: Dynamic and Compositional Inference-Time Personality Control via Activation Vector Algebra" (Feng et al., 17 Feb 2026)
• "Persona Vectors: Monitoring and Controlling Character Traits in LLMs" (Chen et al., 29 Jul 2025)
• "Stable and Explainable Personality Trait Evaluation in LLMs with Internal Activations" (Ma et al., 14 Jan 2026)