Challenges of Existing AI Poster Editing Methods
Although AI tools for automatically generating academic posters, such as Paper2Poster, have emerged in recent years, they generally lack the ability to make secondary edits to initial drafts. Existing editing solutions also have significant drawbacks: on one hand, regeneration-based methods often produce “hallucinations” when handling precise content like academic charts, leading to data inaccuracies or distortions. On the other hand, general-purpose presentation editing agents struggle with the complex layout tasks of dense, structurally intricate academic posters due to a lack of understanding of the academic domain.
APEX: A Solution Based on Multi-Level APIs and a Review Mechanism
To address these challenges, the Planing Lab team at East China Normal University proposed the APEX (Academic Poster Editing Agentic Expert) framework. This framework parses PPTX files, structuring the poster into JSON data. At its core is a “Plan-and-Execute” module that, instead of relying on unstable code generation, calls a set of predefined multi-level APIs to perform incremental edits on the poster. When a command involves paper content, APEX uses a paper comprehension tool to extract factual data from the original text, preventing content fabrication.
APEX’s key innovation is the introduction of a “Review-and-Adjustment” mechanism. After the initial edits are completed, a multimodal agent compares the visual and semantic differences between the before and after versions, checking for deviations from the command or unnecessary changes. If issues are found, the system automatically generates corrective instructions for a second round of adjustments, ensuring the accuracy of the edits.
APEX-Bench: The First Evaluation Benchmark for Academic Poster Editing
To systematically evaluate poster editing capabilities, the team also constructed the first academic poster editing benchmark, APEX-Bench. The benchmark includes 59 papers from top conferences like ICLR and ICML, with 514 corresponding editing commands of varying difficulty levels. These commands cover a range of tasks, including content modification, style adjustment, and layout restructuring. The evaluation is conducted by a multimodal large model (VLM-as-a-judge) across three dimensions: command following, scope of modification control, and visual consistency.
Experimental Validation and Significance
Experimental results on APEX-Bench show that APEX significantly outperforms other compared methods across all key metrics, including direct image generation, XML generation, and general-purpose presentation agents. It particularly excels in controlling the scope of modifications, effectively avoiding unintended changes to irrelevant areas. This demonstrates the superiority of the APEX framework in accurately executing user commands and maintaining design consistency, providing researchers with an efficient and reliable automated editing assistant.
