// FloorplanQA paper landing page — single page with theme toggle in top-right.
const ARXIV_URL = 'https://arxiv.org/abs/2507.07644';
const HF_URL = 'https://huggingface.co/papers/2507.07644';
const CODE_URL = 'https://github.com/OldDeLorean/FloorplanQA';
const PDF_URL = ARXIV_URL.replace('/abs/', '/pdf/') + '.pdf';
const AUTHORS = [
{ name: 'Fedor Rodionov', aff: 1, corresponding: true },
{ name: 'Abdelrahman Eldesokey', aff: 1 },
{ name: 'Michael Birsak', aff: 1 },
{ name: 'John Femiani', aff: 2 },
{ name: 'Bernard Ghanem', aff: 1 },
{ name: 'Peter Wonka', aff: 1 }];
const AFFILIATIONS = [
{ id: 1, name: 'KAUST', full: 'King Abdullah University of Science and Technology' },
{ id: 2, name: 'Miami University', full: 'Miami University, Oxford, OH' }];
const TASKS = [
{ name: 'Pair Distance', fmt: 'N', cat: 'Metric', desc: 'Euclidean distance between two object centroids' },
{ name: 'View Angle', fmt: 'N', cat: 'Metric', desc: 'Smallest angle between inter-object vector and global north' },
{ name: 'Free Space', fmt: 'N', cat: 'Topology', desc: 'Total non-occupied floor area (m²)' },
{ name: 'Max Box', fmt: 'N', cat: 'Topology', desc: 'Area of the largest rectangle fitting inside the room' },
{ name: 'Placement', fmt: 'B', cat: 'Topology', desc: 'Can a w × h object fit fully without overlaps' },
{ name: 'Visibility', fmt: 'L', cat: 'Topology', desc: 'All objects intersecting a centroid-to-centroid ray' },
{ name: 'Repositioning', fmt: 'N', cat: 'Dynamic', desc: 'How far an object can slide in a direction before contact' },
{ name: 'Shortest Path', fmt: 'S', cat: 'Dynamic', desc: 'Collision-free path with required clearance' }];
const BIBTEX = `@inproceedings{rodionov2025floorplanqa,
title = {FloorplanQA: A Benchmark for Spatial Reasoning in LLMs
using Structured Representations},
author = {Rodionov, Fedor and Eldesokey, Abdelrahman and Birsak, Michael
and Femiani, John and Ghanem, Bernard and Wonka, Peter},
booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
year = {2025},
address = {Seoul, South Korea}
}`;
const ABSTRACT = `We introduce FloorplanQA, a diagnostic benchmark for evaluating spatial reasoning in
large language models grounded in structured representations of indoor scenes
(kitchens, living rooms, bedrooms, bathrooms), encoded symbolically in JSON or XML
layouts. The benchmark covers distance measurement, visibility, path finding, and
object placement within constrained spaces. Our results across 15 frontier
open-source and commercial LLMs reveal that while models may succeed on shallow
queries, they often fail to respect physical constraints and preserve spatial
coherence — though they remain mostly robust to small spatial perturbations.`;
const FINDINGS = [
'Models compute the sum of object areas instead of the union — failing on Free Space and Max Box.',
'Distance and angle errors mostly originate from centroid computation on non-axis-aligned polygons (HSSD layouts).',
'Shortest-path planning collapses in cluttered scenes where buffering narrows corridors.',
'A Python code interpreter rescues arithmetic-heavy tasks (+40 pp on Distance / Angle) but does not help planning.',
'Adding a top-down render to JSON gives selective gains but never compensates for removing the JSON.'];
// ---------- icons ----------
const ICONS = {
arxiv:
,
pdf:
,
code:
,
hf:
,
sun:
,
moon:
};
// ---------- theme toggle ----------
function ThemeToggle({ theme, setTheme }) {
return (
Adding a top-down render to the JSON gives selective gains (icons, labeled boxes, photorealistic),
but image-only modalities fall to 19–40% — symbolic JSON remains the dominant signal.
)}