SlideLab: Audience-Centered Scientific Slide Generation and Evaluation

Vidushee Vats; Karun Sharma; Yuxia Wang

cs.CLSep 13, 2026

SlideLab: Audience-Centered Scientific Slide Generation and Evaluation

Authors: Vidushee Vats, Karun Sharma, Yuxia Wang

Organizations: INSAIT, Sofia University “St. Kliment Ohridski”

Abstract

Scientific presentations are more than summaries of research papers. They need to present the work in a coherent sequence, explain the main ideas clearly, and help the audience follow the presentation. We present SlideLab, a training-free multi-agent framework for generating scientific presentations from research papers. SlideLab first plans the presentation narrative, then builds and iteratively refines a shared slide deck using agents for content planning, visual generation, layout refinement, and grounding verification. In a blind human preference study, SlideLab was preferred over both open-source and commercial systems on 77% of papers while using roughly 4 times fewer inference tokens than the strongest open-source baseline. We also introduce ConfArena, an audience-oriented evaluation framework that simulates a conference room and assesses presentations slide by slide. ConfArena matches human system rankings and detects injected presentation problems, including falsified numbers, degraded figures, dropped slides, and shuffled slide order.

Figures & tables

Figure 1: SlideLab Overview: narrative planning, figure design, flow restructuring, layout debugging and grounding.

Figure 2: Slides shortcuts comparison between SlideLab (Ours) and DeepPresenter, Kimi Slides, Manus.

	Per slide				Whole talk
System	Grounding. errors $\downarrow$	Fig. errors $\downarrow$	Design $\uparrow$	Figure use $\uparrow$	Narrative. errors $\downarrow$	Coverage $\uparrow$	Env. score $\uparrow$
PPTAgent	0.98	0.64	3.21	3.18	5.40	0.61	0.42
DeepPresenter	2.08	0.53	3.49	3.27	5.19	0.94	0.32
Kimi Slides	1.48	0.17	3.67	3.08	4.56	0.91	0.58
Manus	1.81	–	3.59	–	4.14	0.89	0.37
SlideLab (ours)	0.71	0.26	4.21	3.80	3.29	0.99	0.73

Table 1: ConfArena results on 100 papers, with three retained decks per paper for SlideLab, DeepPresenter, and PPTAgent, and one for Kimi Slides and Manus. Best values are bold.

System	Papers picked best	% of 30
SlideLab (ours)	23	77%
Kimi Slides	7	23%
DeepPresenter	0	0%
Manus	0	0%

Table 2: Results of the blind human preference study on 30 research papers. Annotators selected the presentation they would use to deliver the paper at a conference after evaluating factual correctness, coverage, narrative flow, visual design, and figure usage.

Configuration	Env. score
SlideLab (full)	0.70
$-$ Planner	0.51
$-$ LayoutDebugger	0.58
$-$ Compositor	0.65
$-$ Custom visuals (paper figs only)	0.63

Table 3: Component ablations of SlideLab. The Planner and LayoutDebugger are the two stages whose removal changes the result significantly. Each configuration is run once per paper on the

100

-paper set, unlike Table 1 , which averages three independent runs per paper.

	PPTEval (1–5)			PresentBench	SlidesGen-Bench
System	Content $\uparrow$	Design $\uparrow$	Coherence $\uparrow$	pass % $\uparrow$	quiz acc. $\uparrow$
PPTAgent	3.5	3.8	3.6	63	0.74
DeepPresenter	3.7	3.7	3.4	64	0.76
Kimi Slides	3.9	4.2	4.0	73	0.79
Manus	3.6	3.8	3.5	61	0.72
SlideLab (ours)	4.1	4.4	4.2	71	0.84

Table 4: Evaluations on the

100

-paper set using other three methods. PPTEval: LLM-judge ratings of content, design, and coherence (

1

–

5

). PresentBench: fraction of decks passing the benchmark’s checks (%). SlidesGenBench: accuracy on quizzes the benchmark generates from the source paper.

\uparrow

higher is better. Best in bold.

System	Human	ConfArena	PPTEval	PresentBench	SlidesGenB.
SlideLab	1	1	1	2	1
Kimi Slides	2	2	2	1	2
Manus	3	3	3	4	4
DeepPresenter	3	4	4	3	3

Table 5: Comparison with human preferences. (a) Aggregate system rankings (

1

= best). Manus and DeepPresenter tie in the human study; PPTEval ranks use the mean of its three dimensions. (b) Frequency with which ConfArena selects the same preferred deck as the human majority for a paper or as an individual annotator.

