PPTArena

PPTArena is a benchmark for agentic PowerPoint (.pptx) editing. Each case gives a natural-language editing instruction and a source deck; a system must edit the source so the result matches a human-authored ground-truth deck. It measures whether agents can perform realistic, fine-grained slide edits — translation, restyling, layout fixes, chart/table edits, transitions, master-level changes, and more — while preserving everything they were not asked to change.

• Paper: arXiv:2512.03042
• Code & web app: https://github.com/michaelofengend/PPTArena
• License: MIT
• Cases: 100 paired editing tasks (100 source + 100 ground-truth decks)
• Modality: binary Office files (.pptx)

Dataset structure

PPTArena/
├── evaluation_pairs_refined.json   # 100 cases: prompt + (original, ground_truth) pairs
├── Original/        *.pptx         # 100 source decks to be edited
└── GroundTruth/     *.pptx         # 100 human-authored target decks

The .pptx files are binary Office documents. Use huggingface_hub.snapshot_download to fetch them (see Usage). datasets.load_dataset is not the intended access path for the decks — it would return raw bytes rather than usable files. The JSON index is what joins each prompt to its file pair.

The index: evaluation_pairs_refined.json

A JSON array of 100 objects. Each object has:

Key	Type	Description
name	string	Human-readable case id, e.g. "Case 1: Translate to English But Keep French".
prompt	string	The editing instruction given to the agent.
style_target	string	Detailed description of the intended result (used for evaluation).
original	string	Path to the source deck, e.g. "Original/Foo_TestA.pptx".
ground_truth	string	Path to the target deck, e.g. "GroundTruth/Foo_GroundTruthA.pptx".
category	list[string]	High-level category, e.g. ["Content"], ["Layout"]. May contain more than one.
edit_type	string	Fine-grained edit type, e.g. "Text & Typography", "Charts".
enhancement_notes	string	Notes on how the case/target was refined.

Categories (cases per category; a case may have more than one)

Category	Count
Content	67
Layout	29
Styling	29
Structure	15
Interactivity	4

Edit types (cases per type)

Text & Typography (29), Charts (10), Images & Pictures (10), Theme & Background (9), Alignment/Distribution/Z-order (8), Slide/Section Management & Footers (8), Tables (8), Shapes & Drawing (4), SmartArt & Diagrams (4), Slide Layout & Placeholders (3), Accessibility & Semantics (2), and one case each of Slide Transitions, Hyperlinks & Action Settings, Template & Master-Level Edits, Audio & Video, and Object Animations.

Usage

# pip install -U huggingface_hub
from huggingface_hub import snapshot_download
import json, os

# Download the whole benchmark (decks + index) into the HF cache.
local = snapshot_download(repo_id="<username>/PPTArena", repo_type="dataset")

# Load the 100-case index.
cases = json.load(open(os.path.join(local, "evaluation_pairs_refined.json")))

# Resolve a single case to absolute .pptx paths.
case = cases[0]
original     = os.path.join(local, case["original"])      # the .pptx to edit
ground_truth = os.path.join(local, case["ground_truth"])  # the target .pptx
print(case["prompt"])
# `original` and `ground_truth` are real .pptx files: open them with python-pptx,
# feed them to an agent, etc.

# Optional: fetch only the index + source decks (skip ground truth) to save bandwidth:
# snapshot_download(repo_id="<username>/PPTArena", repo_type="dataset",
#                   allow_patterns=["*.json", "Original/*"])

Evaluation protocol

For each case: apply prompt to the deck at original, producing a candidate deck, then compare the candidate against the deck at ground_truth (the style_target field describes the intended outcome in detail). The benchmark rewards making exactly the requested edit while preserving unrelated content, layout, and formatting. The paper uses a dual VLM-as-judge pipeline scoring instruction-following and visual quality.

Citation

@article{ofengenden2025pptarena,
  title   = {PPTArena: A Benchmark for Agentic PowerPoint Editing},
  author  = {Ofengenden, Michael and Man, Yunze and Pang, Ziqi and Wang, Yu-Xiong},
  journal = {arXiv preprint arXiv:2512.03042},
  year    = {2025}
}

License

Released under the MIT License.