It's Time to Get It Right: Improving Analog Clock Reading and Clock-Hand Spatial Reasoning in Vision-Language Models

Jaeha Choi; Jin Won Lee; Siwoo You; Jangho Lee

cs.CVMar 9, 2026

It's Time to Get It Right: Improving Analog Clock Reading and Clock-Hand Spatial Reasoning in Vision-Language Models

Authors: Jaeha Choi, Jin Won Lee, Siwoo You, Jangho Lee

Organizations: Incheon National University, Incheon, Republic of Korea · McGill University, Montreal, Canada

Abstract

Advances in vision-language models (VLMs) have achieved remarkable success on complex multimodal reasoning tasks, leading to the assumption that they should also excel at reading analog clocks. However, contrary to this expectation, our study reveals that reading analog clocks in real-world environments remains a significant challenge for state-of-the-art VLMs. Existing analog clock datasets are largely synthetic or planar with limited stylistic diversity and minimal background context, failing to capture the visual variability of real-world scenes. As a result, VLMs trained on such data exhibit weak spatiotemporal reasoning, frequently confusing the hour and minute hands and struggling under common visual conditions such as occlusion, lighting variation, and cluttered backgrounds. To address this issue, we introduce TickTockVQA, a human-annotated dataset containing analog clocks in diverse real-world scenarios. TickTockVQA provides explicit hour and minute annotations, and includes an AM/PM tag when it is inferable from the visual context. Furthermore, we propose Swap-DPO, a direct preference optimization-based fine-tuning framework to align model reasoning toward accurate time interpretation. Experimental results demonstrate that our approach substantially enhances clock reading accuracy and robustness under real-world conditions, establishing a foundation for future research on spatiotemporal reasoning and visual understanding in VLMs.

Figures & tables

Figure 1: Impact of training data quality on Qwen2.5-VL-7B performance. We compare Qwen2.5-VL-7B trained on three datasets: TickTockVQA (real-world), SynClock (OpenCV-based synthetic), and CtrlClock (diffusion-generated synthetic). Training on TickTockVQA achieves the best performance with 99.9 minutes MAE.

Figure 2: Comparison of model predictions on the clock reading task. Our model, It’s Time To Get It Right (ITGR), correctly identifies the time, while other large multimodal models (Llama-3.2-11B Zero-shot, GPT-5, Claude Sonnet 4.5, Gemini-2.5 Pro, and Perplexity Pro) produce incorrect results.

Model	Hour Acc	Min. Acc	Full Time Acc	MAE $\downarrow$
SpatialVLM-3B [ 4 ]	12.51	6.44	1.05	161.68
Llama-3.2-11B [ 11 ]	11.51	8.58	1.43	156.96
Gemma3-12B [ 37 ]	13.00	10.62	2.12	156.49
InternVL3-8B [ 43 ]	13.59	9.08	2.20	159.92
Qwen2.5-VL-7B [ 2 ]	17.65	22.44	6.04	148.62
It’s About Time [ 39 ]	28.95	25.00	18.54	135.15

Table 1: Zero-shot performance comparison on the TickTockVQA test set.

Clock Type	Environment			Transformation			Design (multi-label)			Total
Clock Type	Indoor	Outdoor	Unknown	Normal	Flipped	Partial	Arabic	Roman	No numerals	Total
Wall clocks	2107	1622	317	3887	10	149	1953	1465	671	4046
Tower clocks	3	1905	5	1907	2	4	112	925	881	1913
Wristwatches	495	214	529	1214	1	23	795	95	363	1238
Post clocks	62	931	6	992	1	6	321	512	171	999
Alarm/Desk clocks	1235	53	213	1398	-	103	998	282	241	1501

Table 2: Breakdown of TickTockVQA instances by clock type. The Environment and Transformation columns are mutually exclusive, whereas the Design columns may contain multiple labels.

Panel of Figure 3 — Figure 3: Examples of challenging visual variations in the TickTockVQA test set: (a) cropped clock, (b) clock-like object, (c) illumination changes, and (d) horizontally flipped clock. The figure highlights diverse transformations and ambiguities that models must handle for robust clock understanding.

Model and Training Stages	Hour Acc		Minute Acc		Full Time Acc		MAE $\downarrow$ (hour)		MAE $\downarrow$ (minute)		MAE $\downarrow$ (total)
Model and Training Stages	B	S	B	S	B	S	B	S	B	S	B	S
Gemma3-12B [ 37 ]	13.00	20.75	10.62	19.36	2.12	3.05	2.61	2.59	14.43	14.32	156.49	155.10
w/ TickTockVQA	46.37	55.04	54.39	65.68	34.21	37.11	1.35	1.27	6.68	6.34	81.66	77.26
w/ Swap-DPO	46.67	56.01	57.54	67.64	35.32	37.89	1.36	1.29	6.16	5.85	81.86	77.97
Qwen2.5-VL-7B [ 2 ]	17.65	25.26	22.44	32.14	6.04	7.26	2.47	2.44	11.96	11.81	148.62	146.76
w/ SynClock	28.11	37.51	35.68	47.17	16.12	18.28	2.15	2.09	9.60	9.35	129.51	126.34

Table 3: Comprehensive evaluation results for Gemma3-12B, Qwen2.5-VL-7B, and Llama-3.2-11B on the TickTockVQA test set. B and S denote baseline and swap-equivalence evaluation, respectively (Sec 5.2 .)

