PaperlayerView exact revision on arXiv

Source-linked AI summary

PresentAgent-2: Towards Generalist Multimodal Presentation Agents

Wei Wu, Ziyang Xu, Zeyu Zhang, Yang Zhao, Hao Tang

arXiv:2605.11363v1cs.CVcs.CL

TL;DR

Existing presentation systems generally require complete source documents, whereas open-ended queries require research, multimodal retrieval, and coherent video construction. PresentAgent-2 addresses this with a query-driven framework supporting three presentation modes, achieving quiz scores of 4.84, 4.85, and 4.85 across them.

  • Problem

    Existing methods generally convert complete source documents into presentations, leaving query-to-presentation video generation from open-ended requests insufficiently addressed.

  • Method

    PresentAgent-2 researches user queries, retrieves multimodal resources, generates mode-specific slides and scripts, and composes them into videos across three presentation modes.

  • Results

    Quiz scores were 4.84, 4.85, and 4.85 for Single Presentation, Discussion Presentation, and Interaction Presentation, respectively.

  • Takeaways & Limitations

    PresentAgent-2 demonstrates informative, mode-aware presentation video generation from open-ended user queries.

  • Takeaways & Limitations

    Output quality depends on the availability and reliability of retrieved presentation-friendly sources, especially for queries with limited public multimodal resources.

Abstract

from arXiv · show

Presentation generation is moving beyond static slide creation toward end-to-end presentation video generation with research grounding, multimodal media, and interactive delivery. We introduce PresentAgent-2, an agentic framework for generating presentation videos from user queries. Given an open-ended user query and a selected presentation mode, PresentAgent-2 first summarizes the query into a focused topic and performs deep research over presentation-friendly sources to collect multimodal resources, including relevant text, images, GIFs, and videos. It then constructs presentation slides, generates mode-specific scripts, and composes slides, audio, and dynamic media into a complete presentation video. PresentAgent-2 supports three independent presentation modes within a unified framework: Single Presentation, which generates a single-speaker narrated presentation video; Discussion, which creates a multi-speaker presentation with structured speaker roles, such as for asking guiding questions, explaining concepts, clarifying details, and summarizing key points; and Interaction, which independently supports answering audience questions grounded in the generated slides, scripts, retrieved evidence, and presentation context. To evaluate these capabilities, we build a multimodal presentation benchmark covering single presentation, discussion, and interaction scenarios, with task-specific evaluation criteria for content quality, media relevance, dynamic media use, dialogue naturalness, and interaction grounding. Overall, PresentAgent-2 extends presentation generation from document-dependent slide creation to query-driven, research-grounded presentation video generation with multimodal media, dialogue, and interaction. Code: https://github.com/AIGeeksGroup/PresentAgent-2. Website: https://aigeeksgroup.github.io/PresentAgent-2.

1 Introduction

PresentAgent-2 addresses query-to-presentation video generation, where open-ended queries lack the complete content and visual resources needed for structured presentation videos. It combines deep research, multimodal resource retrieval, slide-and-script generation, and video composition in three presentation modes, evaluated by a multimodal benchmark.

  • Motivation: Existing methods usually convert complete documents into presentations, limiting their use for open-ended queries such as requests to explain a topic.These methods assume source content is already provided as a paper, report, or technical blog.
  • Motivation: Query-to-presentation video generation must explain a natural-language request despite lacking the full content and visual resources required for slide construction.The goal is to generate a structured presentation-style video from the query.
  • PresentAgent-2: PresentAgent-2 researches a focused topic, retrieves textual and multimodal resources, generates slides and scripts, and composes them into structured videos while preserving dynamic media.Retrieved resources include webpages, tutorials, demo pages, articles, images, GIFs, and videos.
  • Presentation modes: The unified framework supports Single Presentation, Discussion, and Interaction for single-speaker narration, multi-speaker dialogue, and grounded interactive Q&A.Discussion assigns roles such as asking guiding questions, explaining concepts, clarifying details, and summarizing key points.
  • Evaluation: A multimodal benchmark evaluates query-driven videos across single presentation, discussion, and interaction scenarios using presentation quality, media use, discussion quality, and interaction grounding.The benchmark assesses whether generated videos communicate knowledge through structure and multimodal presentation, not only factual correctness.

