Paper deep dive
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
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/20/2026, 9:47:40 AM
Summary
The paper introduces DeepPresenter, an agentic framework for automated presentation generation that utilizes a dual-agent architecture (Researcher and Presenter) collaborating via a shared environment. It employs environment-grounded reflection, where agents inspect rendered artifacts to identify and correct visual defects, rather than relying on internal self-reflection. The framework achieves state-of-the-art performance on diverse presentation tasks and includes a fine-tuned 9B model (DeepPresenter-9B) that offers competitive results at lower cost.
Entities (8)
Relation Signals (6)
DeepPresenter → uses → Researcher Agent
confidence 95% · DEEPPRESENTER coordinates two specialized agents... The Researcher autonomously explores and compiles a structured manuscript
DeepPresenter → uses → Presenter Agent
confidence 95% · The Presenter converts it into visually coherent slides via content-driven design
DeepPresenter → employs → Environment-Grounded Reflection
confidence 92% · rather than relying on self-reflection over internal signals... our environment-grounded reflection conditions the generation process on perceptual artifact states
DeepPresenter-9B → isfinetunedversionof → DeepPresenter
confidence 90% · We therefore develop DeepPresenter-9B via supervised fine-tuning on curated trajectories
DeepPresenter → uses → inspect tool
confidence 88% · we introduce the inspect tool as an explicit observation interface... Agents then use think to reflect on observed defects
DeepPresenter → outperforms → Gamma
confidence 85% · DEEPPRESENTER achieves an average score of 4.44, surpassing... the commercial system Gamma (4.36)
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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: this https URL
Tags
Links
- Source: https://arxiv.org/abs/2602.22839v3
- Canonical: https://arxiv.org/abs/2602.22839v3
Trouble viewing inline? Open PDF directly →
Full Text
53,451 characters extracted from source content.
Expand or collapse full text
DEEPPRESENTER: Environment-Grounded Reflection for Agentic Presentation Generation Hao Zheng 1,2,∗ , Guozhao Mo 1,2,∗ , Xinru Yan 2 , Qianhao Yuan 1,2 , Wenkai Zhang 1 , Xuanang Chen 1 , Yaojie Lu 1,† , Hongyu Lin 1 , Xianpei Han 1 , Le Sun 1 1 Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences, Beijing, China 2 University of Chinese Academy of Sciences, Beijing, China zhenghao2022,moguozhao2024,chenxuanang,luyaojie,hongyu@iscas.ac.cn Abstract Presentation generation requires deep content research, coherent visual design, and itera- tive refinement based on observation. How- ever, existing presentation agents often rely on predefined workflows and fixed templates. To address this, we present DEEPPRESEN- TER, an agentic framework that adapts to di- verse user intents, enables effective feedback- driven refinement, and generalizes beyond a scripted pipeline. Specifically, DEEPPRESEN- TER autonomously plans, renders, and revises intermediate slide artifacts to support long- horizon refinement with environmental obser- vations. Furthermore, rather than relying on self-reflection over internal signals (e.g., rea- soning traces), our environment-grounded re- flection 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 per- formance, and the fine-tuned DeepPresenter- 9B remains highly competitive at substantially lower cost. Our project is available at:https: //github.com/icip-cas/PPTAgent 1 Introduction Presentations are a primary medium for informa- tion delivery across education, business, and re- search. A high-quality presentation combines well- researched content with coherent visual design, en- abling audiences to grasp complex ideas efficiently. However, creating such presentations remains time- consuming and skill-demanding, motivating recent work that leverages Multimodal Large Language Models (MLLMs) to automate this task (Liang et al., 2025; Yang et al., 2025b; Zheng et al., 2025). * These authors contributed equally to this work. † Corresponding authors. User Instruction: I want to create a presentation about "Attention Is All You Need" for beginners. Manuscript Writing: OK, I'l read the paper and create assets... Visual Design: I'l read the manuscript and craft slides... Researcher Presenter Shared Agent Environment Attention is All You Need search paper.pdf, https://""arxiv.org/pdf/1706.03762 download paper.pdf, pdf read I should create a graph for beginner..."" think python –c "import matplotlib ..." bash manuscript.md, Attention is All Your Need : Ashish Vaswani,... write manuscript.md, manuscript inspect manuscript.md, "fig1.jpg", "images/fig1.jpg" edit manuscript.md, markdown read designplan.md, Slide Design Plan ## 1. Colo r Scheme ... write slide_01.html, <!DOCTYPE html> <html ... write slide_01.html, slide inspect The contrast of the main title seems weak..."" think slide_01.html, color:black, color: #3A86F edit 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. However, existing presentation agents (Sefid et al., 2021; Xu et al., 2025; Yang et al., 2025b) fall short of meeting these demands. First, they rely on predefined workflows (Zheng et al., 2025) and content-agnostic templates (Cachola et al., 2024), limiting adaptability to varying user intents. This yields text-heavy slides with insufficient research depth and visual designs that fail to resonate with the narrative. Second, introspective reflection over internal signals (e.g., code or reasoning traces) can- not detect post-render defects (Kim et al., 2025; Tang et al., 2025), resulting in overlapping ele- ments, truncated text, and broken layouts. To address these limitations, we propose DEEP- PRESENTER, an agentic framework for presenta- tion generation (Figure 1). Unlike prior methods that decouple content and design via rigid tem- arXiv:2602.22839v3 [cs.AI] 20 Apr 2026 plates, DEEPPRESENTER coordinates two special- ized agents through a shared observation space. The Researcher autonomously explores and com- piles a structured manuscript aligned with the user intent, while the Presenter converts it into visually coherent slides via content-driven design rather than template filling. Crucially, instead of intro- spective self-reflection over internal signals, DEEP- PRESENTER grounds reflection in perceptual arti- fact states obtained from environmental observa- tion (Figure 2): agents useinspectto view ren- dered manuscripts and slides, andthinkto plan targeted revisions to correct post-render defects. While our framework achieves strong