WEBVTT 00:00:00.446 --> 00:00:03.081 [MUSIC] 00:00:03.081 --> 00:00:07.160 GRETCHEN HUIZINGA: Welcome to Abstracts,  a Microsoft Research Podcast that puts the   00:00:07.160 --> 00:00:16.120 spotlight on world-class research in brief.  I’m Dr. Gretchen Huizinga. In this series,   00:00:16.120 --> 00:00:20.080 members of the research community at  Microsoft give us a quick snapshot—or   00:00:20.080 --> 00:00:25.987 a podcast abstract—of their  new and noteworthy papers. 00:00:26.000 --> 00:00:31.520 My guest today is Dr. Michel Galley, a senior  principal researcher at Microsoft Research.   00:00:31.520 --> 00:00:37.240 Dr. Galley is the coauthor of a paper called  “MathVista: Evaluating Mathematical Reasoning of   00:00:37.240 --> 00:00:42.945 Foundation Models in Visual Contexts.” Michel,  thanks for joining us on Abstracts today! 00:00:42.945 --> 00:00:44.040 MICHEL GALLEY: Thank you for having me. 00:00:44.040 --> 00:00:47.560 HUIZINGA: So I like to start with a  distillation or sort of an elevator   00:00:47.560 --> 00:00:52.040 pitch of your research. Tell us in  just a couple sentences what problem   00:00:52.040 --> 00:00:56.140 or issue your paper addresses  and why we should care about it. 00:00:56.140 --> 00:00:59.920 GALLEY: So this paper is about  evaluating large foundation models.   00:01:00.560 --> 00:01:06.240 So it's a very important part of researching  large language models because it's a good way   00:01:06.240 --> 00:01:10.920 to evaluate, kind of, the capabilities—what  these models are good at and not good at. And   00:01:10.920 --> 00:01:16.840 a part of the focus of MathVista is to evaluate  these large foundation models in a multimodal   00:01:16.840 --> 00:01:22.760 setup, so when the input to the model is actually  not just text but also text and images. And then,   00:01:22.760 --> 00:01:26.760 an example of a task that such a model  would perform is, like, the input is   00:01:26.760 --> 00:01:31.560 maybe a mathematical question, and then there's  some visual support to that question, let's say,   00:01:31.560 --> 00:01:37.560 of an image of a graph, and then the model has to  respond to something related to that. And why this   00:01:37.560 --> 00:01:42.440 is important … there has been a lot of work, of  course, on large foundation model. Especially when   00:01:42.440 --> 00:01:46.940 it comes to reasoning tasks, like mathematical  reasoning, a lot has focused more on written form. 00:01:46.940 --> 00:01:47.419 HUIZINGA: Yeah … 00:01:47.419 --> 00:01:49.080 GALLEY: So MathVista is one  of the very first datasets   00:01:49.080 --> 00:01:52.600 that has input that is both images and text. 00:01:52.600 --> 00:01:56.760 HUIZINGA: Yeah, yeah. Well, reading your  paper, it seems like this is an area that   00:01:56.760 --> 00:02:00.320 hasn't been studied systematically.  In fact, you actually say that!   00:02:00.320 --> 00:02:06.800 And say that the field is largely unexplored. But  quickly tell us what has been done in this field,   00:02:06.800 --> 00:02:11.060 and then tell us how your research addresses  the proverbial gap in the literature. 00:02:11.060 --> 00:02:15.920 GALLEY: Well, there has been a lot of work  on vision and language in other problems,   00:02:15.920 --> 00:02:21.760 like not just about reasoning. Maybe let me just  mention why reasoning is important. So one reason   00:02:21.760 --> 00:02:25.840 I think it's very interesting to evaluate these  large language models in terms of reasoning skill   00:02:25.840 --> 00:02:31.760 is that we evaluate their capabilities beyond  just memorization. So as many of your listeners   00:02:31.760 --> 00:02:36.400 probably know, these large foundation models  are trained on large amounts of text that is   00:02:36.400 --> 00:02:41.360 public data from various sources. So when you  ask a question to a large foundation model,   00:02:41.360 --> 00:02:45.898 it could be the case, in many cases, that