APEX: Academic Poster Editing Agentic Expert

TL;DR

APEX addresses the gap between automated, one-shot paper-to-poster generation and professional, iterative poster editing by introducing an interactive agentic framework for poster editing. It combines a robust multi-level API editing system with a review-and-adjustment loop and pairs it with APEX-Bench, a benchmark of 514 editing instructions across 59 papers, evaluated via a multi-dimensional VLM-as-judge protocol. Experimental results show that APEX outperforms regeneration-based and generic slide-editing baselines in instruction fulfillment and visual consistency while maintaining reasonable modification scope and cost. The approach enables reliable, intent-aligned poster edits in high-density layouts, with careful attention to data provenance and ethics, and outlines future work on handling external visual assets and broader model exploration.

Abstract

Designing academic posters is a labor-intensive process requiring the precise balance of high-density content and sophisticated layout. While existing paper-to-poster generation methods automate initial drafting, they are typically single-pass and non-interactive, often fail to align with complex, subjective user intent. To bridge this gap, we propose APEX (Academic Poster Editing agentic eXpert), the first agentic framework for interactive academic poster editing, supporting fine-grained control with robust multi-level API-based editing and a review-and-adjustment Mechanism. In addition, we introduce APEX-Bench, the first systematic benchmark comprising 514 academic poster editing instructions, categorized by a multi-dimensional taxonomy including operation type, difficulty, and abstraction level, constructed via reference-guided and reference-free strategies to ensure realism and diversity. We further establish a multi-dimensional VLM-as-a-judge evaluation protocol to assess instruction fulfillment, modification scope, and visual consistency & harmony. Experimental results demonstrate that APEX significantly outperforms baseline methods. Our implementation is available at https://github.com/Breesiu/APEX.

Figures (18)

• Figure 1: Overview of the Multi-Agent Poster Editing Pipeline. The framework comprises three collaborative stages: (1) Semantic Parsing and Element Serialization: The source poster ${P}_\text{src}$ is parsed into structured JSON data $P^J_\text{src}$, extracting all element attributes for fine-grained control. (2) Planning and Execution: A centralized agent synthesizes user instructions $I$, visual representation $P^V_\text{src}$, and JSON data $P^J_\text{src}$ to generate an execution plan. It optionally invokes a paper understanding tool to extract content from the paper $M$ before calling multi-level APIs to modify the .pptx file. (3) Review and Adjustment: A quality assurance agent evaluates the edited visual output $P^V_\text{edited}$ against the original $P^V_\text{src}$ and the user's instruction $I$, specifically verifying the modified elements, and performs adjustments through additional API calls to ensure visual fidelity and instruction compliance.
• Figure 2: The data construction pipeline of APEX-Bench. Adopting a "Model-assisted, Human-refined" strategy, the workflow consists of three phases: (1) Data Sources Preparation, where initial drafts are synthesized from source papers via PosterGen; (2) AI Instruction Generation, where an VLM performs gap analysis and aesthetic optimization to derive preliminary editing commands; and (3) Human Refinement, where experts verify and adjust instructions to ensure feasibility and high quality.
• Figure 3: Performance across varying difficulty levels on three evaluation metrics. I.F.: Instruction Fulfillment, M.S.: Modification Scope, V.C.: Visual Consistency.
• Figure 4: Performance across varying difficulty levels on three evaluation metrics.
• Figure 5: Comparison of different methods under a complex editing instruction, involving swapping sections, restructuring content into columns, removing specific images, and adjusting vertical spacing.