perfor- mance with proprietary models, their high cost motivates a more efficient alternative. We there- fore develop DeepPresenter-9B via supervised fine- tuning on curated trajectories (Figure 3). We first construct diverse presentation tasks from Per- sonaHub (Ge et al., 2024), arXiv, and FinePDFs (Kydl ́ ı ˇ cek et al., 2025), augmented with verifiable constraints. During trajectory synthesis, we miti- gate self-verification bias (Stechly et al., 2024) with extrinsic verification: an independent critic evalu- ates artifacts in isolation and provides reasoning traces that steer targeted refinements, improving the quality of synthesized trajectories. We evaluate our method on a held-out set of 128 diverse presentation tasks across three dimen- sions: constraint satisfaction, content quality, and visual style. With proprietary backbones, DEEP- PRESENTER achieves an average score of 4.44, surpassing open-source baselines and the commer- cial system Gamma (4.36). Our specialized agen- tic design yields richer content and coherent de- sign, while environment-grounded reflection re- duces post-render defects by revising against ob- served perceptual artifact states. DeepPresenter-9B scores 4.19, outperforming all open-source base- lines and approaching GPT-5 (4.22) at lower cost. In summary, our contributions are threefold: •We propose DEEPPRESENTER, an agentic presentation framework that coordinates Re- searcher and Presenter agents via a shared observation space, enabling autonomous in- formation research and topic-aware design. •We introduce environment-grounded reflec- tion that grounds self-correction in perceptual artifact states obtained from post-render ob- servations, reducing defects that are not de- tectable from internal signals alone. •Results on the evaluation set covering diverse presentation-generation scenarios show that DEEPPRESENTER achieves state-of-the-art performance, and the distilled DeepPresenter- 9B remains highly competitive at substantially lower cost. 2DEEPPRESENTER In this section, we present DEEPPRESENTER, a dual-agent framework for presentation generation. We first formulate the task as an interactive agen- tic process, then describe the Researcher-Presenter collaboration and the environment-grounded reflec- tion mechanism, as illustrated in Figure 2. 2.1 Task Formulation We formulate presentation generation as an inter- active agentic task. Given an instructionIand an agent environmentEequipped with a tool li- braryTand a file systemF, the system aims to generate a high-quality presentationP. The gener- ation process can be modeled as a multi-step trajec- toryτ = (r 1 ,a 1 ,o 1 ),..., (r T ,a T ,o T ), where at each stept, the agent generates a reasoning trace r t , selects an actiona t ∈T, and receives observa- tiono t fromE. We decompose the trajectory into two sequential phases:τ = τ R ◦ τ P , whereτ R andτ P denote the Researcher and Presenter trajec- tories, respectively. The two agents communicate throughF, where the Researcher persists a struc- tured manuscriptMand associated assets for the Presenter to consume. Appendix C lists the tools. 2.2 Dual-Agent Collaboration Presentation generation requires both information research and visual design, which demand differ- ent planning and tool use. We split these roles between two specialized agents while sharing the same backbone model. Researcher Agent GivenI, the Researcher au- tonomously plans its exploration instead of follow- ing a predefined workflow. It executes multiple steps duringτ R , invoking tools fromTto retrieve and synthesize supporting materials and to create auxiliary assets as needed. The exploration depth and strategy adapt to user intent: a technical presen- tation may require surveying related work, while a general-audience talk may prioritize accessible examples and vivid illustrations. Finally, the Re- searcher compiles slide text and associated assets Rendered SlideImage Manuscript Diagnostics LLM Agent Self-Reflection Shared Environment Trigger: Uncertain ❓ Input: Uncertain ❓ Reflection: I am doing great so far ... Shared Environment Inspection Trigger: Inspect Tool Action Action Input: Observation Grounded-Reflection Researcher Presenter Self-Reflection Env-Grounded Reflection AI Era Think Inoticedthattheimage "fig1.jpg"ismissingthe "images/"pathprefix. Num Slide: 14 Language: EN Missing: fig1.jpg 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. into a structured markdown manuscriptMorga- nized by narrative flow, and persists it toF . Presenter Agent Rather than populating prede- fined templates, the Presenter generates slides from scratch duringτ P . GivenMfromF, the agent first develops a global design plan, establishing color themes and typography that resonate with the topic. It then generates each slide as a stan- dalone HTML file, translating manuscript content into visual elements following the design plan. This content-driven approach enables stylistic choices aligned with the presentation topic, such as earthy palettes for sustainability or minimalist layouts for academic tutorials. 2.3 Environment-Grounded Reflection We ground agent reflection in environmental ob- servations rather than introspective reasoning over internal signals (He et al., 2025). The key issue with self-reflection is state mismatch: agents op- erate on intermediate representations (e.g., HTML or markdown), while users perceive only rendered artifacts. As a result, many defects manifest only in perceptual states (e.g., broken images, overflow, or low contrast), leaving introspective reflection operating in a mismatched observation space. To make perceptual artifact states observable to the agent, we introduce theinspecttool as an explicit observation interface. For the Presenter, inspectrenders an HTML file into image pixels, exposing post-render defects such as overflow, over- DimensionCategoryCountRatio (%) Language English60352.34 Chinese54947.66 Source PersonaHub58650.87 FinePDFs36231.42 arXiv20417.71 Aspect Ratio 16:9 Widescreen32728.39 4:3 Standard30426.39 A1 Poster302.60 Free49142.62 Slide Count 11-2024921.61 1-1032027.78 Free58350.61 Total1,152100.00 Table 1: Statistics of the constructed presentation tasks by language, source, aspect ratio, and slide count. “Free” indicates no constraint is specified. lap, and low contrast; for the Researcher,inspect returns structured diagnostics of the manuscript and file state, including slide count, asset availability, and detected language. Agents then usethinkto reflect on observed defects and plan targeted edits. This forms an observe–reflect–revise loop where agent observations align with user perception. 