Figure 4: Qualitative examples of hand-swap error correction by Swap-DPO. SFT incorrectly swaps the hour and minute hands, whereas Swap-DPO successfully corrects this systematic error pattern.

Figure 5: Quantitative comparison of clock reading accuracy. Each plot visualizes the relationship between ground truth (x-axis) and model-predicted time (y-axis) in minutes. The gray dashed line (

y{=}x

) indicates perfect predictions. Left: Zero-shot baseline. Right: Our ITGR model with the Swap-DPO framework.

Figure 6: Statistical analysis of time reading errors. Left: Distribution of prediction errors (in minutes) with histogram and kernel density estimation (KDE). ITGR reduces the heavy tail of large errors. Right: Cumulative probability of samples within a given absolute error threshold. Compares Zero-shot baseline vs. ITGR model.

Dataset Scale	Hour Acc	Minute Acc	Full Time Acc	MAE $\downarrow$
SynClock 12k	25.93	31.60	7.07	144.05
SynClock 50k	30.19	39.74	10.58	136.65
SynClock 100k	36.35	45.67	17.56	131.45
SynClock 1M	37.51	47.17	18.28	126.34

Table 4: Effect of SynClock scale on Qwen2.5-VL-7B under SFT.

Figure 7: Qualitative comparison of synthetic data. (Left) The SynClock examples, generated using an OpenCV-based approach, exhibit limited realism and flat textures. (Right) The CtrlClock examples, generated through a diffusion-based pipeline, achieve much higher photographic realism and contextual diversity.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Test Sources
Source	Images	Split
COCO	2,063	Test
Open Images (OID)	1,940	Test
Clock Movies	1,244	Test
Training Sources
Visual Genome	1,246	Train

Appendix

Table S1: TickTockVQA data source composition and train/test split. Only COCO, Open Images, and Clock Movies are used for testing; all other sources are reserved for training. This separation ensures evaluation on out-of-distribution sources.

Figure S1 : Clock annotation density heatmap. Distribution of labeled times across all hours (1–12) and minutes (0–59). Color intensity represents $\ln(1+\text{count})$ to balance visibility across frequency ranges. Darker regions indicate higher sample density, with notable concentration around 10:10.

Panel of Figure 14 — Figure S2 : Marginal temporal distributions. (a) Hour distribution shows reasonable balance with CV=26.9%. Hour 10 is overrepresented due to the 10:10 aesthetic bias. (b) Minute distribution reveals expected peaks at canonical positions (0, 15, 30, 45) but maintains coverage across all 60 minutes.

Figure S3 : ITGR accuracy breakdown by clock type. Performance varies significantly across categories. Graphic/Illustrated clocks achieve the highest accuracy (62.71%) due to high contrast and clean contours. Wristwatches show the lowest performance (27.99%) due to small size, glass reflections, and occlusion. Bubble size represents sample count for each category. The dashed line indicates overall average accuracy (46.52%).

Figure S4 : ITGR accuracy breakdown across three categorical dimensions. (a) Environment: Performance is stable across indoor (48.5%), outdoor (44.9%), and unknown (45.8%) settings, indicating robustness to background context and lighting variation. (b) Transformation: Severe degradation for flipped clocks (23.1%), revealing fragility to unusual orientations. Partial occlusion (37.3%) also degrades performance. (c) Design: Performance is consistent across Arabic (48.2%) and Roman (46.8%) numerals but degrades for clocks without numerals (36.2%), suggesting reliance on numerical markers for spatial reference. Dashed lines indicate overall average accuracy.

Average across models:
	SFT			Random-DPO			Swap-DPO
Model	B	S	$\Delta$	B	S	$\Delta$	B	S	$\Delta$
Qwen2.5-VL-7B	20.3	22.8	+2.42	20.6	23.4	+2.75	23.1	25.1	+2.02
Gemma3-12B	34.2	37.1	+2.90	35.0	38.0	+3.00	35.3	37.9	+2.57
Llama-3.2-11B	45.8	48.1	+2.32	45.6	47.7	+2.06	46.2	48.5	+2.26
Mean	33.4	36.0	+2.55	33.7	36.4	+2.60	34.9	37.2	+2.28

Appendix

Table S2: Full-time accuracy across SFT, Random-DPO, and Swap-DPO. B denotes Baseline accuracy (strict), S denotes Swap-equivalence accuracy (allowing hand swaps), and

\Delta

is the hand-swap gap (

S-B

). A smaller

\Delta

indicates less hand confusion. Swap-DPO consistently reduces

\Delta

compared to both SFT and Random-DPO while improving overall accuracy.

