Macaron-A2UI: A Model for Generative UI in Personal Agents

Fancy Kong, Congjie Zheng, Murphy Zhuang, Rio Yang, Sueky Zhang, Hao Fu, Gene Jin, Song Cao, Kaijie Chen, Andrew Chen, Pony Ma

arXiv:2605.24830 · 2026-05-26 공개 · arXiv · PDF

reinforcement-learning parameter-efficient lora-finetuning personal-agents a2ui-bench generative-ui large-models ui-actions

Abstract

As personal agents evolve to handle complex, user-centric tasks, static plain-text chat is rapidly becoming a bottleneck. Generative UI emerges as the necessary new interface layer, dynamically synthesizing the right controls, options, and state from the interaction context in real time. We present Macaron-A2UI, a model for Generative UI in personal agents. Our goal is to move beyond text-only interaction by enabling agents to generate natural language together with lightweight, executable UI actions for information collection, preference refinement, confirmation, and multi-goal organization. We build a large-scale Generative UI corpus from heterogeneous dialogue sources, introduce A2UI-Bench for controlled evaluation, and train 30B, 235B and 754B models with parameter-efficient LoRA-based supervised fine-tuning followed by reward-driven reinforcement learning. The best Macaron-A2UI model reaches 75.6 overall on A2UI-Bench without explicit schema hints, surpassing the strongest full-schema frontier baseline. We release the models, benchmark, and evaluation protocol to support future work on Generative UI for personal agents.

한국어 요약

📋 한 줄 요약

**[Generative UI / Personal Agent]** Macaron-A2UI가 personal agent의 정적 plain-text chat 한계를 넘어 lightweight executable UI action 생성 — 30B/235B/754B 모델을 LoRA SFT + RL로 학습, A2UI-Bench에서 schema hint 없이 75.6점·full-schema baseline 능가.

🎯 핵심 기여도

💡 핵심 아이디어

Personal agent의 진정한 interaction은 plain text가 아니라 generative UI action을 동반해야 하며, heterogeneous dialogue corpus 위에 LoRA-SFT + reward RL 학습 + explicit schema에 의존하지 않는 inference가 결합되면 schema hint를 제공받는 강력 baseline을 능가할 수 있다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Personal agent의 interface paradigm을 generative UI로 확장, schema-free inference로 실용 배포 단순화, 모델·벤치·프로토콜 공개로 후속 연구 가속. **한계**: A2UI-Bench가 모든 personal agent 시나리오 커버 미보장, 754B 규모 학습의 자원 부담, UI action의 실제 사용자 만족도와 점수 간 격차 가능.

🚀 실용적 활용