3 Frontier Presentation Agent Model This section presents our training pipeline as shown in Figure 3: task dataset construction, trajectory synthesis with extrinsic verification to elicit high- quality reflective behaviors, and multi-stage filter- ing for quality. 3.1 Query Construction We construct a task collection for training our com- pact model and evaluating our framework. To cover diverse presentation scenarios in both intent-driven and document-conditioned settings, we draw task seeds from PersonaHub (Ge et al., 2024), arXiv, and FinePDFs-Edu (Kydl ́ ı ˇ cek et al., 2025). Each task is augmented with verifiable constraints (e.g., slide count, language, aspect ratio) to capture fine- grained user-specified requirements. For Person- aHub, we prompt GLM-4.6 to synthesize presen- tation tasks conditioned on persona descriptions; for arXiv and FinePDFs-Edu, we construct tasks that require generating presentations based on pro- vided documents. Each task is further augmented with verifiable constraints, including slide count, language, and aspect ratio. In total, this task collec- tion contains 1,152 tasks, with 1,024 for trajectory Trajectory SynthesisTask Construction Researcher Presenter Shared Environment Reflection Critic Trajectory Filter Trajectories High-Quality Trajectories Reasoning Trace I noticed that “AI Era” has low contrast with the background. I should adjust the text color to improve contrast. A Bystander Sees More Than a Player <Reasoning Trace> I'l change the text color to black ... Self Verification Bias: Over Confidence This slide looks perfect already Reflect with Extrinsic Verification Inspection Rule Verification Num Slides: 14 Aspect Ratio: 4:3 Language: Chinese LLM Verification Consistency: Yes Style Score: 4.2 Content Score: 4.7 Inspect tool Observation Aspect RatioNum SlidesLanguage Constraint Type I'm an English teacher and I need to create a presentation about American War of Independence. Please create a 14-slidepresentation in 4:3 aspect ratio, with all content in Chinese. PersonaHubArxivFinePDF Persona: An American English teacher working at a middle school in China. 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. sampling and 128 held out for evaluation. Detailed statistics are shown in Table 1. 3.2 Verification-Guided Trajectory Synthesis When sampling agentic trajectories, self-reflection is susceptible to self-verification bias (Jiang et al., 2025): the agent judges its own intermediate out- puts from within the same trajectory state that pro- duced them. This coupling entangles verification with self-justification, resulting in flawed outputs being accepted. To break this coupling, we in- troduce extrinsic verification, where verification signals are produced in an isolated context. As illustrated in Figure 3, after the agent invokes inspectand obtains an observationo t , an indepen- dent critic performs verification conditioned ono t and the corresponding intermediate artifacts. The critic outputs a reasoning trace that identifies de- fects (e.g., low contrast) and specifies actionable adjustments (e.g., adjust text color). We append this trace to the agent context as athinkcall, guid- ing targeted revisions before continuing the rollout. 3.3 Trajectory Filtering We adopt a three-stage filtering pipeline to ensure trajectory quality. First, we verify constraint com- pliance through a rule-based system. Second, we evaluate consistency using GLM-4.6, removing tra- jectories that fail to follow the extrinsic-verification trace with aligned revisions (i.e., reflection–action inconsistency). Third, we assess output quality using GLM-4.6V, filtering out trajectories with crit- ical defects such as element overlap or broken im- ages. 4 Experiment In this section, we evaluate our method on presen- tation generation and analyze our key components. 4.1 Setup Implementation DetailsWe sample trajectories by running DEEPPRESENTER with Gemini-3-Pro as the backbone and critic model on 1,024 training tasks, with a maximum context window of 50K tokens. 802 trajectories pass our filtering pipeline and are used for supervised fine-tuning. We fine- tune GLM-4.6V-Flash on these trajectories using MS-SWIFT (Zhao et al., 2024), with a batch size of 32 and learning rate of 1e-5 for 5 epochs. Train- ing takes approximately 80 GPU hours on 8 A800 GPUs. Models and Baselines We compare against one commercial system, Gamma 1 , and two academic frameworks: PPTAgent (Zheng et al., 2025) and KCTV (Cachola et al., 2024). For backbone mod- els, we evaluate with proprietary GPT-5 (OpenAI, 2025), Gemini-3-Pro (Comanici et al., 2025), and Claude-Sonnet-4.5 (Anthropic, 2025), as well as open-source GLM-4.6 (Zeng et al., 2025a). For DEEPPRESENTER, we additionally evaluate with GLM-4.6V and GLM-4.6V-Flash (Team et al., 2025), as our framework leverages visual feedback through the inspect tool. Evaluation ProtocolWe hold out 128 tasks from the constructed task collection and evaluate gener- ated presentations using the following metrics: 1 https://gamma.app/ FrameworkModelConstraintContentStyleAvg.Diversity Close-sourced Baseline Gamma–4.934.084.084.360.52 Open-sourced Baseline PPTAgent GPT-53.963.004.073.680.35 Gemini-3-Pro4.223.094.30 3.870.19 Claude-Sonnet-4.53.722.934.153.600.17 GLM-4.64.023.174.243.810.30 KCTV GPT-54.952.843.633.810.21 Gemini-3-Pro4.583.013.903.830.27 Claude-Sonnet-4.54.882.903.993.920.20 GLM-4.64.662.833.943.810.25 Ours DEEPPRESENTER GPT-54.803.794.074.220.56 Gemini-3-Pro4.704.254.374.440.79 Claude-Sonnet-4.54.904.054.274.41 0.49 GLM-4.6V4.693.253.753.900.58 GLM-4.6V-Flash4.673.113.693.820.47 DeepPresenter-9B4.773.524.294.190.53 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. •Constraint scores each presentation by the fraction of user-specified constraints it satisfies, covering slide count, language, and aspect ratio, verified through rule-based checking. •Content & Style evaluate the quality of slide content and visual design. We adopt the MLLM- based evaluation framework from Zheng et al. (2025) with GPT-5 as the judge, which has been validated to correlate well with human judgments. •Diversity quantifies visual style variance across generated presentations using the Vendi Score (Friedman and Dieng, 2022), which com- putes diversity based on the eigenvalue entropy of feature similarity matrices extracted by DINOv2 (Oquab et al., 2023). We report Avg. as the mean of Constraint, Con- tent, and Style (scaled 0–5), while Diversity (scaled 0–1) measures cross-presentation variation. 