Haojie Hu, Chenhao Dang, Yaojia Liu, Hengrui Kang, Conghui He, Weijia Li
6 min
Abstract
Scientific poster construction compresses a long multimodal paper into a readable, editable canvas. Existing systems hide request-level failures by scoring only completed outputs; direct image generation is not element-editable, while coding-agent workflows are costly. PosterMELD is a template-conditioned multi-agent pipeline: capacity-aware slots guide writing before rendering, and deterministic gates plus vision-language model (VLM) review route failures to bounded repair. Each accepted request exports editable PowerPoint (PPTX) and Portable Network Graphics (PNG) artifacts; explicit design controls yield same-paper variants. Across 621 papers, Print-Ready Rate (PRR) counts requests passing geometric, readability, asset-integrity, and obvious-factual-error checks, with native editability reported separately. A frozen VLM assigns conditional Craftsmanship-Harmony-Expressiveness (CHE) scores to print-ready outputs. PosterMELD attains 81.3% PRR, 3.4 times P2P's rate and 5.2 times PosterGen's, and the highest conditional CHE among generated methods with multiple print-ready outputs. Native editability and explicit design controls are retained at a mean cost of USD 0.38 per request, 3.5% of Codex+Skill's. Code and resources are available at https://github.com/Shannon4Science/PosterMELD.
Sam: That's a meaningful gap. And once the poster is generated, can a researcher still go in and make changes?
Alex: Yes, and this was clearly a design priority. The system exports a standard PowerPoint file, so every element—text, images, layout—remains fully editable. The authors describe this as keeping the file "natively editable," meaning no specialist software is needed to make adjustments.
Sam: I'm curious about the templates themselves. Did the researchers just design them by hand?
Alex: No, they took a more data-driven approach. They used a tool called MinerU to extract structural data from a large collection of existing academic posters, and then grouped those posters by their visual structure using a clustering method. The idea was to let real-world examples define what a good layout actually looks like, rather than guessing.
Sam: And how many distinct layouts did they end up with?
Alex: Twenty-four distinct layout shapes, which they call "topologies." Each one represents a proven arrangement that real researchers have actually used. They then tested the whole system against a benchmark of over six hundred papers drawn from ten different research fields—biology, social sciences, and others—to make sure the templates could handle the different conventions each field uses.
Sam: Different fields do have very different styles. A biology paper with lots of figures needs a very different layout than a philosophy paper that's mostly text.
Alex: Exactly. That breadth of testing is what makes the results more credible. A tool that only works for one type of paper isn't particularly useful in practice.
Sam: You mentioned the system uses automated checks to grade the posters. How do they keep those grades consistent across different tests?
Alex: They use a large language model as a judge—essentially an AI that scores each poster against a fixed rubric. The key word is "fixed." By keeping that judge frozen, meaning its scoring criteria don't change between experiments, they ensure that every poster is being evaluated by the same standard. Without that, you couldn't reliably compare one method against another.
Sam: That's an important methodological detail. It also points to a broader limitation, though. The system can check whether a poster looks right, but it can't check whether the science inside it is accurate.
Alex: That's a key limitation the authors acknowledge. The system can guarantee visual and geometric integrity—that nothing overflows, that figures are placed correctly, that the file is editable. But whether the scientific content is being interpreted faithfully from the original paper is a separate problem, and one that remains the researcher's responsibility.
Sam: So it's a tool for presentation, not for verification. It handles the "how" of communicating research, not the "whether it's correct."
Alex: That's a precise way to put it. And within that scope, it does represent a clear improvement over manual methods. Researchers spend significant time reformatting papers for conferences and presentations. A system that handles the tedious structural work—reliably and with an editable output—frees them to focus on the content itself.
Sam: It's a practical solution to a genuinely time-consuming problem. Thanks for walking me through it, Alex.
Alex: My pleasure. That's our look at PosterMELD. Thanks for listening to ResearchPod.