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 an agentic framework for generating personalized presentation slides from scientific papers, conditioned on implicit user preferences supplied as a paper–slides example pair and a .pptx template. It introduces a dual-branch preference distillation pipeline plus a chain-of-speech mechanism that co-generates speech scripts during slide planning. It outperforms all baselines on the newly introduced PSP benchmark with an overall score of 75.8%.
Problem¶
Existing paper-to-slides systems treat generation as unconditioned document conversion, ignoring that presentation style is subjective: users differ in narrative structure, section emphasis, conciseness, and visual aesthetics. Asking users to write detailed textual preference instructions is burdensome and unnatural. Prior personalization attempts (e.g., Persona-Aware-D2S) were limited to four fixed preference categories, failing to capture open-ended, fine-grained user variation.
Method¶
SlideTailor operates in three stages:
-
Implicit Preference Distillation — a dual-branch process extracts content preferences (PC) from a reference paper–slides pair by prompting an LLM (GPT-4.1) to infer how content is selected, emphasized, omitted, or reordered; aesthetic preferences (PA) are extracted from the
.pptxtemplate via a VLM plus raw metadata (bounding boxes, font sizes). The union P = PC ∪ PA forms a structured symbolic preference profile. -
Preference-Guided Slide Planning — three LLM agents operate sequentially: (a) a paper reorganizer restructures the input paper according to PC; (b) a slide outline designer segments content into per-slide plans and simultaneously drafts speech scripts via the chain-of-speech mechanism (inspired by how presenters rehearse narration alongside layout design); (c) a template selector assigns the most semantically appropriate layout from PA to each slide.
-
Slide Realization — a layout-aware agent maps planned content to template elements, and a code agent generates Python/pptx code to edit the
.pptxfile, preserving the original visual theme while producing fully editable output.
Key Contributions¶
- Formal definition of preference-guided paper-to-slides generation with two orthogonal preference dimensions (content and aesthetic).
- SlideTailor: a training-free agentic pipeline with dual-branch implicit preference distillation from unlabeled examples.
- Chain-of-speech mechanism: simultaneous generation of per-slide speech scripts during outline construction, improving content coherence and enabling video presentation downstream.
- PSP dataset: 200 target papers × 50 sample pairs × 10 templates = up to 100,000 unique input combinations, spanning 7 AI conferences and 4 scientific journals; largest existing paper-to-slides benchmark.
- Six interpretable evaluation metrics (4 preference-based: Coverage, Flow, Content Structure, Aesthetic; 2 preference-independent: Content, Aesthetic), scored via LLM/VLM-as-judge with human correlation validated at Pearson r = 0.64.
Results¶
- SlideTailor (GPT-4.1) achieves 75.8% overall vs. PPTAgent 67.3%, AutoPresent 48.8%, ChatGPT 62.9%.
- Aesthetic preference alignment: 98.0% (ours) vs. 97.2% (PPTAgent) vs. 22.4% (AutoPresent).
- Content Structure preference: 72.8% (ours) vs. 57.6% (PPTAgent).
- Open-source variant (Qwen2.5-72B + Qwen2.5-VL-72B): 69.2% overall, competitive with PPTAgent without model-specific tuning.
- Human evaluation: 81.63% win rate vs. PPTAgent across 60 independent ratings.
- Ablation — removing content preference degrades Coverage/Flow/Content Structure by ~10%; removing chain-of-speech drops preference-independent Content from 66.4% to 47.3%.
- Cost: $0.665/deck (GPT-4.1) or $0.016/deck (Qwen).
Limitations¶
- Benchmark restricted to scientific papers; no coverage of business, educational, or non-academic domains.
- Training-free design; end-to-end multimodal preference-conditioned training unexplored.
- MLLM-based evaluation shows alignment with human judgment (r = 0.64) but notable gap remains; MLLMs exhibit self-bias and lack fine-grained perceptual sensitivity of human raters.
- Template selection is purely matching-based; no mechanism for generating novel layouts when the template set is impoverished.
Relevance to Harnesses / Meta-Harnesses¶
SlideTailor is a concrete instance of a document-processing meta-harness: it orchestrates multiple specialized LLM/VLM agents (reorganizer, outline designer, template selector, code renderer) in a fixed pipeline with structured inter-stage communication, directly analogous to how digest harnesses chain fetch → summarize → structure → publish agents. The dual-branch preference distillation stage is architecturally similar to a harness configuration layer that extracts user intent from examples rather than explicit config, a pattern applicable to any harness that must adapt behavior to implicit user preferences. The chain-of-speech mechanism demonstrates that injecting auxiliary planning signals (narration) into an intermediate stage improves downstream output quality — a generalizable design principle for multi-stage agentic pipelines.