4.2 Main Results Table 2 presents the main experimental results. DEEPPRESENTER achieves state-of-the-art per- formance Across all backbone models, DEEPP- RESENTER consistently outperforms open-source baselines. With Gemini-3-Pro as the backbone, DEEPPRESENTER attains an average score of 4.44, surpassing the best open-source baseline (KCTV + Claude-Sonnet-4.5, 3.92) by 13.3% and the com- mercial product Gamma (4.36). The improve- ments stem from two aspects: (1) Content qual- ity improves most because Researcher performs intent-adaptive information seeking and synthesis, rather than relying on fixed workflows or user- provided inputs. Baseline frameworks depend on user-provided materials and lack deep retrieval ca- pability, while our agent searches, retrieves, and synthesizes information from diverse sources. (2) Style scores improve through content-aware design and environment-grounded reflection. Our frame- work enables Presenter to align design decisions with the narrative, while environment-grounded re- flection mitigates free-form generation failures by revising against post-render defects. Free-form generation enables greater visual di- versity, with DEEPPRESENTER achieving a di- versity score of 0.79. Under our diversity met- ric, DEEPPRESENTER more than doubles template- based baselines by generating slides in a free-form manner. Baseline frameworks achieve diversity scores of only 0.17 to 0.35, as fixed templates con- ConfigurationCons.ContentStyleAvg. Gemini-3-Pro DEEPPRESENTER4.704.254.374.44 w/o Grounded Reflection4.524.154.314.32 w/o Dual-Agent3.943.964.224.04 DeepPresenter-9B DeepPresenter-9B4.773.524.294.19 w/o Grounded Reflection4.213.234.013.82 w/o Dual-Agent3.652.933.113.23 w/o Trajectory Filtering4.673.304.124.03 Table 3: Ablation study on framework components and training strategy. Cons. denotes constraint satisfaction. ConfigurationCons.ContentStyleAvg.∆ GLM-4.6V-Flash4.673.113.693.82– + Fine-tuning4.713.193.923.94+0.12 + Extrinsic Verification4.743.284.034.02+0.20 Table 4: Effect of extrinsic verification on model perfor- mance. Both fine-tuned variants use 300 trajectories.∆ denotes improvement over the base model. strain visual variation. PPTAgent, in particular, shows lower constraint scores because its style de- cisions are predetermined by the workflow, limiting task-specific adaptation. Even Gamma, despite its commercial polish, achieves only 0.52. In contrast, our framework maintains high constraint compli- ance while enabling greater visual diversity (0.79). DeepPresenter-9B surpasses all open-source baselines with high efficiency. With only 802 trajectories, our compact model achieves an aver- age score of 4.19, outperforming open-source base- lines and matching GPT-5 (4.22) at substantially lower cost. These results support the effectiveness of our verification-guided trajectory synthesis and suggest that compact models can acquire agentic behaviors from limited but high-quality samples. 4.3 Ablation Study We ablate key components of DEEPPRESENTER on Gemini-3-Pro and DeepPresenter-9B, as shown in Table 3. (1) Environment-grounded reflection is critical because it extends observation space to post-render perceptual artifact states.Dis- ablinginspectconfines reflection to pre-render artifacts and degrades performance from 4.44 to 4.32 on Gemini-3-Pro and from 4.19 to 3.82 on DeepPresenter-9B. (2) Dual-agent collaboration contributes significantly by decomposing long- horizon execution into specialized sub-tasks. With- out it, performance drops substantially on both ConstraintIntegrityFormat (a) Manuscript 0 20 40 60 80 56 76 27 35 26 31 LayoutRenderStyle (b) Slide 0 50 100 150 200 250 300 212 308 43 101 53 69 Self-VerificationExtrinsic-Verification Figure 4: Distribution of defects identified by self- verification and extrinsic verification for manuscripts (left) and slides (right), respectively. backbones. (3) Trajectory filtering effectively pre- vents biased and low-quality patterns from being distilled during fine-tuning. Removing it drops DeepPresenter-9B from 4.19 to 4.03. 5 Analysis We analyze the effectiveness of the extrinsic evalua- tion, examine failure modes in trajectory synthesis, and present efficiency comparisons alongside qual- itative case studies. 5.1 Effect of Extrinsic Verification Extrinsic verification improves trajectory syn- thesis by mitigating self-verification bias. To quantify its impact, we train two variants on 300 tra- jectories sampled from the same set of tasks, with and without extrinsic verification during trajectory synthesis. As shown in Table 4, adding extrinsic verification yields a 67% larger gain in Avg. (0.20 vs. 0.12) than fine-tuning alone. This indicates that, even with environment-grounded observations, re- vision signals produced solely within the agent’s own trajectory state can be biased, leading to sub- optimal refinements being distilled during learning. Extrinsic verification mitigates self-verification bias by strengthening defect-triggered revision signals. We categorize reflection-triggered de- fects into three manuscript types: integrity (e.g., missing asset references), constraint (e.g., mis- matched slide count), and format (e.g., invalid markup); and three slide types: layout (e.g., over- lap), render (e.g., blank slides), and style (e.g., low contrast). Figure 4 compares defects identified on the same 300 trajectories under self-verification versus extrinsic verification. Extrinsic verification consistently yields more defect detections across categories, with the largest gaps on slides (e.g., 308 vs. 212 for layout and 101 vs. 43 for render). 010203040 Percentage of Total Errors (%) Consistency Constraint Environment Quality 11.2% 13.5% 32.3% 43.0% Figure 5: Failure distribution in synthesized trajectories before filtering $0.01$0.10$1.0 Cost per Task ($) 3.6 3.8 4.0 4.2 4.4 4.6 Average Score Gamma GPT-5 Gemini-3-Pro Claude-Sonnet-4.5 GLM-4.6 GPT-5 Gemini-3-Pro Claude-Sonnet-4.5 GLM-4.6 GPT-5 Gemini-3-Pro Claude-Sonnet-4.5 GLM-4.6V GLM-4.6V-Flash Ours Gamma PPTAgent KCTV DeepPresenter Pareto Frontier Significant Performance Gain Figure 6: Performance vs. Price scatter plot with Pareto frontier representation. Different colors represent differ- ent frameworks This pattern indicates a systematic failure in self- verification: when verification is conducted within the generating trajectory state, the agent tends to ra- tionalize defects, producing biased judgment (Jiang et al., 2025; Stechly et al., 2024). By decoupling verification from the agent’s own trajectory state, extrinsic verification mitigates this bias and pro- vides stronger signals to trigger corrective revisions during synthesis. 