Configuration	Qwen2.5-VL-7B	Llama-3.2-11B	Gemma3-12B	Notes
DPO-Specific Parameters
Loss function	sigmoid	sigmoid	sigmoid	Standard DPO loss
$\beta$ (temperature)	0.3	0.3	0.3	Controls preference strength
Precompute ref. logprobs	false	false	false	Compute on-the-fly
LoRA Configuration
LoRA enabled	✗	✓	✓	Qwen uses full fine-tuning

Appendix

Table S3 : Complete DPO training hyperparameters and configurations. We report all settings used for Direct Preference Optimization across three VLM backbones. Model-specific differences are highlighted. All models use 8

\times

NVIDIA A6000 GPUs (48GB each).

Parameter	Value
SFT Checkpoint Selection
Qwen2.5-VL-7B	checkpoint-145 (epoch 5)
Llama-3.2-11B	checkpoint-174 (epoch 6)
Gemma3-12B	checkpoint-203 (epoch 7)
SFT Accuracy at Selected Checkpoint
Qwen2.5-VL-7B	20.3% Full Time Acc

Appendix

Table S4 : DPO preference data generation configuration.

Common Core Instructions (present in all prompts):
Prompt ID	Full Prompt Text
Prompt A	Identify the single most prominent analog clock in the image (ignore digital displays). If multiple clocks are visible, choose the largest clearly visible face. Read hour = shorter thicker hand and minute = longer thinner hand ; ignore any seconds hand. If a hand lies between marks, use the lower hour and the nearest minute. Output only HH:MM (12-hour, leading zero, no AM/PM). If no analog clock is visible, output NO CLOCK.
Prompt B	Find the primary analog clock (exclude digital). Select the largest clearly visible face when multiple are present. Hour = shorter thicker hand ; minute = longer thinner hand ; ignore seconds. If a hand is between ticks, choose the lower hour and nearest minute. Answer strictly as HH:MM (12-hour, leading zero). If no analog clock is found, answer NO CLOCK.
Prompt C	Locate the most visible analog clock and ignore any digital displays. If several clocks exist, pick the biggest clear dial. Use the short thick hand for hours and the long thin hand for minutes; ignore seconds. Between ticks: take the lower hour and nearest minute. Return only HH:MM (12-hour, leading zero). If none is present, return NO CLOCK.
1. Select the most prominent analog clock (ignore digital displays)
2. Hour hand = short/thick; Minute hand = long/thin (ignore seconds)

Appendix

Table S5 : Training prompt variations for supervised fine-tuning. We rotate through three semantically equivalent prompts to prevent overfitting while maintaining consistent instruction semantics. Key phrase variations are shown; all prompts share the core instructions for hand identification, ambiguity resolution, and output formatting.

Inference Prompt

Find the most prominent analog clock in the image. The hour hand is the short, thick one, and the minute hand is the long, thin one. If a hand is between marks, choose the lower hour and the nearest minute. Your answer must be only in HH:MM format (e.g., 08:05).

Appendix

Table S6 : Inference prompt for evaluation. This prompt is used consistently across all test evaluations, ablation studies, and model comparisons. It omits training-specific instructions (multi-clock selection, NO CLOCK fallback) while preserving core task requirements.

Prompt Type	Qwen	Llama	Gemma
Training Prompt A	92	91	92
Training Prompt B	75	74	76
Training Prompt C	75	74	75
Inference Prompt	68	66	69

Appendix

Table S7 : Prompt length analysis across model tokenizers. Token counts vary slightly due to different tokenization schemes (SentencePiece for Llama, tiktoken for Qwen, custom for Gemma). Inference prompt averages 68 tokens.

Visual Spatial Reasoning

Visual spatial reasoning studies how AI systems infer geometry, object relations, viewpoints, and physical change from images and video.

See all 251 papers

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
TimeSpot: Benchmarking Geo-Temporal Understanding in Vision-Language Models in Real-World Settings
Mar 4, 2026Azmine Toushik Wasi, Shahriyar Zaman Ridoy, Koushik Ahamed Tonmoy +5Spatiotemporal Reasoning Geospatial Reasoning Benchmarks
Computational Intelligence and Operations Laboratory (CIOL), Bangladesh · Shahjalal University of Science and Technology (SUST), Sylhet, Bangladesh · North South University (NSU), Dhaka, Bangladesh +1

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

Visual Spatial Reasoning

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

TimeSpot: Benchmarking Geo-Temporal Understanding in Vision-Language Models in Real-World Settings