160 - Jennifer Hu: From Human Minds to Artificial Minds

0:00Welcome back to the Stanford Psychology Podcast. I'm Sue, and I'm so happy to share my conversation with Professor Jennifer Hu. Dr. Hu was a research fellow at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. Starting July 2025, she started her position as an assistant professor of cognitive science at Johns Hopkins University, directing the group for language and intelligence. Her work aims to understand the computational principles that underlie human language and how language and cognition might be achieved by artificial models. In her work to answer these questions, she combines cognitive science and machine learning with the dual goals of understanding the human mind and safely advancing artificial intelligence.

0:42In today's episode, we discuss her research background together with her recent paper Signatures of Human-like Processing in Transformer Forward Passes, which is essentially a paper comparing how large language models and humans process information in real time when solving cognitive tasks. Also, we would like to note that this conversation happened before Jen started her position. Let's keep that in mind. So, without further ado, here is our conversation.

1:35Thank you so much for joining me. I'm really excited to be hosting you today. And I would like to begin by getting to know your research a little bit better. Currently, you are a research fellow at the Harvard University, and you have recently founded the group for language and intelligence. Your work spans cognitive science, AI, and linguistics with the central goal of understanding how language operates in both organic minds and machines. To start off, can you tell us a bit more about how you first got interested in the questions you study today? What initially drew you to the intersection of these seemingly apart but actually very intertwined fields?

2:18Yeah, well first of all, thank you so much for having me. I'm honored to be here. I think looking back now, there were kind of two main ingredients that drew me to this field. So first, growing up, I was really like absolutely enthralled by learning foreign languages. And then second, I got really interested in physics as well. And I wasn't sure how these two passions would ever be combined. And then when I got to college, kind of on a whim, I took my first linguistics class. I'd never really heard of linguistics before, but that was when it all kind of clicked for me. So I realized that I could study this fascinating, complex, and messy human system, but with the precision of a natural scientist trying to describe the way the world works.

3:05And I ended up majoring in mathematics and linguistics. And then thanks to a lot of wonderful mentors, I eventually discovered the field of computational cognitive science, which led me to where I am today. That sounds really interesting, from language and physics to all the way here. Also, historically, you probably know this, but for our audience here, cognitive science has had a great impact on the birth of the AI we know now, all these language models or just like vision models. And with the advancement in both of these fields today, there's a great interest in studying these fields together again.

3:46What has it been like building your research identity across these domains? And also, I assume, keeping this identity in this ever-evolving first-paced space? Yeah, yeah, that's a great question. Yeah, as you said, COGSI and AI have historically been intertwined to varying degrees across the past century. But I think today we're at this super exciting pivotal moment for both fields. So from the COGSI perspective, you know, AI models are kind of the first model organism that we've ever had for studying complex behaviors like language or reasoning. And we can poke and prod these models unless we can't for humans.

4:26And that just opens up so many opportunities for us to use AI to investigate COGSI questions. And then in the opposite direction, I think COGSI is becoming increasingly important for AI. in the sense that the goals of AI are starting to look a lot like the goals of COGSI. So we want to be able to understand the mental constructs of a complicated black box system, but we don't ever have direct access to any of the internal mental states. So the tools and theories that COGSI has developed over the past few decades, I think, are essential for AI researchers to measure the progress of their models.

5:02You also mentioned the fast pace of these fields, which I think is also very salient these days. So you've probably heard the saying that doing a PhD is a marathon, not a sprint. And I feel like doing a PhD in AI is like you're sprinting a marathon. Being at the intersection of these two fields is really exciting, but can also sometimes feel like the worst of both worlds. because you're trying to do like careful theory-driven science, but you also have the pressure to keep up with all these latest models and methods. And so it can feel like a really delicate balancing act. But at the end of the day, I think it's just a really cool, exciting feeling to be like in the wild west of a new emerging field and being able to make new discoveries so quickly.

