A brand new paper from researchers in China and Spain finds that even superior multimodal AI fashions equivalent to GPT-4.1 wrestle to inform the time from pictures of analog clocks. Small visible adjustments within the clocks could cause main interpretation errors, and fine-tuning solely helps with acquainted examples. The outcomes elevate issues concerning the reliability of those fashions when confronted with unfamiliar pictures in real-world duties.
When people develop a deep sufficient understanding of a site, equivalent to gravity or different fundamental bodily ideas, we transfer past particular examples to know the underlying abstractions. This permits us to use that data creatively throughout contexts and to acknowledge new situations, even these we now have by no means seen earlier than, by figuring out the precept in motion.
When a site carries sufficient significance, we could even start to understand it the place it doesn’t exist, as with pareidolia, pushed by the excessive price of failing to acknowledge an actual occasion. So sturdy is that this pattern-recognizing survival mechanism that it even disposes us to discover a wider vary of patterns the place there are none.
The sooner and extra repetitively a site is instilled in us, the deeper its grounding and lifelong persistence; and one of many earliest visible datasets that we’re uncovered to as kids comes within the type of teaching-clocks, the place printed materials or interactive analog clocks are used to show us methods to inform time:
Instructing aids to assist kids be taught to inform time. Supply: https://www.youtube.com/watch?v=IBBQXBhSNUs
Although altering fashions in watch design could generally problem us, the resilience of this early domain-mastery is sort of spectacular, permitting us to discern analogue clock faces even within the face of complicated or ‘eccentric’ design decisions:
Some difficult faces in watch couture. Supply: https://www.ablogtowatch.com/wait-a-minute-legibility-is-the-most-important-part-of-watch-design/
People don’t want hundreds of examples to find out how clocks work; as soon as the fundamental idea is grasped, we will acknowledge it in nearly any kind, even when distorted or abstracted.
The issue that AI fashions face with this job, in contrast, highlights a deeper problem: their obvious power could rely extra on high-volume publicity than on understanding.
Past the Imitation Recreation?
The stress between surface-level efficiency and real ‘understanding’ has surfaced repeatedly in latest investigations of huge fashions. Final month Zhejiang College and Westlake College re-framed the query in a paper titled Do PhD-level LLMs Actually Grasp Elementary Addition? (not the main focus of this text), concluding:
‘Regardless of spectacular benchmarks, fashions present essential reliance on sample matching fairly than true understanding, evidenced by failures with symbolic representations and violations of fundamental properties.
‘Express rule provision impairing efficiency suggests inherent architectural constraints. These insights reveal analysis gaps and spotlight the necessity for architectures able to real mathematical reasoning past sample recognition.’
This week the query arises once more, now in a collaboration between Nanjing College of Aeronautics and Astronautics and the Universidad Politécnica de Madrid in Spain. Titled Have Multimodal Massive Language Fashions (MLLMs) Actually Discovered to Inform the Time on Analog Clocks?, the new paper explores how properly multimodal fashions perceive time-telling.
Although the progress of the analysis is roofed solely in broad element within the paper, the researchers’ preliminary checks established that OpenAI’s GPT-4.1 multimodal language mannequin struggled to accurately learn the time from a various set of clock pictures, usually giving incorrect solutions even on easy circumstances.
This factors to a potential hole within the mannequin’s coaching knowledge, elevating the necessity for a extra balanced dataset, to check whether or not the mannequin can really be taught the underlying idea. Due to this fact the authors curated an artificial dataset of analog clocks, evenly overlaying each potential time, and avoiding the same old biases present in web pictures:
An instance from the researchers’ artificial analog clock dataset, used to fine-tune a GPT mannequin within the new work. Supply: https://huggingface.co/datasets/migonsa/analog_watches_finetune
Earlier than fine-tuning on the brand new dataset, GPT-4.1 persistently didn’t learn these clocks. After some publicity to the brand new assortment, nevertheless, its efficiency improved – however solely when the brand new pictures appeared like ones it had already seen.
When the form of the clock or the fashion of the palms modified, accuracy fell sharply; even small tweaks, equivalent to thinner palms or arrowheads (rightmost picture beneath), have been sufficient to throw it off; and GPT-4.1 struggled moreover to interpret Dali-esque ‘melting clocks’:
Clock pictures with commonplace design (left), distorted form (center), and modified palms (proper), alongside the occasions returned by GPT-4.1 earlier than and after fine-tuning. Supply: https://arxiv.org/pdf/2505.10862
The authors deduce that present fashions equivalent to GPT-4.1 could subsequently be studying clock-reading primarily by means of visible sample matching, fairly than any deeper idea of time, asserting:
‘[GPT 4.1] fails when the clock is deformed or when the palms are modified to be thinner and to have an arrowhead. The Imply Absolute Error (MAE) within the time estimate over 150 random occasions was 232.48s for the preliminary clocks, 1380.69s when the form is deformed and 3726.93s when palms are modified.
‘These outcomes recommend that the MLLM has not realized to inform the time however fairly memorized patterns.’
Sufficient Time
Most coaching datasets depend on scraped net pictures, which are inclined to repeat sure occasions – particularly 10:10, a well-liked setting in watch commercials:
From the brand new paper, an instance of the prevalence of the ‘ten previous ten’ time in analog clock pictures.
On account of this restricted vary of occasions depicted, the mannequin might even see solely a slender vary of potential clock configurations, limiting its capacity to generalize past these repetitive patterns.