2 Related Work

Related work progresses from multimodal document summarization and layout-aware slide generation toward end-to-end presentation video systems. General multimodal generation and interactive visual instruction models provide enabling components, but are often evaluated as standalone modules.

  • Multimodal document summarization: Early presentation systems frame creation as multimodal document summarization with document understanding, content abstraction, and visual layout prediction.Doc2PPT establishes slide-quality evaluation criteria, while SlideGen and Paper2Poster improve slide or poster generation through multimodal agents and layout-aware organization.
  • General multimodal generation: General multimodal models contribute video, speech, temporal alignment, motion, long-sequence modeling, and multimodal evaluation capabilities for presentation synthesis.Interactive visual instruction models additionally support multimodal instruction following and visual question answering.
  • End-to-end presentation video generation: Recent systems move toward end-to-end presentation video generation by coordinating slide assembly, script generation, narration, subtitles, animation rendering, and audio-visual synchronization.PresentAgent converts long documents into narrated videos, while Paper2Video and VideoAgent create scientific explanation videos from academic papers.

3 PresentEval: A Multimodal Presentation Benchmark

PresentEval benchmarks query-to-presentation video generation across Single Presentation, Discussion, and Interaction modes using open-ended queries paired with human-created reference videos. It evaluates both knowledge delivery through objective quizzes and mode-specific presentation quality through subjective scoring.

  • Benchmark Setup: The benchmark evaluates three independent modes: Single Presentation, Discussion, and Interaction, using open-ended user queries rather than source documents.Each example provides a query and human-created reference presentation video, while systems receive only the query during generation.
  • Benchmark Setup: The benchmark contains 60 query–reference video pairs organized into three mode-specific sets of 20 examples each.Reference videos come from public video platforms, educational repositories, and professional presentation archives.
  • Evaluation Framework: Evaluation combines objective quiz assessment of query-relevant knowledge delivery with subjective scoring of presentation quality for the selected mode.Together, these components assess audience comprehension and mode-specific presentation requirements.
  • Objective Quiz Evaluation: For each query–reference pair, five four-option multiple-choice questions are constructed from the reference video and expected query knowledge points.The VLM answers the quizzes as an audience member, and scores are averaged across examples for each mode and model.
  • Subjective Mode-specific Evaluation: Subjective evaluation has a VLM judge independently score three mode-specific metrics on a 1–5 scale using the query, videos, retrieved resources, and generated transcript.The judge receives the generated and reference videos alongside retrieved resources and the transcript transcribed from generated audio.

4 Method: PresentAgent-2

PresentAgent-2 generates presentation videos from open-ended user queries by combining deep research over multimodal, presentation-friendly resources with shared slide and video generation. It supports Single Presentation, Discussion, and Interaction modes that vary in script and delivery style.

  • Query-to-presentation generation: PresentAgent-2 maps a natural-language query and selected mode to a presentation video, unlike document-to-presentation systems that require a complete source document.The system summarizes the query into a focused topic and retrieves multimodal resources before generation.
  • Deep research: Deep research retrieves presentation-friendly pages containing sufficiently complete text and rich media, including images, GIFs, and videos.Filtered resources are extracted into a multimodal set used to select content for the final slides and video.
  • Presentation generation: The presentation generator plans slide structure, uses textual resources for titles and explanations, inserts images, and preserves GIFs and videos as playable overlays during composition.Scripts are converted into audio, which is composed with slide visuals and dynamic media into the final presentation video.
  • Presentation modes: Single Presentation delivers slide-ordered narration from one speaker, while Discussion creates multi-speaker dialogue with roles for questioning, explanation, clarification, and summary.All modes share the same research, presentation generation, and video composition pipeline but differ in script and delivery style.
  • Presentation modes: Interaction adds audience question answering grounded in slides, scripts, and researched resources, with the system able to jump to the relevant slide.This mode allows users to participate during the presentation through detailed answers.

5 Experiments

PresentAgent-2 is evaluated across Single Presentation, Discussion Presentation, and Interaction Presentation using capability comparisons, benchmark scoring, and qualitative demonstrations. It achieves high quiz and subjective scores while adapting content delivery to each mode.