it  just memorizes things it has seen in the data. 00:02:45.898 --> 00:02:49.480 GALLEY: So what makes it interesting in  terms of reasoning, the answer oftentimes   00:02:49.480 --> 00:02:52.920 is not there in the data. So it needs to  develop this ability to connect the dots   00:02:52.920 --> 00:02:57.800 between various pieces of information  to come up with a new answer. So the   00:02:57.800 --> 00:03:01.960 focus of our paper is really on mathematical  reasoning, but it goes also a bit beyond that   00:03:01.960 --> 00:03:05.960 because what is also represented in the  data is also science question and so on. 00:03:05.960 --> 00:03:06.672 HUIZINGA: Yeah … 00:03:06.672 --> 00:03:12.894 GALLEY: So this reasoning part has largely  focused, until MathVista, on text-only modalities. 00:03:12.912 --> 00:03:17.280 GALLEY: So it's one of our very first ones  that combines text and images in terms of   00:03:17.280 --> 00:03:21.880 evaluating these large foundation models.  So you ask about what was done before. So,   00:03:21.880 --> 00:03:26.560 yes, there has been a lot of work, text only,  on reasoning, for example, the mathematical   00:03:26.560 --> 00:03:31.160 question that's just based on text. And there  has been a different stream of work that was   00:03:31.160 --> 00:03:36.050 much more focused on vision. A lot of work has  been on tasks such as visual question answering … 00:03:36.050 --> 00:03:36.068 HUIZINGA: Yeah … 00:03:36.068 --> 00:03:39.760 GALLEY: … where basically, you have an image  and the question is about answer a question   00:03:39.760 --> 00:03:44.853 about this image. So, yes, we’re trying  to fuse the two lines of research here. 00:03:44.853 --> 00:03:44.872 HUIZINGA: Right … 00:03:44.872 --> 00:03:46.880 GALLEY: And that's one of the  first works that does that. 00:03:46.880 --> 00:03:50.560 HUIZINGA: Yeah. Well, let's talk about  your methodology for a minute. Tell   00:03:50.560 --> 00:03:54.320 us how you went about conducting this  research, and what methods did you use? 00:03:54.320 --> 00:03:59.080 GALLEY: Yes, sure. So that's a bit different  from a typical, kind of, machine learning   00:03:59.080 --> 00:04:03.000 paper because the focus on this work is really on  benchmarking on the dataset. So the methodology is   00:04:03.000 --> 00:04:08.560 more about how we collect the data, process it.  So they have two components to doing that. One   00:04:08.560 --> 00:04:13.600 was to look at existing data that already combines  vision and text. And there are existing datasets   00:04:13.600 --> 00:04:19.320 that are actually already fairly big but that were  not focused on reasoning. So we use those existing   00:04:19.320 --> 00:04:25.320 datasets and look for instances in the data that  actually include some mathematical or science   00:04:25.320 --> 00:04:29.920 reasoning. And so that part is leveraging existing  datasets, but the important part is, like,   00:04:29.920 --> 00:04:34.760 we really want to carve out what was interesting  piece in terms of reasoning. And we had different   00:04:34.760 --> 00:04:41.040 stages of processing the data to identify the  subset that was reasoning-based. So one first   00:04:41.040 --> 00:04:48.120 step was basically to apply some automatic filter  to determine whether or not a given example, let's   00:04:48.120 --> 00:04:52.360 say something that is visual and text, is actually  … involves some mathematical reasoning. So we have   00:04:52.360 --> 00:04:56.960 different strategy. For example, if the answer is  numerical, it's likely that it might be something   00:04:56.960 --> 00:05:02.000 mathematically related. But that's just the first  stage. And the second stage, we actually had   00:05:02.000 --> 00:05:07.480 humans, annotators, just certify that the selected  data is actually of high quality. So we do have an   00:05:07.480 --> 00:05:12.800 example of, “Oh, this is mathematical, and that's  either