Concerning why fashions fail to accurately interpret the distorted clocks, the paper states:
‘Though GPT-4.1 performs exceptionally properly with commonplace clock pictures, it’s stunning that modifying the clock palms by making them thinner and including arrowheads results in a major drop in its accuracy.
‘Intuitively, one may count on that the extra visually complicated change – a distorted dial –would have a better affect on efficiency, but this modification appears to have a comparatively smaller impact.
‘This raises a query: how do MLLMs interpret clocks, and why do they fail? One chance is that thinner palms impair the mannequin’s capacity to understand route, weakening its understanding of spatial orientation.
‘Alternatively, there could possibly be different elements that trigger confusion when the mannequin makes an attempt to mix the hour, minute, and second palms into an correct time studying.’
The authors contend that figuring out the foundation trigger of those failures is essential to advancing multimodal fashions: if the problem lies in how the mannequin perceives spatial route, fine-tuning could provide a easy repair; but when the issue stems from a broader issue in integrating a number of visible cues, it factors to a extra elementary weak point in how these methods course of info.
Fantastic-Tuning Exams
To check whether or not the mannequin’s failures could possibly be overcome with publicity, GPT-4.1 was fine-tuned on the aforementioned and complete artificial dataset. Earlier than fine-tuning, its predictions have been extensively scattered, with important errors throughout all sorts of clocks. After fine-tuning on the gathering, accuracy improved sharply on commonplace clock faces, and, to a lesser extent, on distorted ones.
Nonetheless, clocks with modified palms, equivalent to thinner shapes or arrowheads, continued to supply giant errors.
Two distinct failure modes emerged: on regular and distorted clocks, the mannequin sometimes misjudged the route of the palms; however on clocks with altered hand types, it usually confused the perform of every hand, mistaking hour for minute or minute for second.
A comparability illustrating the mannequin’s preliminary weak point, and the partial features achieved by means of fine-tuning, displaying predicted vs. precise time, in seconds, for 150 randomly chosen clocks. On the left, earlier than fine-tuning, GPT-4.1’s predictions are scattered and sometimes removed from the proper values, indicated by the pink diagonal line. On the proper, after fine-tuning on a balanced artificial dataset, the predictions align way more intently with the bottom reality, though some errors stay.
This means that the mannequin had realized to affiliate visible options like hand thickness with particular roles, and struggled when these cues modified.
The restricted enchancment on unfamiliar designs raises additional doubts about whether or not a mannequin of this sort learns the summary idea of time-telling, or merely refines its pattern-matching.
Hand Indicators
So, though fine-tuning improved GPT-4.1’s efficiency on typical analog clocks, it had far much less affect on clocks with thinner palms or arrowhead shapes, elevating the likelihood that the mannequin’s failures stemmed much less from summary reasoning and extra from confusion over which hand was which.
To check whether or not accuracy may enhance if that confusion have been eliminated, a brand new evaluation was carried out on the mannequin’s predictions for the ‘modified-hand’ dataset. The outputs have been divided into two teams: circumstances the place GPT-4.1 accurately acknowledged the hour, minute, and second palms; and circumstances the place it didn’t.
The predictions have been evaluated for Imply Absolute Error (MAE) earlier than and after fine-tuning, and the outcomes in comparison with these from commonplace clocks; angular error was additionally measured for every hand utilizing dial place as a baseline:
Error comparability for clocks with and with out hand-role confusion within the modified-hand dataset earlier than and after fine-tuning.
Complicated the roles of the clock palms led to the biggest errors. When GPT-4.1 mistook the hour hand for the minute hand or vice versa, the ensuing time estimates have been usually far off. In distinction, errors brought on by misjudging the route of a accurately recognized hand have been smaller. Among the many three palms, the hour hand confirmed the best angular error earlier than fine-tuning, whereas the second hand confirmed the bottom.
Angular error by hand sort for predictions with and with out hand-role confusion, earlier than and after fine-tuning, within the modified-hand dataset.
To give attention to directional errors alone, the evaluation was restricted to circumstances the place the mannequin accurately recognized every hand’s perform. If the mannequin had internalized a common idea of time-telling, its efficiency on these examples ought to have matched its accuracy on commonplace clocks. It didn’t, and accuracy remained noticeably worse.
To look at whether or not hand form interfered with the mannequin’s sense of route, a second experiment was run: two new datasets have been created, every containing sixty artificial clocks with solely an hour hand, pointing to a special minute mark. One set used the unique hand design, and the opposite the altered model. The mannequin was requested to call the tick mark that the hand was pointing to.
Outcomes confirmed a slight drop in accuracy with the modified palms, however not sufficient to account for the mannequin’s broader failures. A single unfamiliar visible function appeared able to disrupting the mannequin’s total interpretation, even in duties it had beforehand carried out properly.
Overview of GPT-4.1’s efficiency earlier than and after fine-tuning throughout commonplace, distorted, and modified-hand clocks, highlighting uneven features and chronic weaknesses.
Conclusion
Whereas the paper’s focus could seem trivial at first look, it doesn’t particularly matter if vision-language fashions ever be taught to learn analog clocks at 100% accuracy. What provides the work weight is its give attention to a deeper recurring query: whether or not saturating fashions with extra (and extra various) knowledge can result in the sort of area understanding people purchase by means of abstraction and generalization; or whether or not the one viable path is to flood the area with sufficient examples to anticipate each seemingly variation at inference.
Both route raises doubts about what present architectures are really able to studying.
First printed Monday, Might 19, 2025
