Select Page

Steve Battle

Design is a world-building activity, often spatial in nature. We have heard much about the incredible power of artificial intelligence, which is apparently threatening to take our jobs and overrun civilisation. Specifically, we have seen the emergence of large language models (LLMs). Are LLMs ready to engage in design conversations? These conversations might address product design or a city plan. However, conventional LLMs are particularly poor at the kind of spatial reasoning this might involve. This paper looks at how we might go about augmenting an LLM conversational agent with models that allow us to build shared conversational representations through dialogue.

However, a generative LLM will confabulate as readily as speak the truth, like a creative brain without a critical faculty. The approach taken here to mitigate this effect is to create a system that includes multiple voices inspired by Gordon Pask’s conversation theory. We consider the scenario of an LLM in conversation with a designer, coupled with a critical voice that focuses on alignment with a shared model. The critic is a simple procedural agent, able to pick out simple phrases from the conversation and build a corresponding model. Its purpose is to detect inaccuracies and provide corrective feedback to the LLM.

The combined effect of these entangled conversations is a more stable and performant system than either the LLM or its critic can achieve alone.

KEYWORDS: cybernetics, conversation theory, large language models, mental models

RSD CATEGORY: Sociotechnical Systems




Citation Data

Author(s): Steve Battle
Year: 2023
Title: Conversation Theory for Design Agents
Published in: Proceedings of Relating Systems Thinking and Design
Volume: RSD12
Article No.: pre-release
Host: Georgetown University
Location: Washington DC, USA
Symposium Dates: October 6–20, 2023
First published: 30 September 2023
Last update: no update
Publisher Identification: ISSN 2371-8404
Verified by MonsterInsights