mathematical or scientific,” and so on.   00:05:12.800 --> 00:05:18.280 And that's one part of the effort. The other part  is that we realized while we collected the data,   00:05:18.280 --> 00:05:22.400 there are certain types of mathematical reasoning  or related to mathematical reasoning that were   00:05:22.400 --> 00:05:28.040 not represented in the data. So we created three  new datasets as part of MathVista. So when I said   00:05:28.040 --> 00:05:33.040 dataset, it's more like, think of MathVista as  like an aggregate of different types of data, and   00:05:33.040 --> 00:05:39.880 we added three of them, three new types of data.  One is what you call PaperQA, which is basically   00:05:39.880 --> 00:05:46.040 data that is collected from scientific papers on  arXiv, and that had questions asking about that   00:05:46.040 --> 00:05:52.734 paper and that included some visual components  from the paper, typically a plot or a figure. 00:05:52.734 --> 00:05:52.752 HUIZINGA: Yeah … 00:05:52.752 --> 00:05:55.120 GALLEY: And then we had IQTest,  which is basically, I mean,   00:05:55.120 --> 00:05:59.240 it's vaguely related mathematically,  but basically it also, kind of,   00:06:00.080 --> 00:06:07.040 tried to see maybe more abstractive thinking about  maybe some input that is both text and visual. And   00:06:07.040 --> 00:06:12.340 the final is about FunctionQA, that is basically  algebraic reasoning and function plots and so on. 00:06:12.340 --> 00:06:12.952 HUIZINGA: OK … 00:06:12.952 --> 00:06:17.760 GALLEY: The important part was actually  to identify among vast amounts of data   00:06:17.760 --> 00:06:20.894 what is actually very interesting  in terms of mathematical reasoning. 00:06:20.894 --> 00:06:20.912 HUIZINGA: Yeah … 00:06:20.912 --> 00:06:23.640 GALLEY: So that part, I think,  was quite a big part of doing   00:06:23.640 --> 00:06:27.180 that work—finding existing data  but also creating new data. 00:06:27.180 --> 00:06:30.760 HUIZINGA: Yeah, yeah. Well, my favorite  part of a research paper is where it says,   00:06:30.760 --> 00:06:36.080 “and what we found was … ,” so talk a little  bit about your results. What did you find? 00:06:36.080 --> 00:06:44.280 GALLEY: So we evaluated a wide variety of models,  including GPT-4, Claude 2, GPT-4V, multimodal   00:06:44.280 --> 00:06:50.840 Bard, and LLaVA, and we categorized them into  three categories. So one is text only. So,   00:06:50.840 --> 00:06:57.120 basically, you take a model that is by default  just text, and we give it the text part of the   00:06:57.120 --> 00:07:01.870 question and ask it to answer the question.  Of course, that's, kind of, a bit of a, it’s   00:07:01.870 --> 00:07:05.640 a difficult task because oftentimes [LAUGHTER]  we crucially build these questions so that you   00:07:05.640 --> 00:07:10.440 have to rely on the vision part. But that's for,  you know, scientific investigation to know how   00:07:10.440 --> 00:07:15.720 well they can do, and so that's one category of  model. A different category is still text only   00:07:15.720 --> 00:07:24.360 but that is given the detection from the image. So  on the image, we do OCR. So we convert those words   00:07:24.360 --> 00:07:30.400 from images to text. It’s kind of an extension of  the text-based model, except that what was images   00:07:30.400 --> 00:07:35.280 is translated into text, and then the input to  the model is word only, and that's a different   00:07:35.280 --> 00:07:40.080 category of model. And the third one is basically  truly multimodal model. And what we found, I mean,   00:07:40.080 --> 00:07:44.360 not surprisingly, it’s, kind of, the one that was  doing most poorly is the one that is text only.   00:07:44.360 --> 00:07:50.760 The second is text plus OCR. And then finally,  the one that does best is the multimodal like   00:07:50.760 --> 00:07:56.120 GPT-4V. But while the ordering between these three  categories makes sense, it was a bit surprising   00:07:56.120 --> 00:08:04.320 that maybe the gap between multimodal and text  plus OCR was not bigger. Well, it’s big, but maybe   00:08:04.320 --> 00:08:09.920 not as big as we were expecting. So, for example,  the best detection from the images model achieved   00:08:09.920 --> 00:08:16.160 like 35 percent accuracy while GPT-4V was 50  percent. So it's a substantial gap but not huge. 00:08:16.160 --> 00:08:20.640 HUIZINGA: Right. Just to clarify, you're  saying OCR. What does that stand for? 00:08:20.640 --> 00:08:22.815 GALLEY: [Optical] character recognition. 