Large Language Models (LLMs) are increasingly being explored as a new substrate for supporting group discussions and collaborative decision-making.However, the design space of how LLM-based systems intervene in human-human interaction remains fragmented.To map this emerging landscape, we conducted a scoping review of 11 papers identified through a screening process that combined author screening with LLM-assisted parallel screening.Our thematic analysis identifies three typologies of LLM-based intervention: ( 1 ) Active Discursive Intervention, in which the system generates discursive contributions (e.g., facilitator prompts or devil' s-advocate critiques) intended to mitigate social inhibitors; ( 2 ) Visual and Cognitive Scaffolding, which externalizes discussion content into structured or multimodal representations to support sense-making; and ( 3 ) Process and Temporal Management, which supports continuity of work across synchronous and asynchronous contexts.We further discuss challenges associated with more socially interactive forms of LLM support, including issues of human agency and trust.Based on these insights, we discuss design implications for future systems: tunable agency, multimodal co-creation, and human-on-the-loop architectures.Finally, we propose an agency-oriented design space for LLM-based group discussion support that characterizes systems along a five-level continuum from off-session support to semi-automated, human-controlled intervention and, conceptually, to out-of-the-loop autonomous co-participation.
光太郎 鎌田 (Tue,) studied this question.