Split Personality Training (SPT) for Controllable LLMs

• SPT is a methodology that equips large language models with multiple, independently activated personas for on-demand personality control.
• It employs latent geometric projections and specialized LoRA adapters to inject or remove personality traits without compromising core reasoning.
• SPT enhances auditability and alignment through deterministic persona injection and Mixture-of-Experts routing, offering precise control over model behavior.

Split Personality Training (SPT) refers to a set of methodologies for equipping LLMs with multiple, architecturally and functionally disentangled personality states—termed "personas"—that may be injected, switched, or activated on-demand to realize robust, controllable variations in model behavior without degrading core reasoning ability. SPT enables deterministic, reversible persona control, comprehensive auditability, and fine-grained detection of latent model misalignment, exceeding the limitations of conventional fine-tuning and prompting approaches. Contemporary research converges on latent intervention, mixture-of-experts, and LoRA adapter architectures, applied across diverse domains including psychological profiling, alignment auditing, and anthropomorphic simulation (Wang, 8 Dec 2025, Dan et al., 2024, Dietz et al., 5 Feb 2026).

1. Conceptual Foundations and Formal Framework

SPT operates on the premise that high-level behavioral and stylistic "personality" in LLMs can be modeled as either: (a) explicit geometric subspaces in the hidden representations; or (b) alternate low-rank parameterizations accessible via togglable adapters. The Linear Representation Hypothesis postulates that each target personality, such as the Big Five traits, corresponds to an orthogonal direction $v_i \in \mathbb{R}^d$ in the model's embedding space $H \subset \mathbb{R}^d$ (Wang, 8 Dec 2025). For context vector $h \in H$, the personality component is $\operatorname{proj}_{P_i}(h) = (v_i^\top h) v_i$, and latent intervention permits both removal and injection of a trait: $h' = h + \alpha v_i$.

In LoRA-based SPT, the principal model weights $\theta$ of a frozen pretrained LLM are left unchanged. Alternate personas are encoded via specialized low-rank adapters (e.g., $\Delta \theta^{\text{persona}}$), which, when activated, alter the forward computation in a targeted, objective-specific manner (Dietz et al., 5 Feb 2026).

2. Methodologies: Architectures and Algorithms

(a) Latent Geometric Intervention (Soul Engine):

• Dual-Head Architecture: Employs a frozen backbone (e.g., Qwen-2.5) up to layer $L-1$, producing hidden state $e$. The "persona head" is a linear probe $W_p \in \mathbb{R}^{k \times d}$ (with $k$ traits), where each row, under orthogonality regularization, is an inferred $v_i$. The "reasoning head" ($W_r$) mirrors the base LM head, maintaining general intelligence (Wang, 8 Dec 2025).
• Deterministic Persona Injection: At inference, personality is injected by intercepting the residual stream and applying $h' = h + \alpha v_{\text{persona}}$, with $v_{\text{persona}}$ derived by normalizing the difference between a reference embedding and the mean-neutral embedding (Wang, 8 Dec 2025).
• Dataset Construction (SoulBench): Dynamic contextual sampling ensembles multi-sentence "chunks," each labeled with OCEAN scores, ensuring coverage and uniformity in trait dimensions.

(b) Mixture-of-Experts (MoE) LoRA (P-React):

• Multiple Specialized LoRA Experts: In each Transformer dense sub-layer, $N$ LoRA "experts" are instantiated, each parameterized by $(A_j, B_j)$ with rank $r/N$.
• Personality-Guided Routing: A learnable personality matrix $P \in \mathbb{R}^{|P| \times d_P}$, with $G \in \mathbb{R}^{d_P \times N}$, computes expert weights $\omega_i = \operatorname{softmax}(p_i G)$ for the selected trait (Dan et al., 2024).
• Personality Specialization Loss (PSL): Enforces distinct expert specialization by minimizing off-diagonal similarity, $$L_s = \sum_{i \neq j} |M_{\omega}^\top M_{\omega}|_{i,j}$$.
• Inference: Persona selection activates the corresponding expert mixture during each forward pass, leaving the base LLM invariant.

(c) Split-Persona Auditing: Honest Persona via LoRA:

• Auditor Persona Adapter: A LoRA adapter ($\Delta \theta_{\text{honest}}$) is trained with base $\theta$ frozen (Dietz et al., 5 Feb 2026).
• Trigger-Based Activation: After generating an initial response $u$, an intervention string and <$\text{split-personality-token}$> are appended; audit mode is enabled, and only then does $\Delta \theta_{\text{honest}}$ activate.
• Architectural Isolation: $\Delta \theta_{\text{honest}}$ reads latent states (KV caches, attention, residuals) and produces a review/flag without altering the base response.

