presentation skills

Existing document-to-presentation generation methods largely neglect visual design principles and structural coherence, limiting their practical utility. To address this, we propose an end-to-end, two-stage editing-based generation framework: Stage I leverages large language models to learn document structural patterns and infer design constraints; Stage II employs code-driven atomic editing actions to jointly optimize content selection, layout arrangement, and stylistic consistency across slides. Furthermore, we introduce PPTEval—the first three-dimensional evaluation benchmark covering content accuracy, visual design合理性, and cross-slide coherence—equipped with interpretable, multi-faceted metrics. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art baselines across all three dimensions. The source code, dataset, and evaluation toolkit are fully open-sourced.

Existing automated media presentation generation methods often suffer from narrative discontinuity and suboptimal visual layout, failing to meet professional quality standards. To address these issues, we propose RCPS—a reflective, multi-agent framework integrating deep structured narrative planning, adaptive layout generation, and an iterative optimization loop. Furthermore, we introduce PREVAL, a preference-based evaluation system that jointly optimizes content consistency, coherence, and visual design across multiple dimensions. Experimental results demonstrate that RCPS-generated presentations significantly outperform baseline approaches across all quantitative metrics, achieving overall quality comparable to human experts. Crucially, PREVAL’s assessments exhibit strong agreement with human judgments (Spearman’s ρ > 0.92), validating its effectiveness as a reliable, automated quality evaluator.

Existing automated presentation generation methods heavily rely on research papers, limiting generalizability to generic Word documents, and lack end-to-end joint generation of slides and speech—resulting in time-consuming manual authoring, incoherent narration, and poor immersion. This paper proposes the first end-to-end framework for joint slide and spoken commentary generation from generic documents. It integrates information extraction, structured summarization, multimodal content planning, and text-to-speech (TTS) synthesis, driven by a fine-tuned large language model (LLM) for document understanding and narrative generation. We also introduce the first automated evaluation metric encompassing three dimensions: relevance, coherence, and redundancy. Experiments demonstrate significant quality improvements, with strong agreement between automatic scores and human ratings (Spearman’s ρ > 0.92). The code and dataset are publicly released to ensure reproducibility.

This work addresses the challenge of automatically generating human-like, audiovisually synchronized presentation videos from long documents. We propose an end-to-end multimodal generation framework comprising a modular pipeline: large language models perform segment-wise document understanding and content distillation; text-to-speech, visual generation, and audiovisual alignment modules jointly produce讲解-style videos with precise image-text matching, semantic coherence, and temporal accuracy. To enable rigorous evaluation, we introduce PresentEval—a unified benchmark that quantitatively assesses content fidelity, visual clarity, and audience comprehension—marking the first such framework for presentation video generation. Experiments on 30 real-world document–presentation pairs demonstrate that our generated videos approach human-produced quality across multiple metrics, significantly advancing the conversion of lengthy textual content into accessible, highly engaging dynamic presentations.

Current AI-based presentation training tools suffer from fragmented functionality, insufficient high-quality exemplars, and a lack of personalized, immersive feedback. To address these limitations, we propose a dual-agent collaborative framework: an Ideal Demonstration Agent generates personalized demonstration videos from input slides, while a Coach Agent delivers structured, multimodal feedback—analyzing speech, visual behavior, and verbal content—using a novel Observation-Impact-Suggestion (OIS) format. A third Audience Agent simulates authentic listener responses to enhance feedback humanization and contextual immersion. The system integrates slide understanding, vision-language modeling, text-to-speech synthesis, voice cloning, and multimodal speech analysis to establish an end-to-end training loop. Experimental results demonstrate significant improvements in user engagement, pedagogical effectiveness, and skill acquisition efficiency. The framework exhibits strong scalability and practical applicability in both educational and professional training settings.

This work addresses the challenge of generating high-quality, pedagogically coherent questions from multimodal lecture slides, which requires integrating dispersed textual and visual content while respecting the global instructional logic rather than treating slides in isolation. The authors propose a four-stage large language model pipeline—comprising window planning, lecture synthesis, slide annotation, and global coordination—that jointly models lecture-level teaching objectives and multimodal content for the first time. By incorporating quota allocation and cross-slide deduplication mechanisms, the system produces structured, low-redundancy question sets. Implemented in Flask, it supports PDF text and image extraction, multi-stage inference, and pedagogical evaluation. Experiments on two technical lecture datasets demonstrate its effectiveness in filtering non-instructional content and generating logically coherent, high-fidelity questions, along with comprehensive annotations including learning objectives, structure, summaries, and quality scores.

This work addresses a critical gap in current paper-to-video systems, which typically evaluate only content coverage while neglecting audience comprehension of core scientific ideas. To bridge this gap, the authors propose EffectivePresentationScorer, a novel framework that, for the first time, focuses on the pedagogical effectiveness of scientific presentation videos. By integrating natural language understanding with principles from educational assessment, the framework establishes a multidimensional scoring mechanism to evaluate script quality across key dimensions such as logical structure, contextual framing, and clarity of technical explanations. Experimental results reveal that while automatically generated videos often cover topical content and maintain structural coherence, they consistently fail to clearly articulate prerequisite concepts and methodological rationale. These findings expose a significant blind spot in existing evaluation metrics and establish the first automated approach for assessing the instructional quality of research communication videos.

This work addresses the challenge of automatically transforming static scientific papers into dynamic presentation formats—such as posters, slides, and videos—while preserving semantic consistency across modalities. We formalize this as the Unified Presentation Suite Generation task and propose a centralized content planning framework grounded in renderable HTML. To ensure coherence among multimodal outputs, we introduce a self-correcting verification-and-repair loop. Our contributions include OmniPreBench, a large-scale dataset; a vision-language model–based evaluation protocol; and a method for aligning multimodal content. Experimental results demonstrate that our approach significantly outperforms strong baselines in both factual accuracy and visual appeal, enabling the automatic generation of high-quality, semantically consistent scientific communication materials.

Existing agents lack fine-grained evaluation methodologies for complex, multimodal PowerPoint tasks, making it difficult to assess partial completion and diverse correct solutions. This work proposes the first fine-grained evaluation framework specifically designed for PowerPoint manipulation, introducing a benchmark comprising 120 tasks and a multidimensional scoring mechanism based on human-designed rubrics. The framework supports partial credit for content creation and editing tasks, incorporates aesthetic penalties, detects redundant operations, and provides natural language feedback. It achieves a Kendall’s τ-b correlation of 0.77 with human judgments. Experimental results reveal that even state-of-the-art models, such as Claude-4.5-Opus, attain only a 45% full-task success rate and a 57% average partial score, highlighting significant limitations in current agent capabilities.

This work addresses the limitations of existing presentation generation methods, which are largely confined to static slides and struggle to respond to open-ended queries, integrate multimodal content, or support interactivity. We propose the first unified agent framework for query-driven, research-informed, end-to-end multimodal presentation video generation. Given a user query, the system autonomously identifies the core topic, retrieves relevant multimodal resources, and synthesizes a complete video incorporating slides, spoken narration, and dynamic media. It supports three distinct modes: solo presentation, multi-person discussion, and audience interaction. To evaluate performance across these scenarios, we introduce a comprehensive multimodal benchmark, on which our approach demonstrates strong results in content quality, media relevance, dynamic delivery, conversational naturalness, and interaction accuracy.