PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation

Hao Zheng, Guozhao Mo, Xinru Yan, Qianhao Yuan, Wenkai Zhang, Xuanang Chen, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun

arXiv:2602.22839v3cs.AI

TL;DR

Existing presentation agents are limited by rigid workflows, fixed templates, and reflection that misses post-render defects. DeepPresenter coordinates research and design agents with environment-grounded reflection, and achieves state-of-the-art performance while its 9B model remains competitive at lower cost.

  • Problem

    Existing presentation agents rely on predefined workflows and fixed templates, while internal-signal reflection cannot detect many post-render presentation defects.

  • Method

    DeepPresenter coordinates Researcher and Presenter agents and grounds iterative self-correction in perceptual artifact states obtained through environmental observation.

  • Results

    DEEPPRESENTER achieves an average score of 4.44, surpassing Gamma (4.36), while DeepPresenter-9B scores 4.19 and approaches GPT-5 (4.22) at lower cost.

  • Takeaways & Limitations

    The framework supports adaptive presentation generation and feedback-driven refinement beyond rigid scripted workflows, with a lower-cost compact model remaining competitive.

  • Takeaways & Limitations

    DEEPPRESENTER relies on multi-step tool-using rollouts that increase inference cost and are sensitive to environment instability.

Abstract

from arXiv · show

Presentation generation requires deep content research, coherent visual design, and iterative refinement based on observation. However, existing presentation agents often rely on predefined workflows and fixed templates. To address this, we present DeepPresenter, an agentic framework that adapts to diverse user intents, enables effective feedback-driven refinement, and generalizes beyond a scripted pipeline. Specifically, DeepPresenter autonomously plans, renders, and revises intermediate slide artifacts to support long-horizon refinement with environmental observations. Furthermore, rather than relying on self-reflection over internal signals (e.g., reasoning traces), our environment-grounded reflection conditions the generation process on perceptual artifact states (e.g., rendered slides), enabling the system to identify and correct presentation-specific issues during execution. Results on the evaluation set covering diverse presentation-generation scenarios show that DeepPresenter achieves state-of-the-art performance, and the fine-tuned 9B model remains highly competitive at substantially lower cost. Our project is available at: https://github.com/icip-cas/PPTAgent

1 Introduction

DeepPresenter addresses the limited adaptability and post-render defect detection of existing presentation agents with coordinated research, content-driven design, and environment-grounded reflection. It achieves state-of-the-art performance, while DeepPresenter-9B remains competitive at substantially lower cost.

  • Motivation: Existing agents rely on predefined workflows and content-agnostic templates, limiting adaptation to user intents and producing shallow, text-heavy, visually mismatched slides.Introspective reflection over internal signals also cannot detect defects that appear only after rendering.
  • Efficient Model: DeepPresenter-9B is trained through supervised fine-tuning on curated trajectories synthesized with extrinsic verification.An independent critic evaluates artifacts and provides reasoning traces that steer targeted refinements during trajectory synthesis.
  • Framework: DeepPresenter coordinates Researcher and Presenter agents through a shared observation space for autonomous information research and topic-aware design.The Researcher compiles a structured manuscript, while the Presenter converts it into visually coherent slides through content-driven design.
  • Reflection: Environment-grounded reflection uses post-render perceptual artifact states to identify and correct defects that internal signals alone cannot detect.Agents inspect rendered manuscripts and slides, then plan targeted revisions.
  • Results: DEEPPRESENTER achieves state-of-the-art performance on diverse presentation-generation scenarios, while the distilled DeepPresenter-9B remains highly competitive at substantially lower cost.The evaluation covers held-out tasks across constraint satisfaction, content quality, and visual style.

2 DEEPPRESENTER

DEEPPRESENTER formulates presentation generation as an interactive process in which specialized Researcher and Presenter agents collaborate through shared files and environmental observations. Its reflection loop aligns agent feedback with rendered artifacts and user perception.

  • Task Formulation: Presentation generation is modeled as a multi-step trajectory in which agents produce reasoning traces, select tools, and receive environment observations.The trajectory is divided into sequential Researcher and Presenter phases communicating through the file system.
  • Dual-Agent Collaboration: Two specialized agents split information research and visual design while sharing the same backbone model.This separation reflects their different planning and tool-use requirements.
  • Researcher Agent: The Researcher autonomously adapts exploration depth and strategy to user intent, synthesizes supporting materials, and saves a structured manuscript with associated assets.Technical topics may require related-work surveys, whereas general-audience talks may prioritize accessible examples and illustrations.
  • Presenter Agent: The Presenter generates slides from scratch using a global design plan and translates manuscript content into standalone HTML files rather than populating predefined templates.Design choices such as palettes and layouts can align with the presentation topic.
  • Environment-Grounded Reflection: Environment-grounded reflection addresses state mismatch by exposing rendered artifacts and structured diagnostics through inspect, followed by think-driven targeted revisions.This observe–reflect–revise loop reveals defects such as overflow, overlap, low contrast, missing assets, and language or slide-count issues.