5.2 Trajectory Failure Analysis Following the categories in Section 3.3, we ana- lyze failures in synthesized trajectories before fil- tering (Figure 5). Quality errors are most prevalent (43.0%), underscoring the difficulty of sustaining high standards under free-form generation. Envi- ronment failures are also common (32.3%), reflect- ing long-horizon fragility from context overflow and infrastructure disruptions. The remaining cases include Constraint violations (13.5%) and Consis- tency errors (11.2%), which are less frequent but still non-negligible. 5.3 Efficiency Analysis Figure 6 presents the cost-performance 2 trade-off across frameworks and models. (1) DeepPresenter- 9B advances the Pareto frontier, significantly out- performing the prior frontier point at comparable cost. Compared to KCTV + Gemini-3-Pro (3.83), DeepPresenter-9B achieves 4.19 at a similar price, a significant improvement in cost-efficiency. (2) DEEPPRESENTER establishes a new upper bound for presentation generation, surpassing the previ- ous best system Gamma. With an average score of 4.44 versus Gamma’s 4.36, DEEPPRESENTER delivers the strongest result in our evaluation. Notably, baseline frameworks exhibit flat perfor- mance across backbone models, whereas DEEPP- RESENTER demonstrates substantial variation (3.82 to 4.44). This pattern is consistent with baselines being limited by their fixed pipelines, while DEEP- PRESENTER can better leverage stronger model capacity. 5.4 Case Study We present qualitative examples in Figure 7. (1) DEEPPRESENTER produces visually rich slides through diverse asset sources, while baselines tend to yield text-heavy outputs. Gamma includes more imagery than academic baselines. However, it re- lies heavily on AI-generated images and often mis- handles figures embedded in source documents (e.g., inappropriate scaling of architectural dia- grams). Open-source baselines rarely retrieve or create supporting visuals, resulting in predomi- nantly textual content. (2) DEEPPRESENTER gen- erates visual themes that resonate with content, whereas baselines rely on fixed templates. For ex- ample, DEEPPRESENTER employs green tones for environmental topics and minimalist layouts for academic presentations, while baseline methods exhibit limited topical alignment due to template- driven generation. 6 Related Work Presentation generation has attracted increasing at- tention due to its practical value for information delivery. Before the emergence of large language models, presentation generation was primarily for- mulated as a document summarization task. These approaches employed extractive summarization to select salient sentences using neural networks (Fu 2 The cost of fine-tuned models is estimated using the per- token pricing of their base models on OpenRouter. KCTVPPTAgentGammaDeepPresenter-9BDeepPresenter (Gemini) KCTVPPTAgentGammaDeepPresenter-9BDeepPresenter (Gemini) 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. et al., 2022; Hu and Wan, 2014; Sun et al., 2021) or phrase-based methods (Wang et al., 2017). How- ever, the limited reasoning capabilities of pre-LLM models constrained their ability to handle diverse user intents and produce visually engaging outputs. The emergence of LLMs has shifted the paradigm toward agent-based approaches that lever- age stronger reasoning and generalization capabili- ties. Recent work explores multi-agent collabora- tion for content extraction and layout planning (Ca- chola et al., 2024; Ge et al., 2025; Liang et al., 2025; Xu et al., 2025; Yang et al., 2025b), aesthetic-aware generation (Liu et al., 2025), as well as slide under- standing and editing (Huang et al., 2025; Jung et al., 2025; Zeng et al., 2025b; Zheng et al., 2025). How- ever, these approaches often focus on predefined workflows and fixed templates, limiting adaptation to user intent and iterative refinement with environ- mental feedback. Compared with previous methods, DEEPPRE- SENTER formulates presentation generation as an autonomous exploration and collaboration process between two specialized agents. The Researcher- Presenter decomposition enables adaptive plan- ning based on task complexity, while environment- grounded reflection allows agents to verify and refine artifacts through rendered slides and file sys- tem states (Jiang et al., 2025; Tang et al., 2025). 7 Conclusion In this work, we propose DEEPPRESENTER, an agentic framework for presentation generation in which agents plan autonomously and adapt to di- verse user intents. Our framework grounds self- reflection in perceptual artifact states from environ- mental observations, enabling agents to iteratively identify and fix post-render defects. We further train DeepPresenter-9B on trajectories synthesized with extrinsic verification, which mitigates self- verification bias and strengthens reflective behav- iors. Results show that DEEPPRESENTER achieves state-of-the-art performance, while DeepPresenter- 9B remains competitive at substantially lower cost. Acknowledgments We sincerely thank the reviewers for their insight- ful comments and valuable suggestions. This work was supported by Beijing Natural Science Founda- tion (L243006), the Natural Science Foundation of China (No. 62476265, 62306303, 62506354). Limitations While DEEPPRESENTER demonstrates strong per- formance, several limitations remain. First, DEEP- PRESENTER relies on multi-step, tool-using roll- outs, which increase inference cost and are sensi- tive to environment instability (e.g., context over- flow and infrastructure failures) observed in our trajectory analysis. Second, extrinsic verification is only used during trajectory synthesis. We do not employ an external critic at inference time, as critic- provided reflection signals can introduce reflection– action inconsistency and additional overhead. Fu- ture work can explore mitigating self-verification bias at inference time. References Anthropic. 2025.Introducing claude sonnet 4.5.https://w.anthropic.com/news/ claude-sonnet-4-5. [Accessed 18-11-2025]. Isabel Alyssa Cachola, Silviu Cucerzan, Allen Herring, Vuksan Mijovic, Erik Oveson, and Sujay Kumar Jauhar. 2024. Knowledge-centric templatic views of documents. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 15460–15476, Miami, Florida, USA. Association for Computational Linguistics. Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Mar- cel Blistein, Ori Ram, Dan Zhang, Evan Rosen, Luke Marris, Sam Petulla, Colin Gaffney, Asaf Aha- roni, Nathan Lintz, Tiago Cardal Pais, Henrik Ja- cobsson, Idan Szpektor, Nan-Jiang Jiang, and oth- ers. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. Preprint, arXiv:2507.06261. Dan Friedman and Adji Bousso Dieng. 