Perturbation	PresentB.	PPTEval	SlidesG.	ConfArena
Falsified num.	$\times$	✓	$\times$	✓
Degraded fig.	✓	✓	✓	✓
Dropped slide	$\times$	$\circ$	✓	✓
Shuffled order	✓	✓	✓	✓
Caught /4	2	3	3	4

Table 6: Can evaluation frameworks detect planted errors in slides? ✓

=

the framework’s relevant metric moved in the expected direction;

\times

=

did not;

\circ

=

the framework has no metric targeting that failure.

	Mean change (positive $=$ worse)
Damage applied	Arc defects	Grounding errors	Figure errors
Shuffle slide order	+7.25	+0.11	-0.01
Inject one false number	+0.50	+1.51	+0.20
Shrink figures by 50%	+0.25	-0.17	+0.51
Drop Slides	+2.50	+0.01	-0.01

Table 7: We damage

15

SlideLab decks in targeted ways and re-run the ConfArena. Cells show the mean change per metric; bold marks the metric the damage targets. Each damage moves mainly its own metric.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Stage	Sub-stage	Model	Max rounds / calls
Planner	Candidate plan generation	GPT-5.5	60
	Candidate critic and selector	GLM-5.2	1
	Plan grounding verifier	GPT-5.5	1
Slide Generator	Slide generation	MiMo-v2.5-Pro	50
Visual Generator	Visual instruction / prompt generation	MiMo-v2.5-Pro	15
	Image generation	GPT-Image-2	1 per generated visual

Appendix

Table 8: Models and maximum round each SlideLab stage.

	ConfArena score $\uparrow$
Configuration	GPT-5.4	Gemini-3.8- Flash
SlideLab (full) (MiMo-V2.5-Pro)	0.74	0.78
$-$ Planner	0.53	0.62
$-$ LayoutDebugger	0.60	0.69
$-$ Compositor	0.70	0.74
$-$ Custom visuals (paper figures only)	0.69	0.70

Appendix

Table 9: Component ablations on 30 papers using MiMo-V2.5-Pro for all text agents. Columns report environment scores under two ConfArena runs with different model. Each minus denotes removal of one component from the full pipeline.

	Per slide				Whole talk
System	Grounding errors $\downarrow$	Figure errors $\downarrow$	Design $\uparrow$	Figure use $\uparrow$	Narrative errors $\downarrow$	Coverage $\uparrow$	Env. score $\uparrow$
PPTAgent (GPT-5.5)	1.462	0.443	3.441	3.090	5.13	0.64	0.38
DeepPresenter (GPT-5.5)	1.882	0.316	3.725	3.368	4.77	0.753	0.30
Kimi Slides	1.162	0.156	4.430	2.830	3.94	0.924	0.61
Manus	1.650	–	3.990	–	3.20	0.691	0.31
SlideLab (ours) (MiMo-V2.5-Pro, all text stages)	1.244	0.140	4.673	4.525	2.400	0.927	0.784

Appendix

Table 10: System comparison on 30 human-annotated papers using Gemini-3.8-Flash for all ConfArena evaluation roles. SlideLab uses MiMo-V2.5-Pro for all text agents. Metric definitions follow Table 1 . Bold marks the best value in each column; dashes indicate unavailable figure metrics.

Figure 3: ConfArena evaluation architecture: sequential slide review by the examiner and three independent attendees, audience Q&A and question diagnosis, and aggregation of per-slide and whole-talk evidence into the environment score.

Attendee	Focus and Question Trigger
Expert reviewer	Focuses on scientific rigor and correctness. Asks a question when a claim, result, or numerical value appears unsupported, inconsistent, or overstated.
Learner	Focuses on whether the core ideas, methodology, and motivation are easy to follow. Asks a question when an important concept or step is insufficiently explained.
Cross-field attendee	Focuses on accessibility without deep subfield knowledge. Asks a question when jargon or field-specific assumptions are introduced without adequate explanation.

Appendix

Table 11: Attendee personas used in ConfArena.

Metric	Pearson correlation ( $r$ )
Grounding errors	0.939
Figure errors	0.952
Design	0.981
Figure use	0.917
Coverage	0.637
Narrative errors	0.552

Appendix

Table 12: Correlation between ConfArena scores and scores from an examiner that evaluates slides separately. Both use Gemini-3.8-Flash.

	Narrative errors $\downarrow$
Evaluator	SlideLab	Kimi Slides
Static examiner	0.90	0.80
ConfArena	2.40	3.94

Appendix

Table 13: Mean narrative errors per deck under the static examiner and ConfArena.

Evaluator condition		Narrative errors per deck
	Full ConfArena	3.88
	Neutral attendee roles	3.40
	Without Q&A	2.30
	Examiner only	1.75

Appendix

Table 14: Mean narrative errors per deck under different ConfArena configurations.

Stage	Time (s)	Cost ($)	Tokens (M)
Planner	96	0.10	0.09
Slide Generator	264	0.03	0.23
ImageVisualGenerator + Compositor	261	0.42	0.32
LayoutDebugger	122	0.03	0.05
SlideLab total	743	0.58	0.69
DeepPresenter	1620	2.1	2.9

Appendix

Table 15: Average wall-clock time (end to end), API cost, and token usage per deck.

