PresentAgent: Modular Multimodal Presentation System

• PresentAgent is a modular multimodal agent system that converts long-form documents into fully narrated, slide-based videos with high controllability and interpretability.
• The pipeline integrates document segmentation, LLM-driven slide planning, TTS narration, and precise audio-visual synchronization to produce human-like presentations.
• Its PresentEval framework objectively benchmarks content fidelity, visual clarity, and audience comprehension, influencing subsequent research in automated presentation generation.

PresentAgent is a modular multimodal agent system for converting long-form documents into fully narrated, slide-based presentation videos. It unifies LLMs, vision-LLMs (VLMs), and text-to-speech (TTS) technology in a pipeline designed for high controllability, interpretability, and human-style audio-visual synchronization. The PresentAgent framework has influenced subsequent research in automated presentation video generation and multimodal agent architectures (Shi et al., 5 Jul 2025).

1. System Architecture and Pipeline

PresentAgent’s architecture consists of four main processing stages:

1. Document Segmentation: The input text document $D$—such as a scientific paper, web page, or technical report—is segmented into $K$ coherent content chunks $C_k$ via outline planning. A lightweight LLM generates a top-level outline that specifies section titles and paragraph spans. Optionally, an embedding-based topical shift detection mechanism can refine segmentation, delineating segment boundaries at significant embedding cosine drops ($s_i = \cos(e_i, e_{i+1}) < \tau$).
2. Slide Planning & Rendering: For each segment $C_k$, PresentAgent selects an appropriate slide type (e.g., bullets, chart, figure + caption) by prompting an LLM. It then generates detailed editing instructions (sequences of operations from a set $\mathcal{O}$ such as $\mathtt{replace\_text}, \mathtt{insert\_image}$, etc.) to populate a selected slide template $T_{\rm type}$. Python-pptx or an equivalent HTML engine executes these structured operations to render static slide frames $S_k$ as images.
3. Narration Generation: Each content segment $C_k$ is rewritten as a spoken script $K$0 using an LLM, with prompts enforcing conversational tone and controlling narration length (e.g., direct targets or soft penalties on word count). The scripts are synthesized into 24 kHz, 16-bit audio $K$1 with MegaTTS3, a TTS model trained to minimize spectrogram reconstruction loss.
4. Audio–Visual Composition: Audio clips $K$2 and slide images $K$3 are precisely aligned in the generated video: slide $K$4 is displayed for exactly the duration of $K$5. The ffmpeg utility concatenates frame–audio pairs with transitions. The output video is formally $K$6.

This pipeline enables videos to “look like” human presentations, combining coherent slide structure, conversational narration, and tight temporal alignment. Intermediate outlines, scripts, and slide previews are exposed for greater interpretability and granular user control (Shi et al., 5 Jul 2025).

2. Slide Design: Template and Layout Optimization

Slide planning relies on LLMs for both type selection and detailed layout population. For each content chunk:

• An LLM determines $K$7.
• Operations from $K$8 are composed to edit the slide—these instructions are parsed into editing scripts and executed by Python-pptx/HTML engines.

Optional visual scoring leverages VLMs to score slide images $K$9:

$C_k$0

where $C_k$1 computes key term coverage, $C_k$2 is a VLM-based visual aesthetics score, and $C_k$3 penalizes excessive text density. This schema enables balancing informativeness, aesthetics, and readability (Shi et al., 5 Jul 2025).

3. Narration Generation and Audio Synthesis

The narration subsystem ensures that each slide has a matching spoken script $C_k$4 with tight length control. Script synthesis uses explicit prompt instructions for conversational tone and target duration; further control is achieved through loss penalties on length during LLM fine-tuning:

$C_k$5

MegaTTS3 synthesizes these scripts into audio, optimizing spectrogram fidelity:

$C_k$6

where $C_k$7 and $C_k$8 are ground-truth and predicted mel-spectrograms, respectively. The output maintains high-fidelity and natural prosody (Shi et al., 5 Jul 2025).

4. Evaluation Methodology: The PresentEval Framework

PresentAgent introduces PresentEval—a comprehensive, VLM-powered evaluation framework that measures:

• Content Fidelity: Objective factual coverage, measured via quiz accuracy. For each video–document pair, a VLM answers $C_k$9 multiple-choice questions; accuracy is

$s_i = \cos(e_i, e_{i+1}) < \tau$0

• Visual Clarity: Subjective assessment of slide design and aesthetics, using VLMs with prompts for visual appeal and coherence.
• Audience Comprehension: Subjective evaluation of how easily the content is understood (e.g., via “Comprehension Difficulty” prompts).

The mean of subjective VLM scores across dimensions provides an aggregate measure. This multi-pronged evaluation enables both objective and subjective benchmarking across content and delivery axes (Shi et al., 5 Jul 2025).

5. Quantitative Results and Comparative Analysis

Experiments conducted on the Doc2Present benchmark (30 document–presentation pairs, four domains; documents with 3,000–8,000 words) demonstrated that PresentAgent-generated videos (5–10 slides, 1–2 min) approach or surpass human references in several metrics. The table below summarizes key results from evaluation on a 5-document subset:

Method	Model	QuizAcc	Video Mean	Audio Mean
Human (reference)	Human	0.56	4.47	4.80
PresentAgent	Claude-3.7-Sonnet	0.64	4.00	4.53
PresentAgent	Qwen-VL-Max	0.52	4.47	4.60
PresentAgent	GPT-4o-Mini	0.64	4.67	4.40

• PresentAgent variants can exceed human quiz accuracy (0.64 vs. 0.56), supporting strong factual coverage.
• Subjective scores for video and audio quality are close to human, with the highest PresentAgent video mean (4.67) marginally exceeding the reference (4.47).
• Some trade-offs arise: higher visual design scores can correlate with minor decreases in comprehension score.
• Limitations include generation of static (non-animated) slides and restricted benchmark size due to API costs (Shi et al., 5 Jul 2025).

6. Limitations and Future Directions

PresentAgent’s pipeline is inherently modular and interpretable, but current implementations are limited to static visual frames (no animations) and small-scale evaluations (limited to five documents in quantitative benchmarks due to API constraints). Scalability, dynamic media integration, and real-time or interactive features are not supported in PresentAgent itself.

Subsequent systems have incorporated more sophisticated slide generation (e.g., tree-search layout optimization (Zhu et al., 6 Oct 2025)), integration of multimodal media and dialogue (Wu et al., 12 May 2026), and support for interactive presentations. These advances represent a direct extension of PresentAgent’s pipeline design and evaluation philosophy but introduce additional technical components and benchmark coverage.

7. Impact and Context within Multimodal Agent Research

PresentAgent established a rigorous pipeline for document-to-presentation video generation, providing a clear agentic template—semantic document segmentation, LLM-driven layout planning, controlled script synthesis, and precise audio-visual composition. Its PresentEval framework set a precedent for multimodal, VLM-driven evaluation, influencing subsequent research on automatic presentation video generation and multimodal agent evaluation.

Later works (e.g., PaperTalker (Zhu et al., 6 Oct 2025), PresentAgent-2 (Wu et al., 12 May 2026)) built on these concepts by adding agentic specialization (e.g., dedicated slide, subtitle, and talker builders), multimodal resource integration, dialogue/interaction modes, and improved evaluation criteria reflecting dynamic, research-grounded, and interactive media capabilities.

PresentAgent’s modular approach and focus on alignment between language, vision, and speech continue to inform research in automated knowledge dissemination, educational technology, and agentic systems for human-computer presentation interfaces.