00:08:22.815 --> 00:08:22.832 HUIZINGA: Gotcha. 00:08:22.832 --> 00:08:28.080 GALLEY: So, basically, it's the task of taking  text, sometimes typed, but sometimes written,   00:08:28.080 --> 00:08:33.680 and convert this into the actual text  like you would have in a text file. 00:08:33.680 --> 00:08:38.000 HUIZINGA: Right. Michel, does any of this  have to do with the difficulty of the   00:08:38.000 --> 00:08:42.680 math problems that you present these  models with? I mean, it seems to me,   00:08:42.680 --> 00:08:46.320 similar to humans, that the easier  the problem, the easier it would be   00:08:46.320 --> 00:08:51.920 for the machine. So at what level of  math are we talking for these tests? 00:08:51.920 --> 00:08:55.760 GALLEY: What's nice about MathVista is there's  continuum [of] different difficulties. So the   00:08:55.760 --> 00:09:01.720 spectrum is quite broad, going from elementary  school to more advanced concepts such as   00:09:01.720 --> 00:09:06.800 calculus. So it's quite broad. So in the paper,  we do have this, kind of, broken down by level.   00:09:06.800 --> 00:09:11.260 So the number I gave you, like 50 percent, is  an aggregate over all the difficulties. But … 00:09:11.260 --> 00:09:11.712 HUIZINGA: Gotcha. 00:09:11.712 --> 00:09:14.160 GALLEY: But the goal there was really,  kind of, to compare different models,   00:09:14.160 --> 00:09:17.640 but we do have a fair amount of  analysis in the appendix. Actually,   00:09:17.640 --> 00:09:23.510 we have 100 pages of appendices of plenty of  analysis and so on. So if people, I mean … 00:09:23.510 --> 00:09:26.640 HUIZINGA: I saw that. I saw the  length of the paper, and I'm going,   00:09:26.640 --> 00:09:33.880 what? [LAUGHS] That’s a LONG paper! Well, research  in the lab is one thing, I always like to say,   00:09:33.880 --> 00:09:38.160 but understanding real-world impact  is important, too. So where's this   00:09:38.160 --> 00:09:42.460 work going to make the most difference,  and who does it help most at this point? 00:09:42.460 --> 00:09:46.440 GALLEY: Well, I think perhaps that's the  main point of this kind of line of work   00:09:46.440 --> 00:09:51.720 in terms of reasoning is that when looking at  this difficult problem that are mathematical,   00:09:51.720 --> 00:09:55.840 actually it's a way to, kind of, abstract  away maybe more complex capabilities,   00:09:55.840 --> 00:10:01.200 and I think while thinking just about  mathematics might seem a bit narrow,   00:10:01.200 --> 00:10:08.480 I don't think that really is. It's more about  seeing whether this model has the ability to do,   00:10:08.480 --> 00:10:12.880 kind of, multistep kind of processing  of your input and think maybe somewhat   00:10:12.880 --> 00:10:19.360 intelligently about a given problem. So we  focus mostly on math. There is some science,   00:10:19.360 --> 00:10:22.600 but we would be very interested, especially  in future work, to, kind of, go beyond that. 00:10:22.600 --> 00:10:24.560 HUIZINGA: OK, well, let me press in a little   00:10:24.560 --> 00:10:29.320 bit there because … just say I'm a  regular person using a GPT model.   00:10:30.040 --> 00:10:35.880 Is your work more addressed upstream from that to  the research community to say, how do we get these   00:10:35.880 --> 00:10:42.440 models to be better so that downstream people  like me can be more confident of the models? 