💻 computer science

X+Slides: Benchmarking Audience-Conditioned Slide Generation

The paper introduces X+Slides, a novel benchmark featuring a dynamic, audience-conditioned evaluation framework with four complementary metrics to assess how well large language models generate slide decks that balance source-grounded correctness with the specific informational needs of diverse target audiences.

Original authors: Haodong Chen, Xuanhe Zhou, Wei Zhou, Xinyue Shao, Yanbing Zhu, Bo Wang, Jiawei Hong, Anya Jia, Fan Wu

Published 2026-06-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haodong Chen, Xuanhe Zhou, Wei Zhou, Xinyue Shao, Yanbing Zhu, Bo Wang, Jiawei Hong, Anya Jia, Fan Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, dense encyclopedia about a specific topic. Now, imagine you need to turn that encyclopedia into a short, punchy slide presentation.

The problem is that who is watching the presentation changes everything.

If you are presenting to a scientist, they want the nitty-gritty details, the math proofs, and the tiny exceptions.
If you are presenting to a CEO, they only care about the bottom line, the risks, and the "what do we do next?"
If you are presenting to a student, they just need the big picture and a simple story.

For a long time, computers trying to make these slides have been like a bad waiter who serves the exact same meal to everyone: a giant, uncut steak. It might be "correct" (the steak is real), but it's useless to the vegetarian, the child, or the person who just wants a quick snack.

This paper introduces X+Slides, a new way to test if AI can be a better waiter.

The Core Idea: The "Universal Question Bank"

Instead of asking the AI to guess what the audience wants, the researchers built a massive, neutral "question bank" based on the source document. Think of this like a master list of every possible fact in the encyclopedia.

The Neutral List: First, they generate thousands of questions about the document (e.g., "What was the experiment method?" or "What is the main conclusion?"). These questions are neutral; they don't know who is watching yet.
The Audience Filter: Then, they apply a "filter" based on the audience.
- For the Scientist, the filter says: "The method question is worth 100 points. The 'big picture' question is worth 10 points."
- For the CEO, the filter flips: "The 'big picture' question is worth 100 points. The method question is worth 0 points."
The Score: The AI makes a slide deck. The researchers then check: "Did the AI include the high-value questions for this specific audience?"

The Four Scorecards

X+Slides doesn't just give one score. It uses four different ways to grade the presentation, like a teacher using a rubric:

Audience Coverage (Did they get the right stuff?): This measures how many of the "high-value" questions the audience cares about were actually answered in the slides. If the CEO wanted risks and the AI only gave them history, this score drops.
Domain Coverage (Did they pick the right types of stuff?): This breaks it down further. Did the AI focus too much on "methods" when the audience wanted "implications"? It's like checking if a chef served too much garnish and not enough main course.
Efficiency (Was it too long?): This asks, "How much useful info did I get per minute of my time?" If the slides are 50 pages long but only say what could be said in 5, the efficiency score is low.
Correctness (Did they lie?): This is the safety guardrail. It checks if every claim on the slide is actually supported by the original document. It prevents the AI from making up cool-sounding facts that aren't true.

What They Found (The Results)

The researchers tested three popular AI slide generators (DeepPresenter, SlideTailor, and NotebookLM) using this new system.

The Good News: The AI isn't terrible. It can grab a lot of the important information.
The Bad News: It's still missing the mark on who it's talking to.
- When asked to talk to a Scientist, the AI often missed the deep technical details (the "Level 3 and 4" facts).
- When asked to talk to a CEO, the AI sometimes got bogged down in too much technical jargon.
- NotebookLM (a Google tool) did surprisingly well at picking the right info for CEOs and students, but it struggled a bit with the deep technical details for scientists.

The Big Takeaway

The paper argues that we can no longer just say, "Look how pretty and long this slide deck is!" to judge an AI. A beautiful deck that gives a CEO a 50-page math lecture is a failure, even if the math is 100% correct.

X+Slides proves that to build a truly useful AI for presentations, we need to stop treating all facts as equal. We have to teach the AI that what is important to a scientist is useless to a CEO, and vice versa. The paper provides the ruler to measure if the AI is finally learning that lesson.

Technical Summary: X+Slides: Benchmarking Audience-Conditioned Slide Generation

1. Problem Formulation

The paper identifies a critical gap in existing automated slide generation benchmarks: they primarily assess source completeness, layout, and visual coherence, while implicitly assuming a "one-size-fits-all" paradigm where all source facts are equally valuable. In reality, slide generation is an audience-conditioned task. A single source document (e.g., a research paper) requires fundamentally different information selection depending on the target:

Specialists need rigorous proofs, assumptions, and ablations.
Learners need core ideas, motivations, and examples.
Decision-makers prioritize actionable conclusions, risks, and high-level implications.

Existing metrics fail to distinguish between a deck that faithfully summarizes a source and one that strategically selects information relevant to a specific audience. Furthermore, current evaluations often conflate source alignment with audience utility, making it difficult to isolate whether a system is failing due to poor generation or poor audience adaptation.

2. Methodology: The X+Slides Framework

X+Slides introduces a benchmark and evaluation pipeline designed to decouple source coverage from audience utility. The core innovation is a two-step evaluation process using a shared, audience-agnostic probe bank.