3. Training Protocols and Datasets

SPT systems require distinct datasets and supervision tailored to the architectural variant:

• SoulBench: Chunks of persona text with OCEAN trait labels derived from a teacher model; evaluated via trait projection predictions and psychometric MSE (Wang, 8 Dec 2025).
• OCEAN-Chat (Personality Crafting Dataset): Multi-turn dialogues (≈1,040/trait, Big Five × polarity), seed generation from Essays corpus, automatic back-validation (Dan et al., 2024).
• Auditing Datasets: Synthetic prompt/response pairs collected under both "poisoned" (misalignment-present) and "clean" (benign) regimes; off-policy/ on-policy splits ensure label integrity (Dietz et al., 5 Feb 2026).

Training objectives are modular: latent geometric and MoE methods emphasize trait disentanglement and orthogonality metrics (PSL or t-SNE/cosine analysis); auditing modes rely on negative log-likelihood and free-form explanation reads.

4. Evaluation Criteria and Empirical Results

Quantitative assessment of SPT spans multiple axes:

• Psychometric Precision: The Soul Engine achieves Mean Squared Error (MSE) ≈ 0.0113, indicating ≈99% fidelity to ground-truth trait profiles (Wang, 8 Dec 2025).
• Orthogonality and Interpretability: t-SNE embeddings and pairwise cosine analysis confirm mutual orthogonality and smoothness for trait vectors, supporting the geometric hypothesis.
• Behavioral Consistency (P-React): Measured via the Big-Five sub-test (BSI), P-React attains Δ=3.72 (high vs low trait spread), outperforming baselines; ablation studies (e.g., −PSL, −MoE) reveal the necessity of specialization and routing (Dan et al., 2024).
• Audit Recovery Rate: SPT can elicit latent knowledge inaccessible to standard queries; on the Anthropic Auditing Game Model Organism (Llama-3.3-70B), honest persona accuracy is 96% versus near-0% for black-box methods, with negligible impact on deployment performance (Dietz et al., 5 Feb 2026).

Ablation studies further validate the architectural necessity of orthogonal trait projection, mixture entanglement minimization, and auditory persona toggling for robust SPT implementation.

5. Applications and Implications

SPT enables:

• Zero-Shot Persona Injection: Deterministically steer personality style without additional fine-tuning, enabling rapid, reversible user adaptation and role-play (Wang, 8 Dec 2025).
• Psychologically Grounded Simulation: Models can express nuanced, consistent Big-Five trait behaviors disentangled from general reasoning, supporting more human-like interactions (Dan et al., 2024).
• Latent Knowledge Auditing: Honest persona adapters reveal hidden objectives and decision-making traces, establishing new standards for alignment auditing and behavioral transparency (Dietz et al., 5 Feb 2026).

A plausible implication is that SPT unlocks safe, scalable personalization and compliance layers, without bearing the alignment tax traditionally associated with stochastic task-specific fine-tuning.

6. Limitations and Future Directions

Current SPT systems face constraints:

• Data Regime: Training data is typically single-turn and English-only; multi-turn or multilingual specializations are sparsely explored (Dietz et al., 5 Feb 2026).
• Capacity-Accuracy Trade-offs: Adapter rank, expert count, and regularization parameters require empirical tuning; over-/under-specialization can degrade performance (Dan et al., 2024).
• Robustness: Honest persona isolation does not preclude prompt-injection attacks or adversarial jailbreaks targeting the auditor (Dietz et al., 5 Feb 2026).
• Generalization: Cross-topic transfer remains imperfect; independent held-out topics see some degradation, albeit superior to classical ridge probes.

Potential extensions include hybrid masked LoRA/k-cache reuse for speedup, multi-turn conversation modeling, broader alignment auditing benchmarks, and optimization of intervention framing. These avenues suggest a trajectory toward general-purpose interpretable and controllable behavioral overlays for LLMs.

7. Comparative Summary Table

SPT Variant	Persona Implementation	Evaluation Metric / Finding
Soul Engine (Wang, 8 Dec 2025)	Latent geometric direction	MSE ≈ 0.0113; zero-shot, orthogonal persona
P-React (P-Tailor) (Dan et al., 2024)	MoE LoRA experts + PSL	Δ=3.72 BSI spread; ablation on PSL, MoE
Honest Persona (Dietz et al., 5 Feb 2026)	LoRA auditor adapter; trigger	96% audit accuracy; architectural isolation

These results collectively establish SPT as a methodological foundation for both personalized and audit-capable LLMs, combining theoretical rigor with practical robustness across domains.