2022. The vendi score: A diversity evaluation metric for machine learning. arXiv preprint arXiv:2210.02410. Tsu-Jui Fu, William Yang Wang, Daniel McDuff, and Yale Song. 2022. Doc2ppt: Automatic presentation slides generation from scientific documents. Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, 36(1):634–642. Jiaxin Ge, Zora Zhiruo Wang, Xuhui Zhou, Yi-Hao Peng, Sanjay Subramanian, Qinyue Tan, Maarten Sap, Alane Suhr, Daniel Fried, Graham Neubig, and others. 2025. Autopresent: Designing structured visuals from scratch. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2902–2911. Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu. 2024. Scaling synthetic data cre- ation with 1,000,000,000 personas. arXiv preprint arXiv:2406.20094. Yancheng He, Shilong Li, Jiaheng Liu, Weixun Wang, Xingyuan Bu, Ge Zhang, Z.y. Peng, Zhaoxiang Zhang, Zhicheng Zheng, Wenbo Su, and Bo Zheng. 2025. Can large language models detect errors in long chain-of-thought reasoning? In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18468–18489, Vienna, Austria. Association for Computational Linguistics. Yue Hu and Xiaojun Wan. 2014. Ppsgen: Learning- based presentation slides generation for academic papers. IEEE transactions on knowledge and data engineering, 27(4):1085–1097. Zheng Huang, Xukai Liu, Tianyu Hu, Kai Zhang, and Ye Liu. 2025. Pptbench: Towards holistic eval- uation of large language models for powerpoint layout and design understanding. arXiv preprint arXiv:2512.02624. Dongwei Jiang, Jingyu Zhang, Orion Weller, Nathaniel Weir, Benjamin Van Durme, and Daniel Khashabi. 2025. Self-[in] correct: Llms struggle with discrim- inating self-generated responses. In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 39, pages 24266–24275. Kyudan Jung, Hojun Cho, Jooyeol Yun, Soyoung Yang, Jaehyeok Jang, and Jaegul Choo. 2025. Talk to your slides: Language-driven agents for efficient slide edit- ing. arXiv preprint arXiv:2505.11604. Jeonghye Kim, Sojeong Rhee, Minbeom Kim, Dohyung Kim, Sangmook Lee, Youngchul Sung, and Kyomin Jung. 2025. Reflact: World-grounded decision mak- ing in llm agents via goal-state reflection. arXiv preprint arXiv:2505.15182. Hynek Kydl ́ ı ˇ cek, Guilherme Penedo, and Leandro von Werra. 2025. Finepdfs.https://huggingface.co/ datasets/HuggingFaceFW/finepdfsedu. Xin Liang, Xiang Zhang, Yiwei Xu, Siqi Sun, and Chenyu You. 2025. Slidegen: Collaborative mul- timodal agents for scientific slide generation. arXiv preprint arXiv:2512.04529. Chengzhi Liu, Yuzhe Yang, Kaiwen Zhou, Zhen Zhang, Yue Fan, Yanan Xie, Peng Qi, and Xin Eric Wang. 2025. Presenting a paper is an art: Self-improvement aesthetic agents for academic presentations. arXiv preprint arXiv:2510.05571. OpenAI. 2025. Introducing gpt-5.https://openai. com/index/introducing-gpt-5/ . [Accessed 18- 11-2025]. Maxime Oquab, Timoth ́ e Darcet, Th ́ eo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fer- nandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, and others. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193. Athar Sefid, Prasenjit Mitra, and Lee Giles. 2021. Slide- gen: an abstractive section-based slide generator for scholarly documents. In Proceedings of the 21st ACM Symposium on Document Engineering, pages 1–4. Kaya Stechly, Karthik Valmeekam, and Subbarao Kamb- hampati. 2024. On the self-verification limitations of large language models on reasoning and planning tasks. arXiv preprint arXiv:2402.08115. Edward Sun, Yufang Hou, Dakuo Wang, Yunfeng Zhang, and Nancy XR Wang. 2021. D2s: Document- to-slide generation via query-based text summariza- tion. arXiv preprint arXiv:2105.03664. Zhengyang Tang, Ziniu Li, Zhenyang Xiao, Tian Ding, Ruoyu Sun, Benyou Wang, Dayiheng Liu, Fei Huang, Tianyu Liu, Bowen Yu, and others. 2025. Realcritic: Towards effectiveness-driven evaluation of language model critiques. arXiv preprint arXiv:2501.14492. V Team, Wenyi Hong, Wenmeng Yu, Xiaotao Gu, Guo Wang, Guobing Gan, Haomiao Tang, Jiale Cheng, Ji Qi, Junhui Ji, Lihang Pan, Shuaiqi Duan, Weihan Wang, Yan Wang, Yean Cheng, Zehai He, Zhe Su, Zhen Yang, Ziyang Pan, and others. 2025. Glm- 4.5v and glm-4.1v-thinking: Towards versatile multi- modal reasoning with scalable reinforcement learn- ing. Preprint, arXiv:2507.01006. Sida Wang, Xiaojun Wan, and Shikang Du. 2017. Phrase-based presentation slides generation for aca- demic papers. In Proceedings of the AAAI Confer- ence on Artificial Intelligence, volume 31. Xiaojie Xu, Xinli Xu, Sirui Chen, Haoyu Chen, Fan Zhang, and Ying-Cong Chen. 2025. Pregenie: An agentic framework for high-quality visual presenta- tion generation. arXiv preprint arXiv:2505.21660. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and others. 2025a. Qwen3 technical report. arXiv preprint arXiv:2505.09388. Yuheng Yang, Wenjia Jiang, Yang Wang, Yiwei Wang, and Chi Zhang. 2025b.Auto-slides: An inter- active multi-agent system for creating and cus- tomizing research presentations.arXiv preprint arXiv:2509.11062. Qianhao Yuan, Jie Lou, Zichao Li, Jiawei Chen, Yaojie Lu, Hongyu Lin, Le Sun, Debing Zhang, and Xianpei Han. 2025. Memsearcher: Training llms to reason, search and manage memory via end-to-end reinforce- ment learning. arXiv preprint arXiv:2511.02805. Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, and others. 2025a. Glm-4.5: Agentic, reasoning, and coding (arc) foundation mod- els. arXiv preprint arXiv:2508.06471. Wenzheng Zeng, Mingyu Ouyang, Langyuan Cui, and Hwee Tou Ng. 2025b. Slidetailor: Personalized pre- sentation slide generation for scientific papers. arXiv preprint arXiv:2512.20292. Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, Wenmeng Zhou, and Yingda Chen. 2024. Swift:a scalable lightweight infrastruc- ture for fine-tuning. Preprint, arXiv:2408.05517. Hao Zheng, Xinyan Guan, Hao Kong, Wenkai Zhang, Jia Zheng, Weixiang Zhou, Hongyu Lin, Yaojie Lu, Xianpei Han, and Le Sun. 2025. Pptagent: Gen- erating and evaluating presentations beyond text-to- slides. