What are vision language models considering when attempting to answer a question from a chart image? I am pleased to announce that our recent work exploring this topic has been accepted to #IEEE #VIS. Check out the pre-print below and see you in Vienna: arxiv.org/abs/2504.05445
Those look like sweet vibes though..
I had paper that was rejected from VIS for the same premise (Passing the Data Baton Work - ieeexplore.ieee.org/abstract/doc...). It snuck into to TVCG via VDS and has been reasonably cited. But even when I presented it at VDS the only question I got was "how is this vis?".
I told the students in my class that timelines used to chronological (e.g, not algorithmically curated) and they were surprised. Watching the internet evolve is now wisdoms I impart…
I’m really sorry to hear that. Your work is wonderful and very impactful.
This was a fun paper to write and a fun collaboration.
Very cool!
If we’re writing really long documents that no one wants to write and also no one wants to read, then what is being accomplished with ChatGPT besides procedural performance?
I feel like I have been involuntarily recruited into a reality tv show where they need to manufacture crisis as the basis of a “plot”
We also conducted a preliminary comparison to what humans look at and found interesting alignments with VLMs, particularly those fine-tuned on ChartQA tasks.
Our approach allows you to look at each text token of the input question and examine what the model is `looking' at. To demonstrate this in action, we administer the mini-VLAT & VLAT tests (a standard tool for assessing visualization literacy in humans).
Vision Language Models can jointly reason over images and text, but what is it reasoning about? In this pre-print, we explore the internals of several open-source VLMs to examine what they focus on and what information they prioritize for ChartQA tasks.
🚨 New Pre-print: Probing the Visualization Literacy of Vision Language Models: the Good, the Bad, and the Ugly
arxiv.org/abs/2504.05445
We are seeking individuals to participate in a survey study to understand people's attitudes on collaborative data analysis and AI-assisted data analysis. The survey will take approximately 25 minutes: uwaterloo.ca1.qualtrics.com/jfe/form/SV_.... #datascience #genAI #datawork #collaboration
Saw this in another (lesser impact) situation. It was bots, in high volume. It was bad. It’s probably not meant to boost a specific paper, but more like someone is using conference platforms to experiment with peer review bots.
Somewhere out there a thought leader has now found his chart
An unexpected part of teaching a course on Human-AI interaction is that many students critique papers based upon techniques that could not exist at the time the paper was written. They are quite knowledgeable of the current methods, but, teaching them how we got here has been interesting.
I feel a few of the class discussions with my students this term are “in the before times of AI, here’s how X was done”
I think CHI and CSCW + others are experiencing reviewer fatigue, particularly of senior reviewers. I was an AC and what I saw in the backend of CSCW was concerning (I am AC for CHI too, and it has issues but, but they didn’t seem as bad).
Anecdotal evidence, but, I think also partly reviewer immaturity in the topic (true also at CHI) + long cycles. I withdrew a paper that got an R&R decision because the reviews were so poor we thought it would be a waste of time to resubmit. I’ve had bad reviews before, but this was something else.
I’ve heard this too from a few folks. At what age did your kid start playing?
When I was working with the BC CDC I saw first hand (and for the first time) all the work that dedicated public servants do to keep many critical things running. Politicians are one thing, public servants are another. We can get by without the former but probably not without the latter.
Even though it’s an interesting result, I am somewhat disheartened that research is now “add wait to the prompt and things get better”. I remember when genomics was “turn on the sequencer, get genome, get nature paper”, and while novel, those early genomes had tons of errors in the rush to publish.
If the list is real, it seems like math grants would be flagged but very quickly deemed as not DEI (because they don’t understand it, but they know it’s math), whereas any application with research on females also gets flagged and might be deemed DEI (and so potentially banned).
Pray that it stays that way and that they don't evolve to the point of only contact napping.
This is the closest I got, which was turning everything off:
Science twitter was why I used twitter. I missed loosing it and I am glad it looks like it’s reforming in a new space.
Accurate. You need to be organized enough to have the time to put everything into an (often suboptimal) interface.
That episode is a rare discussion on the rights of AI (especially as we all move toward AGI (allegedly)). To me, it’s immensely important to think about if what we build truly becomes intelligent ( which is also a controversial topic).
I don’t think so, but the reading list is all clickable links to (what I think) are interesting papers.
