From Scratch to Precise Illustration: Automatic Illustration Generation with Iterative Multi-Agent Refinement

ACL ARR 2026 January Submission8285 Authors

Keywords: tool-augmented LLMs, illustration generation, multi-agent collaboration, educational AI

Abstract: Automated generation of illustrations for educational exams can significantly reduce the labor-intensive workload of human exam authors. While recent LLM research has automated textual question generation, visual components remain underexplored despite their prevalence in real-world exams. In this work, we extend tool-augmented LLMs to illustration generation via specialized Tool Sandbox, and introduce Actre, an iterative multi-agent framework that improves tool utilization through feedback-driven refinement. Furthermore, we propose IllustrationBench, a graded benchmark of 320 real college entrance exam questions across four types and three difficulty levels. Experiments show Actre significantly outperforms one-shot baselines, particularly on high-difficulty tasks, demonstrating the effectiveness of iterative tool-augmented agents for educational content generation.

Paper Type: Short

Research Area: AI/LLM Agents

Research Area Keywords: LLM agents, tool use, function calling, multi-modal agents, multi-agent systems, agent communication

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 8285