PosterForest: Hierarchical Multi-Agent Collaboration for Scientific Poster Generation

A new framework called PosterForest generates scientific posters without additional training by using a hierarchical multi-agent system, outperforming prior methods in both automatic and human evaluations, according to a paper posted on arXiv [1]. The framework, detailed in a submission last revised on 17 June 2026, introduces a structured intermediate representation named the Poster Tree [1]. This representation captures document hierarchy and visual-textual semantics across multiple levels [2]. Content and layout agents then perform hierarchical reasoning and recursive refinement, progressively optimizing the poster from global organization to local composition [1]. The joint optimization improves semantic coherence, logical flow, and visual harmony, the authors state [2]. The paper was first submitted on 29 August 2025 by Jiho Choi [1]. The initial version was 39,272 KB, while the second and third versions were 43,632 KB [1]. Existing approaches to automated poster generation often rely on flat summarization or optimize content and layout separately, leading to information loss, weak logical flow, and poor visual balance [2]. PosterForest addresses these limitations through its training-free design, requiring no domain-specific supervision [2]. The method’s performance was measured against prior techniques in both automatic metrics and human assessments [1]. While the paper focuses on scientific communication, the broader field of development communication has long emphasized structured information dissemination and education as techniques to create positive social change [3]. The philosophy of development communication, articulated by scholars such as Nora C. Quebral, frames the practice as “the art and science of human communication applied to the speedy transformation of a country and the mass of its people from poverty to a dynamic state of economic growth” [3]. Although PosterForest operates in a different domain, its emphasis on hierarchical document understanding echoes the structured, purposive communication strategies studied in that field [3]. Advances in materials science have similarly depended on structured, multi-level understanding. Transition-metal dichalcogenide monolayers, atomically thin semiconductors such as MoS2 and WS2, exhibit properties distinctly different from bulk crystals, including a direct band gap useful in transistors and optical emitters [6]. The discovery that thinning a bulk crystal to one atomic layer can unlock new physical properties parallels the way PosterForest’s hierarchical representation unlocks coherent visual-textual planning from a flat research paper [2][6]. These monolayers are often combined with other two-dimensional materials like graphene to build van der Waals heterostructures for devices such as solar cells and photodetectors [6].