  • Experimental Scope: The experiments compare PresentAgent-2 with related systems across input settings, presentation modes, and multimodal resource support.The evaluation covers the three presentation modes and uses objective quiz evaluation alongside subjective mode-specific quality assessment.
  • Capability Analysis: PresentAgent-2 targets open-ended query-to-presentation generation with deep research, unlike systems designed for document-to-presentation, paper-to-video, poster, or general video generation.The system starts from a user query, researches relevant content, and generates presentation videos.
  • Benchmark Evaluation: 4.84, 4.85, and 4.85 are the quiz scores for Single Presentation, Discussion Presentation, and Interaction Presentation using the Qwen3.5-VL-Plus backbone.Quiz scores are averaged on a 0–5 scale.
  • Benchmark Evaluation: 4.47, 4.37, and 4.52 are the mean subjective scores for Single Presentation, Discussion Presentation, and Interaction Presentation, respectively.These results indicate knowledge delivery and mode-aware video generation from user queries; subjective scores use a 1–5 scale.
  • Mode-Specific Results: Single Presentation produces coherent explanations, Discussion Presentation creates complementary multi-speaker dialogue, and Interaction Presentation supports effective audience question answering.The outputs combine retrieved textual and multimodal resources, natural conversational delivery, and generated presentation context.

6 Conclusion

PresentAgent-2 transforms open-ended user queries into multimodal presentation videos through an integrated research, generation, synthesis, and composition pipeline. It supports three presentation modes and introduces PresentEval for objective knowledge delivery and subjective mode-specific quality.

  • Framework: PresentAgent-2 is a query-to-presentation video generation agent that transforms open-ended user queries into multimodal presentation videos.The framework is designed for end-to-end presentation video generation rather than only static slide creation.
  • Framework: The framework integrates deep research, multimodal resource retrieval, slide and script generation, audio synthesis, and video composition.These components form a unified pipeline from query understanding and evidence gathering to completed presentation videos.
  • Presentation modes: PresentAgent-2 supports Single Presentation, Discussion Presentation, and Interaction Presentation modes.The modes cover single-speaker delivery, multi-speaker discussion, and audience interaction within the same framework.
  • Evaluation: PresentEval evaluates objective knowledge delivery and subjective mode-specific quality.The benchmark is designed to assess both content delivery and quality criteria specific to each presentation mode.

A Additional Qualitative Examples

Figures 6 and 7 provide additional qualitative examples of PresentAgent-2, showing representative outputs across generated video frames, slides, scripts, and interaction screenshots.

  • Qualitative Examples: Figures 6 and 7 organize each example by column, with rows showing a representative video frame, generated slide, generated script, and interaction screenshot.The examples cover PresentAgent-2’s Single Presentation and Interaction Presentation outputs.

B Evaluation Prompts and Rubrics

The evaluation combines objective quiz scoring with subjective VLM judgments of presentation quality. Quizzes test knowledge communicated through the generated video and transcript, while subjective rubrics assess mode-specific content, research use, delivery, dialogue, or interaction.

  • B.1 Objective Quiz Format and Scoring: Each query–reference video pair has five four-option multiple-choice questions with one correct answer, constructed from the reference video and expected knowledge points.The quiz record includes the example ID, presentation mode, questions, options, answer key, and expected knowledge point.
  • B.1 Objective Quiz Format and Scoring: Quiz scoring uses exact answer matching, awarding one point per correct response and reporting the average score over examples on a 0–5 scale.Each example contains five questions, so its score ranges from 0 to 5.
  • B.2 Objective Quiz Answering Prompt: For objective answering, the VLM acts as an audience member and may use only the generated presentation video and its transcribed audio transcript.The prompt supplies the user query, generated video, transcript, and quiz questions while excluding the reference video and external information.
  • B.2 Objective Quiz Answering Prompt: The objective prompt forbids reference answers, expected knowledge points, and external assumptions, requiring exactly one choice among A, B, C, or D in valid JSON.Answers must rely only on information clearly communicated in the generated video or transcript.
  • B.3 Subjective Scoring Prompt: For subjective evaluation, a VLM judge independently assigns 1–5 scores to three metrics for the presentation mode using the query, generated and reference videos, retrieved resources, and transcript.The rubric defines 1 as poor performance and 3 as acceptable but incomplete performance.
  • B.3 Subjective Scoring Prompt: Single-presentation criteria assess query answering, deep-research effectiveness, and video delivery quality across content coverage, resource use, narration, visuals, and dynamic media.Delivery quality emphasizes coherence, clarity, and ease of following the combined narration, visual examples, and dynamic media.
  • B.3 Subjective Scoring Prompt: Discussion and interaction criteria evaluate dialogue structure and naturalness, speaker-role complementarity, grounded answer effectiveness, comprehensibility, and helpfulness.Discussion rubrics cover questions, clarifications, comparisons, supplementary explanations, and fluent turntaking; interaction rubrics emphasize presentation-grounded audience answers.