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 14413–14429. A Detailed Analysis A.1 Human Evaluation To address concerns about potential circularity in- troduced by LLM-as-judge evaluation, we conduct a small-scale human study to corroborate the au- tomatic assessments. We recruited two graduate students majoring in computer science to evaluate 32 randomly sampled presentations from the test set. Following the evaluation dimensions in Sec- tion 4, annotators rate Content and Style on a 1–5 Likert scale using the scoring criteria of Zheng et al. (2025), while Constraint satisfaction is verified via rule-based checks consistent with our evaluation protocol. Evaluators were provided with rendered slide images and scored them independently. Ta- ble 5 reports the resulting ratings. Importantly, the relative ranking and overall trends under human judgment align with our automatic evaluation, sug- gesting that the observed improvements are not an artifact of relying solely on GPT-5 as the judge. A.2 Performance by Domain We analyze DEEPPRESENTER with Gemini-3-Pro across domains. PersonaHub shows the strongest content (4.49) and style (4.49) scores, but relatively lower constraint satisfaction (4.38). This is likely because PersonaHub queries are synthesized by an LLM based on persona descriptions, resulting in more diverse and complex constraint specifi- cations that are harder to follow. arXiv achieves near-perfect constraint satisfaction (4.91) but the lowest content (3.84) and style (4.13) scores. The formal nature of academic presentations restricts visual diversity, and accurately conveying technical content requires deeper domain understanding. A.3 Tool Usage Analysis We analyze tool invocation patterns across agents and domains, as shown in Figure 8. For agent roles (Figure 8a), Researcher and Presenter exhibit dis- tinct tool preferences aligned with their responsibil- ities. Researcher relies heavily on Retrieve tools for information gathering, while Presenter focuses on File operations and Reason tools for iterative slide editing and reflection. This specialization validates our dual-agent design, where each agent develops tool usage patterns tailored to its role. Across domains (Figure 8b), Researcher shows adaptable usage patterns reflecting task characteris- tics. PersonaHub tasks exhibit significantly higher Retrieve usage, as persona-driven queries do not MethodCons.ContentStyleAvg. Gamma4.843.523.904.09 PPTAgent3.723.073.603.46 KCTV4.412.843.193.48 DeepPresenter4.563.864.254.22 Table 5: Human evaluation results on 32 randomly sam- pled presentations. DomainCons.ContentStyleAvg. PersonaHub4.384.494.494.45 arXiv4.913.844.134.29 FinePDF4.944.214.384.51 Table 6: Domain performance breakdown. Cons. de- notes constraint satisfaction. provide reference documents, requiring agents to search for relevant materials (Yuan et al., 2025). In contrast, arXiv and FinePDF tasks involve pro- vided source documents, leading to higher File us- age for document processing and lower reliance on retrieval. Tool categories are detailed in Table 8. B Dataset B.1 Data Sources We collect presentation tasks from three sources to ensure diverse scenario coverage. For academic presentations, we pair arXiv papers with requests that specify target audiences (beginners, intermedi- ate learners, domain experts, or peer researchers) and corresponding scenarios (lectures, seminars, defenses, or conference talks). For general educa- tional topics, we sample English and Chinese PDF documents from FinePDFs-Edu (Kydl ́ ı ˇ cek et al., 2025), each accompanied by instructions to create a presentation based on the attachment. For personalized scenarios, we leverage Person- aHub (Ge et al., 2024) and prompt Qwen3-235B- A22B (Yang et al., 2025a) to generate realistic presentation requests grounded in user personas. We adopt two generation strategies: knowledge- grounded generation, which incorporates both per- sona descriptions and synthesized domain knowl- edge, and open-ended generation, which relies solely on persona characteristics. The model is instructed to adopt the persona’s perspective and select the appropriate language based on cultural background. Generated queries undergo language filtering, semantic deduplication, and LLM-based File 37.8% Retrieve 26.8% Reason 18.7% Create 9.7% Control 7.0% Control 3.5% Reason 44.5% File 52.0% read_file download_file execute_command write_file edit_file create_directory list_directory convert_to_markdown move_file search_web search_images image_caption fetch_url search_papers document_analyze thinking inspect_manuscript image_generation finalize todo_create todo_update todo_list finalize thinking inspect_slide list_directory create_directory edit_file read_file write_file (a) Tool Usage by Agent arXiv FinePDF Persona Hub File 50.5% Reason 19.8% Control 11.8% Retrieve 9.6% Create 8.2% File 40.7% Reason 24.0% Retrieve 17.0% Control 9.3% Create 9.0% Retrieve 42.9% File 28.3% Reason 15.1% Create 11.0% Control 2.7% read_file execute_command write_file edit_file download_file convert_to_markdown create_directory list_directory thinking inspect_manuscript todo_create todo_update finalize todo_list search_images image_caption search_web search_papers document_analyze image_generation read_file execute_command edit_file download_file create_directory write_file convert_to_markdown list_directory thinking inspect_manuscript image_caption search_images search_web fetch_url todo_create todo_update finalize todo_list image_generation search_web search_images fetch_url image_caption search_papers download_file write_file create_directory read_file list_directory move_file edit_file execute_command convert_to_markdown thinking inspect_manuscript image_generation finalize (b) Tool Usage by Domain (Researcher) Figure 8: Tool usage analysis. (a) Distribution of tool invocations by agent role. (b) Tool usage patterns of Researcher across different domains. quality control to remove low-quality or inappro- priate samples. B.2 Constraint Augmentation To assess instruction-following capabilities, each task is augmented with verifiable constraints, in- cluding slide count, aspect ratio (widescreen 16:9, standard 4:3, or poster), and language. These con- straints are randomly assigned per task. For auto- mated verification, we parse generated PDFs and validate them against specified constraints using a rule-based system. The constraint satisfaction score is computed as the proportion of constraints successfully met. B.3 Evaluation Set To facilitate replication, we disclose the composi- tion of our 128-task evaluation split and statistics in Table 7. C Agent Framework Presentation creation requires interacting with het- erogeneous resources beyond static web text, in- cluding search results, images, papers, and local files, as well as inspecting intermediate artifacts such as manuscripts and rendered slides. To sup- port this, we organize our toolset into five cate- gories (Table 8): Retrieve for information gather- ing, File for document manipulation, Reason for inspection and reflection, Control for task manage- DimensionCategoryCountRatio (%) Language English7457.81 Chinese5442.19 Source PersonaHub5744.53 FinePDFs3829.69 arXiv3325.78 Aspect Ratio 16:9 Widescreen4232.81 4:3 Standard3426.56 A1 Poster43.12 Free4837.50 Slide Count 11-202620.31 