DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation

Hao Zheng; Guozhao Mo; Xinru Yan; Qianhao Yuan; Wenkai Zhang; Xuanang Chen; Yaojie Lu; Hongyu Lin

cs.AIFeb 26, 2026

DeepPresenter: Environment-Grounded Reflection for Agentic Presentation Generation

Authors: Hao Zheng, Guozhao Mo, Xinru Yan, Qianhao Yuan, Wenkai Zhang, Xuanang Chen, Yaojie Lu, Hongyu Lin,

Organizations: Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China

Abstract

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

Figures & tables

Figure 1: Illustration of DeepPresenter . Given a user instruction, the Researcher gathers information and compiles a structured manuscript, while the Presenter transforms it into visual slides. Both agents interact and collaborate with a shared environment, leveraging grounded observations for reflective refinement.

Figure 2: Comparison between self-reflection and environment-grounded reflection. Self-reflection relies on uncertain triggers and inputs without external signals. DeepPresenter grounds reflection in environmental observations through the inspect tool.

Dimension	Category	Count	Ratio (%)
Language	English	603	52.34
Language	Chinese	549	47.66
Source	PersonaHub	586	50.87
	FinePDFs	362	31.42
	arXiv	204	17.71
Aspect Ratio	16:9 Widescreen	327	28.39

Table 1: Statistics of the constructed presentation tasks by language, source, aspect ratio, and slide count. “Free” indicates no constraint is specified.

Figure 3: Our data synthesis pipeline. The process ensures high-quality trajectories for supervised fine-tuning through three integrated mechanisms: (1) Query Construction augments tasks with verifiable constraints; (2) Extrinsic Verification injects reasoning traces when defects are identified to guide agent self-correction during sampling; and (3) Trajectory Filtering validates constraint compliance and assesses consistency and output quality.

Close-sourced Baseline
Framework	Model	Constraint	Content	Style	Avg.	Diversity
Gamma	–	4.93	4.08	4.08	4.36	0.52
Open-sourced Baseline
PPTAgent	GPT-5	3.96	3.00	4.07	3.68	0.35
	Gemini-3-Pro	4.22	3.09	4.30	3.87	0.19
	Claude-Sonnet-4.5	3.72	2.93	4.15	3.60	0.17

Table 2: Performance comparison of different frameworks and models. The best/second-best scores are bolded / underlined . Quality metrics (Constraint, Content, Style, Avg.) are scaled to 0–5, while Diversity is scaled to 0–1.

Gemini-3-Pro
Configuration	Cons.	Content	Style	Avg.
DeepPresenter	4.70	4.25	4.37	4.44
w/o Grounded Reflection	4.52	4.15	4.31	4.32
w/o Dual-Agent	3.94	3.96	4.22	4.04
DeepPresenter-9B
DeepPresenter-9B	4.77	3.52	4.29	4.19

Table 3: Ablation study on framework components and training strategy. Cons. denotes constraint satisfaction.

Configuration	Cons.	Content	Style	Avg.	$\Delta$
GLM-4.6V-Flash	4.67	3.11	3.69	3.82	–
+ Fine-tuning	4.71	3.19	3.92	3.94	+0.12
+ Extrinsic Verification	4.74	3.28	4.03	4.02	+0.20

Table 4: Effect of extrinsic verification on model performance. Both fine-tuned variants use 300 trajectories.

\Delta

denotes improvement over the base model.

Figure 4: Distribution of defects identified by self-verification and extrinsic verification for manuscripts (left) and slides (right), respectively.

Figure 5: Failure distribution in synthesized trajectories before filtering

Figure 6: Performance vs. Price scatter plot with Pareto frontier representation. Different colors represent different frameworks

Figure 7: Qualitative comparison of presentations generated by different methods. DeepPresenter under Gemini-3-Pro and DeepPresenter-9B produce high-quality slides with styles that resonate with the topic. Baselines rely on document-embedded or AI-generated images with template-based generation, producing text-heavy outputs and misaligned visual themes.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Cons.	Content	Style	Avg.
Gamma	4.84	3.52	3.90	4.09
PPTAgent	3.72	3.07	3.60	3.46
KCTV	4.41	2.84	3.19	3.48
DeepPresenter	4.56	3.86	4.25	4.22

Appendix

Table 5: Human evaluation results on 32 randomly sampled presentations.

Domain	Cons.	Content	Style	Avg.
PersonaHub	4.38	4.49	4.49	4.45
arXiv	4.91	3.84	4.13	4.29
FinePDF	4.94	4.21	4.38	4.51

Appendix

Table 6: Domain performance breakdown. Cons. denotes constraint satisfaction.

Figure 8: Tool usage analysis. (a) Distribution of tool invocations by agent role. (b) Tool usage patterns of Researcher across different domains.

Dimension	Category	Count	Ratio (%)
Language	English	74	57.81
Language	Chinese	54	42.19
Source	PersonaHub	57	44.53
	FinePDFs	38	29.69
	arXiv	33	25.78
Aspect Ratio	16:9 Widescreen	42	32.81

Appendix

Table 7: Evaluation set statistics across language, source, aspect ratio, and slide-count constraints. “Free” indicates no constraint is specified.

Category	Action
Retrieve	search_web, search_images, search_papers, fetch_url, get_paper_authors, get_scholar_details , document_analyze, image_caption
File	convert_to_markdown, read_file, write_file, move_file, edit_file, download_file, execute_command, create_directory, list_directory
Reason	thinking, inspect_slide, inspect_manuscript
Control	todo_create, todo_update, todo_list, finalize
Create	image_generation

Appendix

Table 8: Action Categories

Agentic AI

Agentic artificial intelligence studies systems that plan, use tools, learn from feedback, and act toward goals with varying degrees of autonomy.

See all 211 papers

DeepSlide: From Artifacts to Presentation Delivery
Apr 1, 2026Ming Yang, Zhiwei Zhang, Jiahang Li +3Human-in-the-Loop AI
School of Data Science, Fudan University, Shanghai, China
PresentAgent-2: Towards Generalist Multimodal Presentation Agents
May 12, 2026Wei Wu, Ziyang Xu, Zeyu Zhang +2Video Generation Multimodal Agents
1Peking University · 2La Trobe University
Learning to Present: Inverse Specification Rewards for Agentic Slide Generation
Mar 17, 2026Karthik Ragunath Ananda Kumar, Subrahmanyam ArunachalamLLM Agent Training Tool-Augmented Language Model Agents
Tavus Inc. · University of Texas at Dallas · Texas A&M University

Apr 1, 2026J/K move · Enter open · S save

DeepSlide: From Artifacts to Presentation Delivery

Ming Yang, Zhiwei Zhang, Jiahang Li, Haoseng Liu+2

School of Data Science, Fudan University, Shanghai, China

Human-in-the-Loop AI

Open paper AskPDF arXiv

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. We further introduce a dual-scoreboard benchmark that cleanly separates static artifact quality from dynamic delivery excellence. Across 20 domains and diverse audience profiles, DeepSlide matches strong baselines on artifact quality while consistently achieving larger gains on delivery metrics, improving narrative flow, pacing precision, and slide--script synergy with clearer attention guidance.

Abstract

Agentic AI

Explore similar work

DeepSlide: From Artifacts to Presentation Delivery

PresentAgent-2: Towards Generalist Multimodal Presentation Agents

Learning to Present: Inverse Specification Rewards for Agentic Slide Generation