3 Frontier Presentation Agent Model

The training pipeline builds diverse constrained presentation tasks, synthesizes trajectories with independent verification, and filters them for compliance, consistency, and output quality. This process supports compact-model training and evaluation across varied presentation scenarios.

  • Pipeline: The pipeline combines task dataset construction, extrinsic-verification trajectory synthesis, and multi-stage quality filtering.These mechanisms are designed to produce high-quality trajectories for supervised fine-tuning.
  • Query Construction: Tasks draw from PersonaHub, arXiv, and FinePDFs-Edu to cover intent-driven and document-conditioned presentation scenarios.Each task includes verifiable requirements such as slide count, language, and aspect ratio.
  • Extrinsic Verification: Self-reflection can produce self-verification bias because agents judge outputs from the same trajectory state that generated them.This coupling can cause flawed outputs to be accepted.
  • Extrinsic Verification: An independent critic evaluates inspected artifacts, identifies defects, and supplies actionable reasoning traces that guide targeted revisions during sampling.The trace is appended as a think call before the rollout continues.
  • Trajectory Filtering: The filtering pipeline checks constraint compliance, removes reflection–action inconsistencies, and excludes outputs with critical defects such as overlap or broken images.Rule-based checking and GLM-4.6 or GLM-4.6V assessments support the three filtering stages.

4 Experiment

The evaluation measures constraint satisfaction, content quality, visual style, and diversity across presentation-generation systems. DEEPPRESENTER leads on average performance and diversity, while DeepPresenter-9B remains competitive with fewer training trajectories.

  • Evaluation Setup: The evaluation holds out 128 tasks and scores constraint satisfaction, content, style, and diversity across frameworks and backbone models.Avg. is the mean of Constraint, Content, and Style on a 0–5 scale; Diversity uses a 0–1 scale.
  • Main Results: 4.44 average score makes DEEPPRESENTER the strongest evaluated system, exceeding Gamma’s 4.36 and the best open-source baseline’s 3.92.With Gemini-3-Pro, the 4.44 score is 13.3% above KCTV + Claude-Sonnet-4.5.
  • Main Results: 0.79 diversity score more than doubles template-based baselines scoring 0.17–0.35, while maintaining high constraint compliance.Gamma reaches 0.52 on the same diversity metric.
  • Ablation Study: Disabling environment-grounded reflection reduces performance from 4.44 to 4.32 with Gemini-3-Pro and from 4.19 to 3.82 with DeepPresenter-9B.The ablation removes post-render inspection from the reflection process.
  • Ablation Study: Removing trajectory filtering lowers DeepPresenter-9B from 4.19 to 4.03, supporting filtering as part of the fine-tuning strategy.The ablation also reports substantial drops without dual-agent collaboration.

5 Analysis

The analysis examines extrinsic verification, trajectory failures, cost-performance trade-offs, and qualitative outputs. Extrinsic verification strengthens revision signals, while DeepPresenter-9B improves cost-efficiency and DEEPPRESENTER produces richer, more topic-aligned slides.

  • Extrinsic Verification: Extrinsic verification produces a 67% larger Avg. gain than fine-tuning alone, improving by 0.20 versus 0.12 on matched 300-trajectory variants.The comparison targets self-verification bias during trajectory synthesis.
  • Extrinsic Verification: Extrinsic verification identifies more manuscript and slide defects by evaluating artifacts in an isolated context rather than within the generating trajectory.Defects include integrity, constraint, format, layout, render, and style categories.
  • Failure Analysis: 43.0% of pre-filtering trajectory failures are quality errors and 32.3% are environment failures, with constraint and consistency errors accounting for 13.5% and 11.2%.The distribution highlights quality problems and long-horizon fragility as the most frequent failure categories.
  • Efficiency Analysis: DeepPresenter-9B reaches 4.19 at a similar price to KCTV + Gemini-3-Pro’s 3.83, advancing the cost-performance Pareto frontier.DEEPPRESENTER also exceeds Gamma, scoring 4.44 versus 4.36.
  • Efficiency Analysis: DEEPPRESENTER’s performance varies from 3.82 to 4.44 across backbones, whereas baseline frameworks remain comparatively flat.The analysis attributes this pattern to DEEPPRESENTER leveraging stronger model capacity more effectively than fixed pipelines.
  • Qualitative Analysis: DEEPPRESENTER produces visually rich, topic-aligned slides, while baselines more often generate text-heavy outputs with template-based or mismatched visual themes.Gamma uses more imagery but can mishandle figures embedded in source documents.

6 Related Work

Presentation generation evolved from document summarization toward agent-based systems addressing content, layout, aesthetics, and slide editing. DEEPPRESENTER extends this direction through autonomous exploration, specialized agents, and environment-grounded reflection.