1-103628.12 Free6651.56 Total128100.00 Table 7: Evaluation set statistics across language, source, aspect ratio, and slide-count constraints. “Free” indicates no constraint is specified. ment, and Synthesis for code execution and asset generation. Inspection Tools.The Reason category includes two inspection tools that enable environment- grounded reflection: • inspect manuscript: Parses the markdown manuscript and returns structured diagnostics, including the total slide count, detected con- tent language, and validation results for refer- enced image assets. The tool checks whether CategoryAction Retrieve searchweb, searchimages, searchpapers, fetchurl, getpaperauthors, get scholardetails , documentanalyze, imagecaption File convert tomarkdown, readfile, writefile, movefile, editfile, downloadfile, executecommand, create directory, listdirectory Reason thinking, inspectslide, inspectmanuscript Control todocreate, todoupdate, todolist, finalize Createimagegeneration Table 8: Action Categories each image path exists, flags external URLs that should be downloaded locally, identifies missing alt text, and warns about duplicate image usage. • inspectslide: Renders an HTML slide into a pixel image using a headless browser and returns the image to the agent’s visual context. The tool supports multiple aspect ratios (16:9 widescreen, 4:3 standard, A1 poster) and en- ables agents to perceive visual defects such as contrast issues and element overflow that are invisible at the code level. Each task is executed as a sequence of reasoning- action-observation steps within a maximum context window of 50K tokens. To prevent context over- flow, our system sends warning messages when the accumulated window length reaches 50% and 80% of the maximum capacity, allowing the agent to adjust its strategy accordingly. D Prompts D.1 Data Synthesis Prompts QUERY SYNTHESIS (PERSONAHUB) Generate a slide creation request based on the following information: hint User persona: persona QUERYSYNTHESIS(PERSONAHUB- DETAIL) Assume you are a user with the following characteristics: <persona> persona </persona> You want to create a slide presentation based on the following topic: <presentation topic> synthesizedtext </presentationtopic> hint Please generate a slide creation request based on the above persona and topic. D.2 Extrinsic Verification Prompts EXTRINSIC VERIFICATIONFOR RE- SEARCHER AGENT You are a professional slide content reviewer responsible for checking slide content for issues based on specified dimensions. Review Dimensions Your review authority is strictly limited to the following dimensions: • Image path compliance: Check whether local paths are used and whether captions are included • Language selection compliance: Check whether the document is written in the correct language • Language consistency: Check for unnecessary mixing of Chinese and English, or inconsistent style • Language correctness: Check for grammatical errors, spelling mistakes/typos • Tool-returned warnings: Check warning messages and evaluate their impact on user understanding Problem Description Standards When finding issues, use first person: • Problem Location: ‘I noticed on this page...’ / ‘The tool detected...’ • Improvement Plan: ‘I will...’ Return Format (strict JSON) ‘severity’: <0-3 integer>, ‘thought’: ‘<analysis, less than 30 words>’ EXTRINSIC VERIFICATION FOR PRESEN- TER AGENT You are a professional slide design reviewer, responsible for analyzing the visual design and readability of HTML slides generated by another Design Agent and providing specific improvement guidelines. Review Dimensions 1. Readability • Whether the contrast between text and background is too low, causing reading difficulties • Whether fonts and images render properly • Whether text elements are obscured or overflow 2. Aesthetics • Whether similar elements maintain consistent alignment • Whether color schemes, visual hierarchy, and layout cause visual confusion • You should only check if images display correctly, not their aesthetics or watermarks Problem Description Standards When finding issues, use first person: ‘I noticed on this slide...’ → ‘This will cause...’ → ‘I will...’ Return Format (strict JSON) ‘severity’: <0-3 integer>, ‘thought’: ‘<analysis and improvement actions>’ D.3 Agent System Prompts RESEARCHER AGENT SYSTEM PROMPT You are a professional presentation content expert capable of leveraging various tools for deep and comprehensive information retrieval and collection based on user requirements, then analyzing and highly distilling the information to create high-quality slide content that embodies ‘Information Aesthetics’ Task Instructions • Based on user requirements and their underlying logic, conduct systematic and comprehensive information research, and construct a slide framework with strong narrative tension • After fully completing information collection and organization, organize visual materials guided by information value and content logic • Write the manuscript in Markdown format: Use --- for page separation; images must be downloaded locally and referenced via relative paths • Upon completion, call finalize with the manuscript path as the parameter Important Notes • Use the same language as the user’s instructions for manuscript generation • Your task is limited to manuscript writing and material collection/creation; do not involve slide layout and design work • Leverage ‘thinking‘ to reflect on the current state and next steps, and execute strictly • You are not allowed to interact with the user; all information must be obtained through retrieval and tools PRESENTER AGENT SYSTEM PROMPT You are a professional slide visual design expert, skilled in creating fixed-layout slide designs using HTML/CSS. Your core competency is faithfully transforming manuscripts into visually balanced, overlap-free, high-quality slides suitable for projection display, making full use of all available materials. Task Description • Deeply analyze the manuscript and develop a ‘slide master’ style design plan (including color scheme, fonts, grid system, font size specifications) • Based on the design plan and manuscript content, generate high-quality HTML files page by page • After generating all slides, call the finalize tool to return the slides folder and end the workflow Important Notes • Mandatory Fixed Dimensions: Strictly lock body/html to specified dimensions (e.g., 16:9, 1280px x 720px), and set overflow: hidden • Use the same language as the user’s instructions for thinking and working • Thinking: You can use ‘thinking‘ to reflect on the current state and plan next steps, then execute strictly