2.1 Benchmark Construction

Corpus: The dataset spans 113 diverse topics (50 academic papers, 63 non-academic reports in policy, business, health, etc.) across 7 presentation scenes.
Probe Generation: An LLM generates 8,133 deduplicated, source-grounded probes (question-answer pairs with evidence spans) from the source documents. These probes are audience-agnostic, meaning they are generated without knowledge of the target audience to ensure a consistent factual baseline.
Probe Attributes: Each probe is tagged with:
- Depth Level (1–4): Ranging from context/motivation (Level 1) to implementation details/hyperparameters (Level 4).
- Information Domain: Context, Method, Evidence, Limitations, Implementation, Implications.
- Evidence Modality: Text, table, chart, figure, equation, or multimodal mix.

2.2 Audience-Conditioned Scoring

Instead of generating different questions for different audiences, X+Slides assigns audience-specific utility weights ( $w_j$ ) to the same set of probes based on the audience profile (Specialist, Learner, Decision-maker) and presentation scene.

Utility Threshold ( $\tau_A$ ): Probes with weights above a threshold (e.g., $\tau_A = 0.7$ ) are considered "essential" for that specific audience.
Evaluation Logic: A generated deck is scored based on whether the essential probes for a specific audience can be answered using visible slide content that is supported by the source evidence.

2.3 Evaluation Metrics

The framework reports four complementary metrics:

Audience Coverage: The weighted fraction of audience-essential source information successfully conveyed in the deck.
Domain-wise Coverage: A granular breakdown of coverage across the six information domains (e.g., how well "Method" details are covered for a specialist vs. a learner).
Efficiency: The density of useful information delivered per unit of attention cost (measured by slide count and word count).
Correctness: A guardrail metric verifying that claims in the deck are factually supported by the source, preventing the reward of plausible but hallucinated content.
SafeEfficiency: The product of Efficiency and Correctness, representing useful utility that is also grounded.

3. Key Contributions

Formal Conceptualization: Defines slide generation as a conditional mapping $f: S \times A \times C \to T$ , explicitly framing audience adaptation as a measurable information-selection problem rather than a vague prompting goal.
Decoupled Probe Bank: Develops an evidence-backed, audience-agnostic probe bank that is subsequently weighted. This allows for fair comparison across audiences using the same factual baseline.
Comprehensive Metric Suite: Introduces a four-dimensional evaluation framework (Coverage, Domain-wise Coverage, Efficiency, Correctness) that distinguishes between faithful adaptation, generic summarization, and hallucination.
Reproducible Benchmark: Releases a dataset of 113 topics, 8,133 probes, three audience profiles, and a fully reproducible evaluation pipeline.
Empirical Evaluation: Provides an in-depth analysis of current state-of-the-art systems (DeepPresenter, SlideTailor) and a focused ablation of NotebookLM.

4. Experimental Results

The authors evaluated DeepPresenter, SlideTailor, and NotebookLM (as an ablation) under both audience-agnostic and audience-conditioned settings.

Partial Recovery of Essential Information: Current systems recover a substantial but incomplete portion of audience-essential information.
- At $\tau_A = 0.7$ , DeepPresenter achieved the best Audience Coverage of 0.714 (for Learners).
- SlideTailor reached 0.594 (for Learners).
- NotebookLM (agnostic setting) reached 0.853 for Decision-makers, though its image-based outputs precluded direct efficiency comparisons.
Audience Conditioning Effects:
- Conditioning prompts generally improved Audience Coverage (e.g., DeepPresenter's Specialist coverage rose from 0.413 to 0.496).
- Domain-wise Coverage revealed that conditioning successfully shifts focus (e.g., specialists receive more "Method" and "Evidence" coverage, while decision-makers receive more "Implications" and "Context").
Correctness vs. Efficiency:
- Correctness scores remained high (approx. 0.82–0.85) across systems, indicating most claims are source-supported.
- However, SafeEfficiency varied significantly. For instance, SlideTailor's conditioned learner deck had the highest SafeEfficiency among PPTX-generating systems, while DeepPresenter's conditioned learner deck had the highest Audience Coverage.
Trade-offs: The paper notes that optimizing for one metric (e.g., broad coverage) does not guarantee optimization for others (e.g., efficiency or correctness). Audience prompts successfully change information selection but do not automatically improve all dimensions of quality.

5. Significance and Claims

The paper claims that X+Slides represents a fundamental shift from document alignment to audience utility maximization.

Differentiation: It demonstrates that visual quality and broad topic coverage are insufficient proxies for slide quality; without source-grounded, audience-conditioned evaluation, systems may appear successful while missing critical audience needs or including unsupported claims.
Limitations: The authors modestly acknowledge current limitations, including potential gaps in LLM-generated probes for highly visual documents, the broadness of the three audience profiles, and the reliance on textual scoring which does not directly evaluate graphs or layout-based explanations.
Future Impact: The framework is positioned to support research communication, education, policy briefing, and business decision support, urging the community to treat audience adaptation as a rigorous information-selection challenge rather than a stylistic adjustment.