DeepPresenter: Adaptive Slide Generation

• DeepPresenter is an adaptive presentation generation system defined by its dual-agent (Researcher and Presenter) architecture that separates content research from slide design.
• It employs environment-grounded reflection to inspect and revise rendered slides and manuscripts, ensuring both visual coherence and content accuracy.
• Empirical evaluations show competitive performance with Gemini-3-Pro scoring 4.44 and DeepPresenter-9B achieving 4.19, underscoring its effective design and audience conditioning.

DeepPresenter is an agentic framework for presentation generation that separates long-horizon content research from slide design and revision through a dual-agent architecture composed of a Researcher and a Presenter. Its defining mechanism is environment-grounded reflection: rather than reflecting only on internal reasoning traces, code, or intermediate text, it conditions revision on perceptual artifact states such as rendered slides and manuscript diagnostics. This positions DeepPresenter as a system for adaptive presentation construction under user constraints, not merely document summarization. In evaluation, the Gemini-3-Pro-based framework reaches a 4.44 average score and a 0.79 diversity score, while the compact DeepPresenter-9B reaches 4.19; in the audience-conditioned X+Slides benchmark, DeepPresenter is the strongest PPTX-generating baseline by Audience Coverage but still recovers only a substantial, incomplete portion of audience-essential information, with a best score of 0.714 at $\tau_A=0.7$ (Zheng et al., 26 Feb 2026, Chen et al., 17 Jun 2026).

1. Problem formulation and design rationale

DeepPresenter is motivated by a tension specific to automated presentation generation: a useful deck must do more than summarize source material. It must perform deep content research, satisfy user constraints such as language, slide count, and aspect ratio, and remain visually coherent after rendering. The framework is explicitly presented as an alternative to systems that rely on predefined workflows and fixed templates, which are described as brittle under diverse user intents and prone to text-heavy slides with weak topical alignment.

The same general problem is formalized in X+Slides as audience-conditioned slide generation, where a slide deck is treated as a selective, audience-tailored information artifact rather than a compressed version of its source. Specialists demand technical depth, assumptions, proofs, and limitations; learners prioritize motivation, intuition, and core ideas; decision makers prioritize implications, risks, and actionable conclusions. This framing is important for DeepPresenter because it shifts evaluation away from generic completeness or visual polish toward whether the deck selected the right source-grounded information for the intended audience (Zheng et al., 26 Feb 2026, Chen et al., 17 Jun 2026).

2. Dual-agent architecture and artifact production

DeepPresenter uses a dual-agent, environment-aware architecture in which the Researcher and Presenter share a common tool environment and file system. The Researcher is responsible for autonomous information gathering and manuscript construction. Given a user instruction, it plans its own exploration, retrieves web pages, papers, images, and other resources as needed, and synthesizes them into a structured Markdown manuscript. That manuscript includes slide text and associated assets organized by narrative flow and is written to disk for downstream consumption.

The Presenter converts the manuscript into the visual deck. Rather than populating a fixed template, it creates slides from scratch as HTML files. The process begins with a global slide master style plan specifying color palette, typography, layout grid, and overall visual language, after which the slides are rendered page by page in a design intended to remain consistent with the topic. This yields a long-horizon iterative workflow in which intermediate artifacts can be inspected, revised, and regenerated until the presentation is finalized.

The Researcher–Presenter split is not treated as a convenience layer but as a core quality mechanism. In ablation, removing the dual-agent structure drops the Gemini-3-Pro configuration from 4.44 to 4.04 and the 9B model from 4.19 to 3.23, indicating that separating long-horizon research from visual deck construction materially improves performance (Zheng et al., 26 Feb 2026).

3. Environment-grounded reflection and execution model

The technical core of DeepPresenter is environment-grounded reflection, which differs from self-reflection over internal reasoning traces. The premise is that the true presentation artifact is the rendered slide, not the HTML or Markdown source. Many defects are post-render phenomena: overlapping elements, broken images, text truncation, overflow, low contrast, misalignment, and font or image rendering issues. DeepPresenter therefore adds explicit inspection interfaces that expose artifact state in the environment.

For the Presenter, inspect_slide renders an HTML slide into pixels using a headless browser, allowing the agent to observe real rendering defects. For the Researcher, inspect_manuscript parses the Markdown manuscript and returns structured diagnostics such as slide count, detected language, asset validity, missing local image paths, missing alt text, and duplicate image usage. The reflective loop is then implemented as observe–think–revise: the agent inspects the actual artifact state, uses think to reason about that observed state, and applies targeted edits.

