Within psychology, the sub-field of social intelligence comprises two core components, social awareness and social facility. Social awareness refers to the capacity to understand and empathise with other’s emotions and perspectives, while social facility pertains to the ability to behave effectively in social situations.
Work to date on Willowbrook has shown the LLM propensity for human like behaviours, when generating sensible daily schedules, context-appropriate actions, and in-character conversations.
Recent work has shown that when sufficiently prompted, the communication style of an agent can portray that agents’ personality. But the ability of the agent to demonstrate realistic conversational reactions to certain situations appears to be constrained by the ‘chirpiness’ of the underlying model.
We have also shown that when an agent is engineered to maintain an internal model of its (the character’s) values and boundaries, it is possible for the LLM-agent to hold a view of other agents and their mannerisms, and a synthetic view of how they ‘feel’ about the other agent, which could be seen as a foundation for mock social awareness.
However, the mechanism by which an agent’s conversations and actions were impacted by these models has yet to be achieved, i.e. we need to improve the agent’s social facility by using its social awareness. The ideal outcome would be for two agents to have an argument, or to otherwise show a negative emotion towards another agent, instigated by the simulation (i.e. without direct prompting).
Funding is sought to cover the research costs to explore the internal representations needed to allow the agents actions and/or conversations to be influenced by their inner social awareness model.
Findings will be presented in a Willowbrook report, currently expected to be published via the CETaS website, as a sequel to the original report, which can be found at: https://cetas.turing.ac.uk/publications/welcome-willowbrook
1st October 2024, Dr Sarah Mercer.