SCITEPRESS proceedings
SAC: A Framework for Measuring and Inducing Personality Traits in LLMs with Intensity Control
Introduced Specific Attribute Control (SAC), extending MPI from the Big Five to 16PF so LLM personalities can be measured and steered across 16 fine-grained traits with explicit intensity levels. Experiments show continuous intensity control is more reliable than binary toggles and causes coherent shifts in related traits, enabling more nuanced human-machine interactions.
When you interact with ChatGPT, Claude or Gemini on a daily basis, you have a vague understanding of whether or not they are cooperative, kind, or even stubborn at times. However these perceptions are all subjective. Our research works towards evaluating these personality factors in an objective form. We then go one step beyond to introduce a framework that can steer these personality traits in a way that is more nuanced than just saying 'be more cooperative' or 'be more stubborn'. We introduce the concept of intensity control, which allows us to specify not just the direction of change but also the degree of change in these personality traits. This opens up new possibilities for creating AI agents that can adapt their personalities in a more human-like and contextually appropriate manner.
The use cases of this work are incredibly exciting, imagine a mental health agent that continuously modifies its personality on the basis of the interactions it has with its patient. On a certain day when the patient is feeling low, the warmth level of the agent goes up and on a certain day if the patient is seeking advice, the assertiveness level goes up.
Imagine a multi-agent framework where you have different agents with different personalities and you can control the intensity of their traits to create a more dynamic and engaging interaction. For example, in a customer service scenario, you could have an agent that is more empathetic and cooperative for handling sensitive issues, and another agent that is more assertive and direct for handling straightforward inquiries.
Overall, our work on SAC and intensity control provides a new way to think about and design AI personalities, moving beyond static traits to a more fluid and adaptable model that can better meet the needs of users in a wide range of contexts.
We started this work as a course project in our second year, professors took notice of the value that this work could have and encouraged us to pursue this further. In a year, we managed to conduct full fledged experiments and had results that had direct applications to real world scenarios.
Our work got accepted at the 18th edition of the International Conference on Agents and Artificial Intelligence (ICAART 2026) and we went to Marbella, Spain in March to present our work in front of some of the leading researchers in Artificial Intelligence.
In a sequence of events that kept getting more surreal - we met some of the best researchers from all over the world (who were all so welcoming to two complete noobs), had an incredible audience at our poster booth and we flew back to see that our paper has also been cited now :)