Macaron-A2UI: A Model for Generative UI in Personal Agents
Fresh 2026 arXiv paper focused on runtime Generative UI for personal agents. The authors argue that plain chat becomes a bottleneck for complex agent tasks, introduce a large-scale Generative UI corpus plus A2UI-Bench, and train models that generate natural language together with lightweight executable UI actions for confirmation, preference refinement, and multi-goal coordination.
Efficient Personalization of Generative User Interfaces
Recent arXiv paper on one of the core unsolved GenUI problems: personalization. The authors collect pairwise judgments from trained designers over the same 600 generated UIs, show that design preferences diverge substantially, and propose a sample-efficient preference model that personalizes generated interfaces better than baseline evaluators and direct prompting.
MAIC-UI: Making Interactive Courseware with Generative UI
Recent arXiv paper introducing a zero-code GenUI system for educational courseware. MAIC-UI turns textbooks, PPTs, and PDFs into interactive learning pages, uses a generate-verify-optimize pipeline for pedagogical accuracy, and supports click-to-locate incremental edits instead of expensive full regeneration.
Software as Content: Dynamic Applications as the Human-Agent Interaction Layer
2026 arXiv paper arguing that chat alone is a poor medium for structured, stateful human-agent work. The authors propose Software as Content, where dynamically generated agentic applications become the primary interaction layer, persist across turns, and expose actionable controls instead of forcing everything through linear text.
Gradual Generation of User Interfaces as a Design Method for Malleable Software
2026 arXiv paper proposing "Gradually Generating User Interfaces", a GenUI design method that inserts intermediate interface layers during generation so users can discover and control customization options without being overwhelmed by a giant prompt box or dense menu system.
Generative UI: LLMs are Effective UI Generators
Google Research paper that first systematically articulates the Generative UI concept. The paper defines Generative UI as "AI models generating not just content, but the entire user experience". Experiments show that with the right prompts and tools, LLMs can generate high-quality custom interactive interfaces that are significantly preferred by human users (44% of cases comparable to human experts). The paper also releases the PAGEN dataset for evaluating generated interfaces.
Generative Interfaces for Language Models
ACL 2026 Findings paper proposing a paradigm where LLMs proactively generate task-specific interfaces instead of replying only with text. The work pairs interface-specific representations with iterative refinement and reports that human evaluators preferred generative interfaces over conversational ones by up to 72% across information-dense and exploratory tasks.
Towards a Working Definition of Designing Generative User Interfaces
DIS 2025 conference paper that proposes the first working definition of Generative UI (GenUI) through literature review and case analysis. The research identifies GenUI as a new paradigm of interface creation where "humans and AI collaborate at design-time to generate interfaces, and users interact with AI-generated interfaces at runtime". The paper summarizes five key characteristics of GenUI: co-creation, design space expansion, representational fluidity, contextual adaptation, and generation-first.
Generative and Malleable User Interfaces with Generative AI
CHI 2025 paper that proposes using AI to generate "generative and malleable user interfaces" based on user tasks. The authors designed a task-driven data model where LLMs parse user prompts and generate UI specifications, then map them to concrete interfaces. Experiments demonstrate that this method can dynamically generate interface elements such as forms and visualizations, supporting users to modify interfaces via natural language.