GLANCE: Strategy-Based Visual Mediation for LLM Interaction

Abstract

Interaction with Large Language Models (LLMs) is inherently text-centric and requires users to interpret, compare, and revise generated responses. While visual augmentation can support these activities, existing interfaces implement highlighting or comparison views as fixed UI features without an explicit model controlling how augmentation is applied. This work formalizes LLM interface augmentation as the mediation strategy framework GLANCE, that defines (1) an interpretive task context, (2) an evidence derivation procedure operating on prompt-response artifacts, and (3) visual encodings that externalize the resulting evidence. The framework is instantiated in an interactive workspace that enables users to configure and combine mediation strategies within LLM interaction workflows. The framework is evaluated in a user study with 10 participants performing LLM communication tasks derived from common usage patterns of generation, improvement, and summarization. The findings indicate that strategy-mediated augmentation provides a flexible mechanism for externalizing interpretive intent and supporting revision-oriented interaction with LLM systems.