Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

Technical Summary: Mechanistic Personality Analysis of LLMs

Problem Statement
Large Language Models (LLMs) inherently simulate human-like personality traits (specifically the Big Five/OCEAN traits) due to their training on vast human-generated data. While previous methods have attempted to shape these personalities via prompt engineering or fine-tuning, these approaches suffer from significant drawbacks: inconsistency in output, high resource intensity, and the risk of degrading performance on core language tasks. Furthermore, shaping multiple intercorrelated traits simultaneously remains a challenge, often requiring model capacities that smaller or less optimized models lack. There is a need for a method that can controllably steer personality traits at the mechanistic level without compromising the model's fundamental language modeling abilities.

Methodology
The authors propose a mechanistic interpretability framework that directly intervenes on the model's latent features to modulate personality. The pipeline consists of four main components:

1. Data Generation and Dataset Creation:
   Using the PsyBORGS framework, the authors generate 12,000 Facebook-style status updates with varying levels of Big Five trait expression. Unlike prior work using neutral baselines, this method creates distinct "Positive" (high expression, scores 7–9) and "Negative" (low expression, scores 1–3) sets for each trait to identify both trait-enhancing and trait-suppressing features.

2. Feature Extraction via Sparse Autoencoders (SAEs):
   The authors utilize a pre-trained Sparse Autoencoder (SAE) attached to layer 19 of the DeepSeek-R1-Distill-Llama-8B model. This layer was selected as it represents a "sweet spot" (63% depth) where features are semantically rich yet upstream of final logits.

   • Contrastive Analysis: Mean activations are computed for the positive ($\mu_P$) and negative ($\mu_N$) sets. The difference vector $\Delta h = \mu_P - \mu_N$ is decomposed using the SAE encoder to obtain latent differences $\Delta z$.
   • Feature Ranking: Latent components are ranked by the absolute difference $|\Delta z_i|$. The top-$k$ features driving the trait (positive direction) and suppressing the trait (negative direction) are identified. These are converted into unit feature vectors ($u_{ij}$) via the SAE decoder.

3. Optimization via Grid Search:
   To determine the optimal intervention, the authors employ a bidirectional linear weighting approach. A steering vector is constructed by adding a weighted combination of positive and negative feature vectors to the model's hidden states ($h'$):
   $$h' = h + \sum \delta^{pos} u_{ij} + \sum \delta^{neg} u_{jl}$$
   The coefficients are tuned via a grid search over four parameters: the magnitude of positive shifts ($\alpha_{pos}$), the count of positive features ($k_{pos}$), the magnitude of negative shifts ($\alpha_{neg}$), and the count of negative features ($k_{neg}$).
   The optimization maximizes an objective function $O$ that balances trait expression ($S_{trait}$) against a penalty for performance degradation ($C$), measured by the relative drop in MMLU accuracy:
   $$O = S_{trait} - \lambda \cdot C$$
   where $\lambda$ is fixed at 0.5.

4. Evaluation Framework:
   The efficacy of steering is measured using a multi-faceted approach:

   • Embedding Similarity: Cosine similarity between generated text and reference trait exemplars.
   • LLM-Based Classification: State-of-the-art LLMs (Claude 3.7, GPT-4o) judge which of two texts better exhibits a target trait.
   • Human Evaluation: Parallel human assessment of the same pairs.
   • Performance Metrics: MMLU accuracy and coherence checks to ensure language capabilities are preserved.

Key Contributions

1. Methodology for Feature Extraction: A novel approach to extracting personality-relevant features via contrastive analysis between positive and negative prompt conditions, utilizing SAEs to isolate monosemantic latent features.
2. Multi-Faceted Evaluation: The establishment of a robust evaluation framework combining embedding-based similarity, LLM classification, and human evaluation to measure personality expression.
3. Refined Steering Strategies: The development of a bidirectional linear weighting approach with grid search optimization to compute optimal additive shifts in latent space, balancing trait expression with task performance.

Results

• Detection Rates: The intervention was reliably detected by both LLMs and humans for most traits. Conscientiousness and Neuroticism showed the highest detection rates (>90% for both LLM and human evaluators). Extraversion and Agreeableness showed moderate rates, while Openness proved challenging (50% LLM, 30% human), likely because LLMs are inherently "open" by design or because the trait manifests in subtle patterns.
• Trait-Specific Optimization: Grid search revealed that different traits require different intervention profiles.
  • Extraversion required the fewest features (9 positive, 4 negative), suggesting a concentrated latent representation.
  • Openness required the most features (25 positive) and high magnitude, indicating a distributed encoding.
  • Conscientiousness required a balanced mix of positive and negative features.
• Performance Trade-offs: All traits exhibited a "peak" in the objective function followed by a sharp decline in performance (coherence and MMLU accuracy) as intervention magnitude increased. This highlights a critical threshold where excessive feature manipulation degrades the model's language generation capabilities.
• Negative Shifts: While positive shifts generally increased trait similarity, negative shifts produced mixed results; for Openness, negative shifts sometimes resulted in higher similarity to the trait reference than positive shifts, suggesting complex relationships between feature activation and trait expression.

Significance and Claims
The paper claims to provide an interpretable and fine-grained method for personality steering that bridges high-level psychometric observations and low-level neural mechanisms. The significance lies in:

• Mechanistic Control: Demonstrating that personality can be modulated by directly intervening on specific latent features without retraining the model, offering on-demand modulation.
• Interpretability: The use of SAEs allows for the identification of specific directions in the residual stream corresponding to psychological traits, demystifying how these abstract concepts emerge in the model.
• Responsible AI: The approach offers a tool for controlling undesirable personality manifestations in AI agents, potentially improving safety, fairness, and effectiveness in communication.
• Computational Perspective on Psychology: The findings suggest that the complexity of human personality traits (e.g., the distributed nature of Openness vs. the concentrated nature of Extraversion) may be mirrored in the latent feature structure of LLMs.

The authors remain modest regarding the implications, noting that while the model exhibits behaviors associated with specific traits, this does not equate to the model possessing human-like psychological states. They also acknowledge limitations regarding computational overhead, the difficulty of maintaining coherence at high intervention magnitudes, and the challenge of disentangling cross-trait dependencies due to polysemanticity in the SAE features.