Figure 4: Annotation interface for the human preference study. Annotators see the four anonymized decks side by side, pick the one they would present, then rate it on six dimensions.

Dimension	What annotators are asked
Content Grounding	Do the slides match what the paper says?
Content Coverage	Does the deck cover the paper’s key contributions?
Narrative Structure	Do the slides tell a coherent story?
Visual Design	Does it look like a professional conference talk?
Information Density	Are slides concise, not walls of text?
Figure Usage	Are figures well-chosen and sized correctly?

Appendix

Table 16: The six dimensions annotators rate after picking their preferred deck. Each uses a five-point scale with the anchors shown in the interface.

Perturbation	Benchmark	Metric (direction)	Baseline	Perturbed	$\Delta$	Detected?
Falsified number	PPTEval	content (1–5, $\uparrow$ )	4.30	4.27	$-0.03$	✓
	SlidesGen-Bench	quiz accuracy (0–1, $\uparrow$ )	0.98	0.98	$+0.00$	$\times$
	PresentBench	material-dep. pass % ( $\uparrow$ )	27.5	45.0	$+17.5$	$\times$
	ConfArena	faithfulness (0–1, $\uparrow$ )	0.58	0.57	$-0.01$	✓
Degraded figure	PPTEval	design (1–5, $\uparrow$ )	4.05	3.85	$-0.20$	✓
	SlidesGen-Bench	visual weighted total (0–10, $\uparrow$ )	8.45	8.22	$-0.23$	✓

Appendix

Table 17: Mean baseline and perturbed scores per benchmark per perturbation, across

15

papers.

\Delta

= perturbed minus baseline; for metrics where higher is better, a negative

\Delta

means the benchmark moved in the expected direction. ✓

=

detected,

\times

=

not detected,

\circ

=

no metric exists for that failure mode. Baselines are the mean over the

15

unperturbed decks.

Panel of Figure 5 — Figure 5: Representative qualitative comparisons supporting the quantitative results reported in the main paper.

Explore similar work

X+Slides: Benchmarking Audience-Conditioned Slide Generation
Jun 17, 2026Haodong Chen, Xuanhe Zhou, Wei Zhou +6Evidence-Grounded Generation Conditional Generation
Shanghai Jiao Tong University · Harbin Institute of Technology · SenseTime
SeaSlides: Semantic Abstraction Layer for Agentic Slide Generation
Aug 4, 2026Shengjun Fang, Chenyang Wu, Zongzhang ZhangStructured Output Generation Agentic AI
National Key Laboratory for Novel Software Technology, Nanjing University, China · School of Artificial Intelligence, Nanjing University, 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

Jun 17, 2026, cs.AIJ/K move · Enter open · S save

X+Slides: Benchmarking Audience-Conditioned Slide Generation

Haodong Chen, Xuanhe Zhou, Wei Zhou, Xinyue Shao+5

Shanghai Jiao Tong University · Harbin Institute of Technology · SenseTime

Evidence-Grounded Generation Conditional Generation Generative AI Evaluation

Open paper AskPDF arXiv

Automatically generating slide decks from source documents is an important application of large language models (LLMs). Existing benchmarks primarily assess slide completeness and technical depth, while overlooking the target audience as a critical real-world factor. For instance, specialists demand rigorous proofs, whereas decision-makers prioritize actionable conclusions. To bridge this gap, we introduce X+Slides, a benchmark specifically designed for audience-conditioned slide generation. Built on a diverse corpus spanning 113 topics and seven presentation scenes, X+Slides employs a dynamic evaluation framework constructed from 8,133 deduplicated, source-grounded probes. By assigning audience-specific utility weights to the same source-grounded probes, X+Slides reports four complementary metrics: Audience Coverage measures how much audience-essential information is conveyed, Domain-wise Coverage shows which information types are covered, Efficiency measures delivered utility per unit of attention cost, and Correctness verifies whether slide claims are supported by the source. Experiments on DeepPresenter, SlideTailor, and NotebookLM show that current systems can recover a substantial but still incomplete part of audience-essential information: at

τ_A=0.7

, DeepPresenter reaches a best Audience Coverage of 0.714, SlideTailor reaches 0.594, and the NotebookLM ablation reaches 0.853 while showing clear grounding differences. These results indicate that visual quality and broad topic coverage should not be treated as evidence support without source-grounded evaluation.

Abstract

Explore similar work

X+Slides: Benchmarking Audience-Conditioned Slide Generation

SeaSlides: Semantic Abstraction Layer for Agentic Slide Generation

PresentAgent-2: Towards Generalist Multimodal Presentation Agents