  • Prior Paradigms: Pre-LLM presentation-generation methods primarily treated the task as document summarization through extractive or phrase-based selection.Their limited reasoning capabilities constrained handling of diverse user intents and visually engaging outputs.
  • LLM-Based Methods: Recent LLM-based work explores multi-agent collaboration, aesthetic-aware generation, and slide understanding and editing.These directions broaden presentation generation beyond sentence selection.
  • DEEPPRESENTER: DEEPPRESENTER formulates presentation generation as autonomous exploration and collaboration between specialized Researcher and Presenter agents.The decomposition supports adaptive planning based on task complexity.
  • DEEPPRESENTER: Environment-grounded reflection lets agents verify and refine artifacts through rendered slides and file-system states rather than relying only on internal trajectory signals.This provides perceptual feedback during presentation generation.

7 Conclusion

DEEPPRESENTER autonomously adapts presentation generation to diverse user intents and grounds reflection in perceptual artifact states. Its fine-tuned DeepPresenter-9B variant remains competitive at substantially lower cost.

  • Conclusion: DEEPPRESENTER combines autonomous planning, perceptual artifact-based reflection, and iterative correction of post-render defects.The framework is designed to adapt to diverse user intents.
  • Conclusion: DeepPresenter-9B is trained on trajectories synthesized with extrinsic verification to mitigate self-verification bias and strengthen reflective behaviors.The conclusion reports that the compact model remains competitive at substantially lower cost.

Limitations

DeepPresenter’s limitations concern rollout cost and stability, the absence of external-critic verification at inference, and uneven performance across data domains.

  • System limitations: Multi-step tool-using rollouts increase inference cost and are sensitive to context overflow and infrastructure failures.These environment-instability issues were observed in trajectory analysis.
  • System limitations: Extrinsic verification is used only during trajectory synthesis, not inference, because external critic signals may create reflection–action inconsistency and additional overhead.
  • Evaluation scope: A human study evaluated 32 randomly sampled presentations to corroborate the automatic assessments.Two computer-science graduate students rated content and style on a 1–5 Likert scale, while constraint satisfaction was verified separately.
  • Domain variation: PersonaHub achieves stronger content and style scores, whereas arXiv achieves near-perfect constraint satisfaction but lower content and style scores.PersonaHub scores 4.49 for both content and style, while arXiv scores 4.91 for constraint satisfaction, 3.84 for content, and 4.13 for style.

A.3 Tool Usage Analysis

Tool usage differs by agent role and domain: Researchers retrieve information, while Presenters manipulate files and reason during slide refinement; source availability shifts these patterns.

  • Agent roles: Researchers rely heavily on Retrieve tools, whereas Presenters focus on File operations and Reason tools for iterative slide editing and reflection.The role-specific preferences align with the responsibilities of the two specialized agents.
  • Domains: PersonaHub tasks show significantly higher Retrieve usage because they lack reference documents and require searching for relevant materials.
  • Domains: arXiv and FinePDF tasks show higher File usage and lower retrieval reliance because they provide source documents for processing.

B.1 Data Sources

The evaluation data combines academic, educational, and personalized presentation scenarios, with generated requests filtered for quality and augmented by verifiable constraints.

  • Academic presentations: Academic tasks pair arXiv papers with audience-specific requests spanning lectures, seminars, defenses, and conference talks.Audiences range from beginners and intermediate learners to domain experts and peer researchers.
  • Educational presentations: FinePDFs-Edu supplies English and Chinese PDF documents for general educational presentations based on attached materials.
  • Personalized presentations: PersonaHub supports personalized scenarios by generating presentation requests from user personas with knowledge-grounded and open-ended strategies.Qwen3-235B-A22B generates realistic requests, with language selected according to cultural background.
  • Data quality: Generated PersonaHub queries undergo language filtering, semantic deduplication, and LLM-based quality control to remove unsuitable samples.
  • Constraint design: Each task receives randomly assigned constraints on slide count, aspect ratio, and language, scored by rule-based PDF verification.Constraint satisfaction is the proportion of specified constraints successfully met.
  • Evaluation split: The evaluation split contains 128 tasks, with composition and statistics disclosed for replication.

C Agent Framework

The agent framework organizes heterogeneous presentation resources into tool categories and coordinates content research, visual design, inspection, and iterative revision through environment observations.

  • Tool organization: The toolset covers Retrieve, File, Reason, Control, and Synthesis categories for information gathering, document manipulation, reflection, task management, and execution.
  • Inspection tools: Reason inspection tools ground reflection in artifact states by diagnosing manuscripts and rendering slides for visual inspection.
  • Inspection tools: Inspect slide exposes contrast problems and element overflow by returning rendered HTML slides as pixel images in the agent’s visual context.
  • Execution control: Each task proceeds through reasoning–action–observation steps within a 50K-token context window, with warnings at 50% and 80% capacity.
  • Researcher agent: The Researcher conducts systematic information retrieval, builds a narrative slide framework, and writes a Markdown manuscript while collecting or creating visual materials.Its task excludes slide layout and design, and it uses reasoning to reflect on state and next steps.
  • Presenter agent: The Presenter converts manuscripts into fixed-layout HTML/CSS slides through a slide-master plan specifying style, grid, fonts, and dimensions.It generates slides page by page, uses locked dimensions with hidden overflow, and finalizes the slide folder.
Loading 2602.22839v3…