DeepSlide: A Human-in-the-Loop Multi-Agent System for Full Presentation Preparation (arXiv 2505.10892)

Forum topic · 小凯 · 2026-05-19

Summary

DeepSlide (arXiv:2505.10892) is a human-in-the-loop multi-agent system that addresses a gap in AI slide generation: most tools optimize the artifact (a visually plausible deck) while neglecting the delivery process, including pacing, narrative flow, and rehearsal. DeepSlide supports the complete presentation preparation pipeline, from requirement elicitation and time-budgeted narrative planning to evidence-grounded slide-script generation, attention augmentation, and rehearsal support. Its architecture integrates four components: (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to guarantee renderability. The authors also introduce a dual-scorecard benchmark that cleanly separates static artifact quality from dynamic delivery performance. Evaluated across 20 domains and diverse audiences, DeepSlide matches strong baselines on artifact quality while achieving consistently larger gains on delivery metrics, improving narrative fluency, pacing accuracy, and slide-script alignment.

Paper Overview

Field: NLP Authors: Ming Yang, Zhiwei Zhang, Jiahang Li Published: 2025-05-15 arXiv: 2505.10892

Introduction

Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize only the *artifact* — a visually plausible deck — while under-optimizing the *delivery process*: pacing, narrative, and presentation preparation.

What DeepSlide Does

DeepSlide is a human-in-the-loop multi-agent system that supports preparing the full presentation process:

Requirement elicitation
Time-budgeted narrative planning
Evidence-grounded slide–script generation
Attention augmentation
Rehearsal support

System Architecture

DeepSlide integrates four key components:

1. Controllable logical-chain planner with per-node time budgets 2. Lightweight content-tree retriever for grounding 3. Markov-style sequential rendering with style inheritance 4. Sandboxed execution with minimal repair to ensure renderability

Evaluation

The authors introduce a dual-scorecard benchmark that cleanly separates static artifact quality from dynamic delivery performance.

Across 20 domains and diverse audiences, DeepSlide:

Matches strong baselines on artifact quality
Achieves consistently larger gains on delivery metrics
Improves narrative fluency, pacing accuracy, and slide–script alignment

Abstract (Original)

> Presentations are a primary medium for scholarly communication, yet most AI slide generators optimize the artifact (a visually plausible deck) while under-optimizing the delivery process (pacing, narrative, and presentation preparation). We present DeepSlide, a human-in-the-loop multi-agent system that supports preparing the full presentation process, from requirement elicitation and time-budgeted narrative planning, to evidence-grounded slide–script generation, attention augmentation, and rehearsal support. DeepSlide integrates (i) a controllable logical-chain planner with per-node time budgets, (ii) a lightweight content-tree retriever for grounding, (iii) Markov-style sequential rendering with style inheritance, and (iv) sandboxed execution with minimal repair to ensure renderability. ...

*Full paper: arXiv:2505.10892*