SlideTailor: Personalized Presentation Slide Generation for Scientific Papers¶
🕒 Published (v1): 2025-12-23 12:01 UTC · Source: Arxiv · Venue: AAAI 2026 · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
💬 Ask ChatGPT✦ Ask Claude
TL;DR¶
SlideTailor is a training-free agentic framework for generating personalized presentation slides from scientific papers, conditioned on implicit user preferences encoded in a reference paper-slides pair and a .pptx template. It introduces a dual-branch preference distillation module and a chain-of-speech mechanism that jointly shapes content structure and oral narration. The system achieves 75.8% overall score on the new PSP benchmark, outperforming all baselines, with an 81.63% human-judged win rate against the strongest prior method.
Problem¶
Prior paper-to-slides systems treat slide generation as a one-size-fits-all document conversion task, ignoring that users differ in narrative structure, section emphasis, conciseness, and aesthetic choices. Asking users to specify preferences via detailed text is burdensome; existing preference-aware work (e.g., Persona-Aware-D2S) is restricted to four fixed categories and cannot capture fine-grained, open-ended variation.
Method¶
SlideTailor operates in three stages:
-
Implicit Preference Distillation: A dual-branch module extracts (a) content preferences—narrative flow, section-level emphasis/omission, formatting style—from a reference paper-slides pair via an LLM that infers the latent mapping
f_content: D_ref → S_ref; and (b) aesthetic preferences—layout schemas with bounding box metadata—from a.pptxtemplate via a VLM plus raw XML parsing. These form a structured preference profileP = P_C ∪ P_A. -
Preference-Guided Slide Planning: Three LLM agents act sequentially: a Paper Reorganizer restructures the target paper according to
P_C; a Slide Outline Designer segments reorganized content into slide-wise outlines and simultaneously drafts a per-slide speech script via the chain-of-speech mechanism (mimicking how human presenters rehearse); a Template Selector matches each planned slide to a layout fromP_Ausing the aesthetic schema. -
Slide Realization: A layout-aware agent maps outline content to template element placeholders; a code agent generates executable code to edit the
.pptxfile directly, yielding fully editable output.
The entire pipeline is training-free and uses GPT-4.1 (or open-source Qwen2.5-72B/VL-72B) in zero-shot mode.
Key Contributions¶
- Novel task formulation: Preference-guided paper-to-slides generation conditioned on a (paper, slides) example pair and a
.pptxtemplate, capturing both content and aesthetic preferences in an open-ended, unlabeled setting. - SlideTailor framework: Human-behavior-inspired agentic pipeline with dual-branch implicit preference distillation, preference-guided multi-agent slide planning, and template-editing-based slide realization.
- Chain-of-speech mechanism: Simultaneous per-slide speech script generation during outline construction, aligning visual content with anticipated oral delivery and enabling downstream video presentation generation.
- PSP benchmark: 200 target papers × 50 reference pairs × 10 templates = up to 100,000 unique input combinations, the largest paper-to-slides benchmark with open-ended preference modeling, sourced from 7 AI conferences and 4 scientific journals.
- Interpretable evaluation metrics: Four preference-based metrics (Coverage/IoU, Flow/NGLD, Content Structure, Aesthetic) and two preference-independent metrics (Content, Aesthetic), all LLM/VLM-judged and normalized to 0–100.
Results¶
- SlideTailor (GPT-4.1) vs. baselines on PSP (50 papers):
- Overall: 75.80 vs. PPTAgent 67.30, ChatGPT 62.86, AutoPresent 48.78
- Coverage: 74.47 vs. PPTAgent 64.41, ChatGPT 62.62
- Flow: 66.65 vs. PPTAgent 54.24, ChatGPT 56.84
- Content Structure: 72.80 vs. PPTAgent 57.60, ChatGPT 61.60
- Aesthetic (preference-based): 98.00 vs. PPTAgent 97.20, ChatGPT 80.80
- Preference-independent Content: 67.64 vs. PPTAgent 58.36, ChatGPT 47.00
- Open-source variant (Qwen2.5+VL): Overall 69.21, competitive across all metrics without model-specific tuning.
- Human evaluation: 81.63% win rate vs. PPTAgent (60 ratings, 30 cases, 2 raters each); Pearson correlation 0.64 between MLLM and human scores.
- Ablation (30 samples): Removing content preference drops Coverage by ~9pp, Flow by ~11pp, overall by ~5.7pp; removing chain-of-speech drops preference-independent Content from 66.4 → 47.3.
- Cost: $0.665/deck (GPT-4.1) vs. $0.016/deck (Qwen).
Limitations¶
- Benchmark restricted to scientific papers; no coverage of business, educational, or other document types.
- Training-free design foregoes potential gains from end-to-end multimodal preference learning.
- MLLM-based evaluation exhibits self-bias (models favor their own outputs) and lacks human-level fine-grained perception; cross-judge evaluation is more consistent with human ratings but not yet adopted as default.
- No reported evaluation of the downstream video presentation quality beyond mentioning the capability.
- Evaluation averaged over only 50 target papers from a pool of 200, with each paper paired to a single randomly chosen preference combination.
Relevance to Agentic AI / LLM Agents¶
SlideTailor is a concrete example of a multi-agent document-processing pipeline where specialized LLM/VLM agents—distiller, reorganizer, outline designer, template selector, layout mapper, code executor—collaborate in a fixed sequential topology to accomplish a complex, multi-constraint generation task. It demonstrates how implicit human preferences can be surfaced and operationalized as structured conditioning signals for downstream agents without fine-tuning, a pattern broadly applicable to other agentic document automation workflows. The chain-of-speech mechanism is an instance of chain-of-thought-style intermediate generation used not for reasoning transparency but to align one modality's output (slide content) with an anticipated downstream action (speech), which is relevant to agents that must co-plan across output channels. The framework also highlights the challenge of multi-aspect alignment—a recurrent problem in multi-agent coordination—where independent preference signals must be reconciled into a coherent final artifact.