C Implementation Details

PresentAgent-2 implements a research-to-video pipeline that cleans and filters multimodal web resources before planning slides, generating scripts and narration, and composing presentation videos. Its shared backbone supports distinct single-speaker, multi-speaker discussion, and audience-facing interaction formats, while preserving retrieved dynamic media during composition.

  • Retrieval and filtering: The system summarizes each query into a focused topic, researches presentation-friendly sources, cleans HTML pages, and retains sources with complete text and rich multimodal materials.Cleaning removes boilerplate, navigation, advertisements, and fragmented text; retained materials may include images, GIFs, and videos.
  • Presentation generation: PresentAgent-2 plans slides, generates slide content and mode-specific scripts, synthesizes narration, and composes slides, audio, and retrieved visual resources into a final video.Text supports titles, bullet points, and explanatory scripts, while visual resources are inserted into corresponding slide regions for visual grounding.
  • Dynamic media: GIFs and videos remain dynamic during composition, placed in relevant slide regions and synchronized with narration and the slide sequence.This preserves moving demonstrations, animations, and visual examples instead of converting them into static screenshots.
  • Presentation modes: All three modes share retrieval and presentation generation, but use different delivery formats: single-speaker narration, multi-speaker dialogue, or audience-facing question answering.Discussion roles include asking guiding questions, explaining concepts, clarifying details, and summarizing key points.

D Ablation Study

The ablation study isolates shared resource components, discussion-mode role-aware dialogue generation, and interaction context grounding, using a common backbone and evaluation protocol. It tests multimodal evidence, dynamic media, complementary speaker roles, and presentation-context grounding.

  • Study Design: Ablations use the same backbone model and evaluation protocol as the main experiments, covering shared resources, discussion mode, and interaction grounding.The three groups analyze key design choices in PresentAgent-2.
  • Shared Resource Ablation: Text-only Retrieval removes retrieved images, GIFs, and videos, retaining only textual resources for slide and script generation.This tests the contribution of multimodal evidence beyond text-only retrieval.
  • Shared Resource Ablation: Static-media retains retrieved visual resources but converts GIFs and videos into static frames during video composition.This tests whether preserving dynamic visual resources benefits query-driven presentations.
  • Discussion-Mode Ablation: Random Script Splitting assigns sentences to two virtual speakers instead of using complementary roles for question guidance, explanation, clarification, and summarization.The ablation tests whether discussion benefits from explicit speaker-role complementarity rather than monologue splitting.
  • Interaction Grounding Ablation: Context-Free Interaction removes grounding in structured slides, speaker scripts, and retrieved evidence from audience-oriented responses.The full interaction mode grounds responses in the complete presentation context.

E Limitations

PresentAgent-2 remains limited by the availability and reliability of retrieved presentation-friendly sources, upstream errors affecting interaction answers, and the benchmark’s limited coverage. Its 60 query–reference video pairs do not exhaust presentation domains, audience types, or interaction scenarios.

  • Source and retrieval limitations: Output quality depends on retrieved sources; limited or low-quality multimodal resources can reduce visual evidence and explanation comprehensiveness.Queries with limited public multimodal resources or low-quality search results may produce less informative visual evidence or less comprehensive explanations.
  • Source and retrieval limitations: Interaction answers depend on generated slides, scripts, retrieved evidence, and presentation context, so upstream errors can reduce their correctness or helpfulness.Errors in retrieval, slide generation, or script generation may propagate to the interaction stage.
  • Benchmark limitations: 60 query–reference video pairs span three presentation modes but do not cover all domains, audience types, or interaction scenarios.Future work can add more domains, longer presentations, and more fine-grained human evaluation.
Loading 2605.11363v1…