TimeLineXR: An Interactive Exploration of Time in Vision-Language Models

Nidham Tekaya; Markus Passecker; Matthias Zeppelzauer

cs.CVOct 27, 2026

TimeLineXR: An Interactive Exploration of Time in Vision-Language Models

Authors: Nidham Tekaya, Markus Passecker, Matthias Zeppelzauer

Organizations: Univ. of Applied Sciences St. Pölten TU Wien St. Pölten / Vienna, Austria · Univ. of Applied Sciences St. Pölten St. Pölten, Austria

Abstract

Recently, it has been shown that vision-language models (VLMs) have a certain intrinsic understanding of time. They can estimate the temporal origin of objects and encode time along a low-dimensional, non-linear manifold inside their embedding space~\cite{tekaya2025matter}. That result is quantitative and reproducible, but the structure itself stays out of reach for end users: they cannot see it, query it with their own images, or inspect where a query lands relative to other years on the time axis. We present TimeLineXR, an interactive demonstration that turns this time axis into an explorable 3D scene. A web demo lets users upload or select images and see each placed on a color-coded 3D time axis; a Three.js WebXR scene renders the same time axis as a room-scale representation that can be grabbed, scaled, inspected, and queried by throwing image cards into it. Both clients run the same prediction pipeline against the same cached temporal index: encode the query with CLIP, project it through cosine-kernel KPCA, and snap it to the Bézier time axis. The system turns each image into a point on the time axis, so users can compare where different objects land and read off their estimated first year of appearance.Code source and reproduction steps can be found in: https://github.com/sterngefeuert/timeline-xr.

Figures & tables

Fig. 2: System architecture in two phases. The offline setup encodes text prompts for each year, reduces them to three dimensions with cosine-kernel KPCA, fits a Bézier curve through the reduced time embeddings, and projects the embeddings onto the curve to obtain a uniform time axis, caching the result as a single temporal index. The online prediction encodes the query image with CLIP, snaps it to the Bézier time axis using the cached index, and returns an estimated year that the web view and the XR scene display in their own way. Parallelograms denote inputs and outputs, rounded rectangles denote processes, and the cylinder denotes the cached temporal index.

Backbone	$S=3$ (demo)	$S=13$ (reference)
CLIP ViT-B/32	$15.30$	$9.00$
EVA-CLIP	$20.81$	$7.62$

TABLE I: Dimensionality trade-off on TIME10k: MAE (years) at

S{=}13

(reference) vs.

S{=}3

(TimeLineXR), Bézier(

\mathbf{R}^{S}

, NN) variant of [ 1 ] .

Fig. 3: Browser web demo after selecting an example. The left panel specifies the embedding backbone and query path. The centre panel shows the VLM’s time axis in 3D: the time embeddings projected onto the fitted Bézier curve (drawn in red) as points colour-coded by year, with the query image placed on it. The right panel provides bundled examples.

Stage	Time
Backend prediction (CPU)	$150$ – $300$ ms
Backend prediction (GPU)	$<50$ ms
XR throw animation	$0.5$ – $1.3$ s

TABLE II: Backend prediction latency and XR throw animation duration.

Chronological Blindness: Benchmarking Temporal Reasoning in Vision-Language Models with CHRONOSIGHT
Jun 15, 2026Parthaw Goswami, Jaynto Goswami DeepVLM Evaluation Temporal Reasoning
Department of Computer Science, University of Missouri, Columbia, 65201, MO, USA · SAP, Prague, Czech Republic
Seeing Time: Benchmarking Chronological Reasoning and Shortcut Biases in Vision-Language Models
Jun 4, 2026Haoyu Zhou, Qing Qing, Caichong Li +6VLM Evaluation Vision-Language Models
College of Computer Science and Technology, Jilin University, Changchun 130012, China · College of Computing and Data Science, Nanyang Technological University, 639798, Singapore · School of Computer Science, Wuhan University, China +1
SpaTime: Streaming Vision-Language Models for Spatio-temporal Reasoning
Oct 6, 2026Hairong Yin, Huangying Zhan, Shin-Fang Chng +2Vision-Language Models Spatiotemporal Reasoning
Department of Computer Science, Purdue University, USA · Goertek Alpha Labs, USA

Jun 15, 2026J/K move · Enter open · S save

Chronological Blindness: Benchmarking Temporal Reasoning in Vision-Language Models with CHRONOSIGHT

Parthaw Goswami, Jaynto Goswami Deep

Department of Computer Science, University of Missouri, Columbia, 65201, MO, USA · SAP, Prague, Czech Republic

VLM Evaluation Temporal Reasoning VLM Reasoning Temporal Reasoning in VLMs

Open paper AskPDF arXiv

Human perception of visual scenes is inherently temporal. We instinctively recognise whether a fruit is ripening or rotting, whether construction is progressing or being demolished, and approximately how much time separates two photographs of the same subject. Whether large vision-language models (VLMs) share this competence remains an open and practically important question. We introduce CHRONOSIGHT, a rigorously controlled benchmark evaluating five dimensions of visual temporal reasoning: CHRONORANK (chronological ordering of image sequences), CHRONOLOCATE (ordinal stage localisation from a single image), CHRONODELTA (estimation of time elapsed between two images on a logarithmic scale), CHRONOREVERSE (detection of temporally reversed sequences), and CHRONOODD (identification of a temporal outlier within a set). The benchmark comprises 1{,}000 items across eight process families (biological growth, food transformation, physical weathering, construction, environmental change, human ageing, astronomical phenomena, and urban dynamics) spanning timescales from minutes to millennia. We evaluate eight open-source VLMs (500 M to 19 B parameters) under two prompting regimes and collect human performance baselines. Human performance averages 0.89 across tasks; the best open model (Qwen2.5-VL-7B) reaches 0.40 under direct prompting, a gap we term chronological blindness. Lightweight LoRA fine-tuning on 151 examples raises CHRONODELTA accuracy from near-zero to 0.43, transferring zero-shot to related tasks (CHRONOODD: 0.37; CHRONOREVERSE: 0.64)suggesting the bottleneck is partly instruction following rather than visual perception. Benchmark, code, and predictions will be released upon acceptance.

Abstract

Explore similar work

Chronological Blindness: Benchmarking Temporal Reasoning in Vision-Language Models with CHRONOSIGHT

Seeing Time: Benchmarking Chronological Reasoning and Shortcut Biases in Vision-Language Models

SpaTime: Streaming Vision-Language Models for Spatio-temporal Reasoning