The execution process is written as an interactive agentic trajectory

$$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$

where each step consists of a reasoning trace $r_t$, an action $a_t$ from a tool library $\mathcal{T}$, and an observation $o_t$ from the environment $\mathcal{E}$. The overall trajectory is decomposed as

$\tau = \tau^R \circ \tau^P,$

with $\tau^R$ the Researcher trajectory and $\tau^P$ the Presenter trajectory; communication occurs through the file system $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$0, which stores the manuscript $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$1 and associated assets.

The tool ecosystem is organized into five categories: Retrieve, File, Reason, Control, and Create. Retrieve includes web and document search, URL fetching, paper metadata, and image captions. File includes reading, writing, moving, editing, downloading, executing commands, and directory management. Reason contains inspect_manuscript, inspect_slide, and generic thinking. Control includes todo management and finalize. Create includes image generation. During execution, the system also emits context-length warnings at 50% and 80% of the 50K-token maximum so that the agent can adapt strategy and avoid overflow (Zheng et al., 26 Feb 2026).

4. Training pipeline, extrinsic verification, and DeepPresenter-9B

The compact model, DeepPresenter-9B, is trained from synthesized trajectories rather than from a manually authored fixed workflow. Query construction draws from PersonaHub, arXiv, and FinePDFs-Edu, and these prompts are augmented with verifiable constraints such as slide count, language, and aspect ratio. The task pool contains 1,152 tasks, with 1,024 used for trajectory sampling and 128 held out for evaluation.

A central training claim is that self-verification creates bias: if the agent evaluates its own output in the same trajectory state that produced it, it may rationalize defects instead of correcting them. DeepPresenter addresses this with extrinsic verification. After the agent calls inspect, an independent critic evaluates the observed artifact in isolation and outputs a short reasoning trace with a severity score and actionable advice. That trace is then inserted back into the agent context as a think signal. The goal is to produce targeted revisions guided by artifact-level critique rather than self-justifying introspection.

Before supervised fine-tuning, trajectories pass through a three-stage filtering pipeline. First, a rule-based system checks whether specified constraints are satisfied. Second, GLM-4.6 evaluates whether the trajectory actually follows the extrinsic-verification advice; trajectories with reflection-action inconsistency are removed. Third, GLM-4.6V evaluates output quality and removes trajectories with critical defects such as overlap or broken images. Of the 1,024 sampled trajectories, 802 survive filtering and are used for fine-tuning.

DeepPresenter-9B is trained with supervised fine-tuning using MS-SWIFT, batch size 32, learning rate $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$2, and 5 epochs, requiring about 80 GPU hours on 8 A800 GPUs. In ablation, removing trajectory filtering reduces the 9B model from 4.19 to 4.03, while training with extrinsic verification yields a larger gain than fine-tuning alone on a 300-trajectory comparison. The analysis also reports that extrinsic verification detects more issues than self-verification across manuscript and slide categories, especially layout and render defects (Zheng et al., 26 Feb 2026).

5. Empirical performance and audience-conditioned evaluation

DeepPresenter is evaluated on 128 held-out tasks spanning English and Chinese, persona-driven prompts, arXiv papers, and educational PDFs, with varied slide-count and aspect-ratio constraints. The primary evaluation dimensions are Constraint, Content and Style, and Diversity. Constraint is the fraction of user-specified constraints satisfied, verified by rules that parse generated PDFs. Content and Style are assessed by an MLLM-based framework from prior work, using GPT-5 as judge. Diversity is measured by the Vendi Score, computed from eigenvalue entropy over DINOv2 feature similarity matrices. The reported average score is the mean of Constraint, Content, and Style, each scaled to 0–5, while Diversity is separately scaled to 0–1.

Under this protocol, the full DeepPresenter framework with Gemini-3-Pro as backbone scores 4.44 average, surpassing the best open-source baseline and also beating Gamma, which scores 4.36. It also achieves a 0.79 diversity score. The compact DeepPresenter-9B scores 4.19 average, outperforming all open-source baselines and approaching GPT-5’s 4.22. Removing environment-grounded reflection lowers performance from 4.44 to 4.32 on Gemini-3-Pro and from 4.19 to 3.82 on the 9B model, which supports the claim that post-render, artifact-grounded revision is a substantive contributor to quality (Zheng et al., 26 Feb 2026).

