Artificial intelligence has advanced to the point where we can not only generate coherent text but also model and control the personality of language models. Traditionally, this control has been based on static frameworks like the Big Five traits. However, a deeper and more dynamic approach emerges by applying the eight Jungian cognitive functions. This article explores the geometry of personality in the activation space of LLMs, showing how activation steering allows effective manipulation of these functions. Beyond theory, we analyze practical implications for businesses looking to personalize virtual assistants, AI agents, and customer service systems.
The Jungian functions —Thinking, Feeling, Sensation, and Intuition, each in their introverted and extraverted versions— offer a more nuanced representation of cognitive processes than simple personality traits. Recent research demonstrates that it is possible to extract direction vectors for each function and apply them via steering in models like Llama-3.1-8B, achieving monotonic control over all eight functions. This means we can increase or decrease the intensity of a specific function in the model's responses, opening the door to granular personalization.
One of the most fascinating findings is that personality information concentrates in the middle transformer layers. Additionally, control vectors exhibit structured geometric relationships that reflect the classic distinction between rational and irrational functions. For example, judging functions (Thinking and Feeling) tend to align in certain directions, while perceiving functions (Sensation and Intuition) occupy orthogonal regions in activation space. However, effective directions for multidimensional control cannot be reconstructed as simple linear combinations of individual directions, suggesting complex nonlinear interactions.
For businesses, this ability to modulate the personality of an AI agent has direct applications. Imagine a customer service bot that can switch between an analytical style (extraverted Thinking) for technical problem-solving and an empathetic style (introverted Feeling) for handling emotional complaints. Or a sales assistant that adapts its communication according to the customer's cognitive profile. Q2BSTUDIO, as a leading software and technology company, is already exploring these techniques to offer more human and effective artificial intelligence solutions.
Implementing this type of control requires a solid technical infrastructure. Activation vectors are extracted through experiments with specialized datasets —in this case, over 2,100 character narrations— and then applied during inference. The process involves modifying the model's internal representations at appropriate layers, a task that combines prompt engineering with weight adjustments. This is where Q2BSTUDIO's custom software development services make a difference: we can integrate these capabilities into existing systems, whether on the cloud or on-premises.
Personality is not the only field of application. The same steering technique with Jungian functions could extend to other cognitive aspects, such as creativity or logical rigor. Combined with Business Intelligence and Power BI tools, it would allow analyzing how different personality profiles impact business decision-making. For example, an AI model that simulates a management team with diverse functions could predict conflicts or synergies before a real meeting.
From a technical standpoint, the geometry of personality reveals that the activation space is not a simple Cartesian plane but a complex manifold with its own curvature and topology. This implies that simultaneous control of multiple functions requires more sophisticated optimization algorithms than simple vector addition. Our team at Q2BSTUDIO is researching reinforcement learning and gradient-based optimization methods to find composite directions that preserve model coherence.
In the field of cybersecurity, there are also promising applications. A model with controlled personality can serve as an intelligent honeypot that adapts its behavior to deceive attackers, simulating different roles depending on the threat. Moreover, by better understanding how traits are represented in internal space, we can detect jailbreaking attempts or malicious manipulation that alter the model's personality in unwanted ways.
Adopting these approaches requires scalable cloud infrastructure. Q2BSTUDIO offers AWS/Azure cloud services that enable deploying steered models in real-time, managing demand spikes and ensuring low latency. Combined with process automation solutions, companies can integrate assistants with dynamic personality into their workflows without hassle.
Looking ahead, the geometry of personality opens the door to AI systems that not only understand what we say but also how we say it and which cognitive style we prefer. The collaboration between Q2BSTUDIO and research centers allows us to stay at the forefront of these innovations, offering our clients the most advanced tools to create truly personalized digital experiences.
In conclusion, activation steering based on Jungian functions represents a qualitative leap in the interpretation and modulation of personality in language models. The evidence that information concentrates in middle layers and that control directions have a structured —yet nonlinear— geometry poses new challenges and opportunities. From commercial applications to fundamental research, this technology promises to transform how we interact with AI. At Q2BSTUDIO we are ready to accompany businesses on this journey, providing custom applications that integrate these advances ethically and effectively.