00:10:42.440 --> 00:10:45.040 GALLEY: Yes, I would say at the moment, I mean,   00:10:45.040 --> 00:10:49.120 this line of work is perhaps more geared  towards somewhat more research community,   00:10:49.120 --> 00:10:57.360 but I think it could be some seed for researchers  to think about some applications perhaps that   00:10:57.360 --> 00:11:01.900 also requires some kind of step-by-step  reasoning but perhaps not going beyond math. 00:11:01.900 --> 00:11:05.920 HUIZINGA: Yeah. Michel, if there was  one thing you wanted our listeners to   00:11:05.920 --> 00:11:10.000 take away from this research, kind  of golden nugget, what would it be? 00:11:10.000 --> 00:11:16.000 GALLEY: Well, I would say it’s the challenging  part of these datasets. I think that's what   00:11:16.000 --> 00:11:20.960 makes MathVista stand out compared to other  datasets. By now, there are a few other vision   00:11:20.960 --> 00:11:26.080 and language datasets, and of course, many  that are more text-based. And we've seen,   00:11:26.080 --> 00:11:31.240 for example, some recent papers showing  that actually MathVista remains one of   00:11:31.240 --> 00:11:34.920 the most challenging ones. So I think  it's probably going to stay around for   00:11:34.920 --> 00:11:39.800 a while because of the difficulty it  represents. So it's open source of   00:11:39.800 --> 00:11:44.580 available datasets that everybody can use,  and I very much encourage people to use it. 00:11:44.580 --> 00:11:47.000 HUIZINGA: Is it on GitHub? 00:11:47.000 --> 00:11:48.280 GALLEY: Yes, it's on GitHub. 00:11:48.280 --> 00:11:53.360 HUIZINGA: So what's next on the  research agenda for helping LLMs   00:11:53.360 --> 00:11:57.680 get better at math, Michel? What are the  big challenges in the field yet? I mean,   00:11:57.680 --> 00:12:03.280 you've alluded to many of them already, sort  of, but what's next on your research agenda? 00:12:03.280 --> 00:12:10.520 GALLEY: Well, I would say what we found so far  is these models are very good at processing the   00:12:10.520 --> 00:12:16.040 textual part of problems it's given, to the model,  but you have the equivalent in images actually   00:12:16.040 --> 00:12:21.520 harder somehow. So I think a lot more work needs  to be done in terms of vision capabilities,   00:12:21.520 --> 00:12:26.800 in terms of reasoning over images, because the  capabilities you will see in text are actually   00:12:26.800 --> 00:12:31.800 quite advanced, whereas the equivalent in images  doesn't seem that good. I mean, a fair disclaimer:   00:12:31.800 --> 00:12:35.200 my background is more on the text side, [LAUGHTER]  so some of my colleagues on the paper are more   00:12:35.200 --> 00:12:39.600 on the vision side, so maybe if a listener maybe  run into some of our coauthors at the conference,   00:12:39.600 --> 00:12:45.150 they might want to talk to these vision people  because that's less of my background. [LAUGHS] 00:12:45.150 --> 00:12:47.360 HUIZINGA: Well, and if you think  about Venn diagrams, you know,   00:12:47.360 --> 00:12:50.240 you've got people that are doing  text, people that are doing vision,   00:12:50.240 --> 00:12:55.140 and then the people that are trying to  do both to see how the worlds collide. 00:12:55.140 --> 00:12:55.920 [MUSIC] 00:12:55.920 --> 00:12:59.520 Well, Michel Galley, thanks for  joining us today. And to our listeners,   00:12:59.520 --> 00:13:06.600 thanks for tuning in. If you want to read this  paper, you can find a link at aka.ms/abstracts,   00:13:06.600 --> 00:13:11.200 or you can find it on arXiv. You can also read it  on the website for the International Conference   00:13:11.200 --> 00:13:17.240 on Learning Representations, or ICLR. And if you  happen to be at the ICLR conference this week,   00:13:17.240 --> 00:13:29.280 you can hear more about it there.  See you next time on Abstracts! 00:13:29.280 --> 00:13:36.111 [MUSIC FADES]