X+Slides evaluates DeepPresenter from a different angle: source-grounded, audience-conditioned information selection. The benchmark is constructed from 113 topics, 50 academic papers, 63 non-academic documents, seven presentation scenes, three audience profiles, and 8,133 deduplicated probes. Each probe is an evidence-backed information slice

$$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$3

where $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$4 is the question, $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$5 the expected answer, $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$6 the exact source evidence span, $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$7 the depth level, $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$8 the evidence modality, and $$\tau = \{(r_1, a_1, o_1), \dots, (r_T, a_T, o_T)\},$$9 the information domain. Audience-specific utility is assigned on a four-point rubric, and the main threshold is $r_t$0, retaining only probes judged essential. At that threshold, the benchmark retains on average 42 essential probes for specialists, 31 for learners, and 24 for decision makers per topic. Audience Coverage is defined as

$r_t$1

with $r_t$2 and $r_t$3, so the metric measures how much audience-essential information is conveyed and grounded (Chen et al., 17 Jun 2026).

For DeepPresenter, the audience-conditioned gains are clear but incomplete:

Audience	Agnostic AudCov	Conditioned AudCov
Specialist	0.413	0.496
Learner	0.658	0.714
Decision maker	0.622	0.654

Its best reported Audience Coverage is 0.714 in the conditioned learner setting. The corresponding claim-level Correctness scores remain fairly stable: 0.842 in the agnostic setting, and 0.835 for conditioned specialist, 0.846 for conditioned learner, and 0.819 for conditioned decision maker. The time-based Efficiency values lie roughly between 1.663 and 2.603, with the best conditioned specialist SafeEfficiency at 2.603 and conditioned learner SafeEfficiency at 2.448. The domain-wise analysis reports that conditioning improves DeepPresenter across all six domains, with the conditioned learner row reaching 0.84 context, 0.77 method, 0.70 evidence, 0.73 limitations, 0.75 implementation, and 0.72 implications. These results support a specific interpretation: audience conditioning mainly improves information selection rather than correctness, and specialist-level depth remains a notable gap (Chen et al., 17 Jun 2026).

6. Position in the literature, misconceptions, limitations, and directions

DeepPresenter belongs to the broader family of agentic presentation-generation systems that rely on intermediate representations, rendering, and iterative repair rather than direct slide-image synthesis. A nearby system, PreGenie, uses Slidev and a five-agent decomposition consisting of Text Summarizer, Image Captioner, Code Generator, Code Reviewer, and Page Reviewer, which places it in the same general render-check-repair lineage while emphasizing multimodal document understanding and editable Markdown code (Xu et al., 27 May 2025).

Two recurrent misconceptions are directly challenged by work surrounding DeepPresenter. The first is that visual quality implies evidential quality. X+Slides explicitly argues that visual quality and broad topic coverage should not be treated as evidence support without source-grounded evaluation, and its NotebookLM contrast shows that strong coverage can be separated from claim-level correctness. The second is that reflection alone is sufficient for layout quality. In the controlled prefix-conditioned next-slide generation setting of AeSlides, reflection-heavy methods are presented as expensive and weak on subtle aesthetics; the reported DeepPresenter baseline receives a human score of 3.107, while AeSlides + verifiable rewards reaches 3.435 and the GDPO variant reaches 3.561. AeSlides therefore argues for explicit layout supervision through verifiable rewards rather than relying exclusively on VLM-based reflection for aspect ratio, whitespace, collision, and imbalance (Pan et al., 21 Apr 2026).

The limitations stated for DeepPresenter are operational rather than conceptual. Multi-step tool-using rollouts increase inference cost and create sensitivity to environment instability, including context overflow and infrastructure failures. Extrinsic verification is used during trajectory synthesis but not at inference time, because an external critic online can create reflection-action inconsistency and additional overhead. The analysis of synthesized trajectories also identifies quality errors as the most common problem before filtering, followed by environment failures.

The broader research direction suggested by DeepPresenter’s own results and by X+Slides is toward systems that combine agentic planning, explicit artifact inspection, and stricter source grounding with richer audience models. The benchmark literature specifically calls for finer-grained audience profiles, stronger visual grounding, better treatment of graphs, diagrams, and layout-dependent explanations, and multilingual or cross-cultural presentation settings. Taken together, these findings place DeepPresenter at an important transition point: from workflow- and template-based slide generation toward systems that jointly optimize content research, audience-utility-driven selection, rendered-form revision, and grounded presentation correctness (Chen et al., 17 Jun 2026).