5:53I definitely understand that. and I think it's a really nice combination too when you think about it. As you mentioned, there is this historical treasure called cognitive science, psychology and neuroscience where people have been studying this black box we have in our body for centuries now and with the advancement of recent technology. I can definitely see how exciting that could be but at the same time very tiring as you mentioned. Right. Yeah. now that you have your own group which you will be directing soon how do you see your research direction evolving how does this new group reflect the scientific questions you are most excited about right now yeah so the focus of my new group is to understand the computational and cognitive basis of language in human minds and in machines so some of the questions that we're interested in include how do humans understand language so flexibly and efficiently, even though communication in the real world is full of noise and ambiguity?

7:04How do we reason about other people's goals and beliefs in order to communicate successfully? And how does this all happen in real time, given the limited resources of our minds and brains? And I think in some ways, my scientific questions are pretty constant. It's just that the tools that we have to investigate them are constantly changing and evolving. So for example, I think new tools from the field of mechanistic interpretability are suggesting new ways to use AI models to study human processing, which hopefully we'll talk about more later. And with new language models being optimized for chatting with humans, I think we also have new opportunities to study pragmatic communication in humans and machines.

7:46And there are new data sets like BabyView and these high quality egocentric video recordings from babies. And so we can try training models and more naturalistic data and so on. And so I think with each of these new advances, we can start to push closer and closer to better answers of those core scientific questions, which is super exciting to me. I think one follow-up question I would like to ask again is, you are going to be the director of your own group and probably you will be pushing those questions or methods and how do you think that will differ from being in someone else's lab or directing projects but not like a whole group do you see any differences coming or challenges or benefits yeah definitely I mean yeah there are so many it opens so many new opportunities but also a lot of responsibilities and challenges as well.

8:41So I think, you know, part of what's really important to me as I step into more of a mentorship role is to help my future students and trainees find this balance between pursuing these longer term scientific questions, while also feeling like they're kind of keeping up with the pace of AI and being able to go to NeurIPS and, you know, go to all these conferences and feel like they're part of that while not being sucked into this constant cycle of doing kind of smaller, more incremental projects. And finding that balance is hard personally as a researcher, but I think even harder when you're responsible for helping a brand new researcher find their voice and find their footing in the field.

9:27So these are some of the things that I'm definitely going to be grappling with as I recruit more students and start the lab. That sounds super nice to hear how dedicated you are to mentoring your students and how caring you are with their progress as well. Maybe you should think to your paper. One thread that comes through clearly in your work is a desire to understand the workings, limitations and challenges of human and machine intelligence. Not just from a computational perspective, but from a cognitive and psychological one as well. I would like to dig into one of your recent papers that explores these themes, where you use methods and concepts from cognitive science to study if artificial minds provide good models of human cognition.

10:13I'm talking about signatures of human life processing and transformer forward tests. So far, we often have seen scientists use AI models to predict what humans do, what their final decisions, judgments or behaviors would be. This paper takes a different approach. It looks at how the models arrive at those answers, those outputs, if you want to say, by comparing their internal computational steps to the time course of human responses. It's a move from treating AI as a black box to seeing it as a potential model of human cognitive processes. We see more and more studies shifting to this perspective.

10:56What are your initial thoughts on this idea? Yeah, so I'm really excited about these directions. And of course, this isn't a completely new idea. So people have previously looked at like processing pipelines within AI models and try to compare them to pipelines, whether it's like more cognitive kind of processing pipeline in a human mind, for example, or also like sequential kind of processing in brains as well. And so a lot of that was kind of inspiration or foundation for the work that we did. But I think a lot of the prior work has focused on specific tasks, for example, like object recognition.

11:36and there's still a lot of untapped potential in comparing model and human processing on more complicated tasks like reasoning or complicated kind of like factual recall tasks and looking for higher order patterns. So not just saying, okay, at this time slice, the model should be doing this kind of computation and the next time slice, it should be doing a different kind of computation, but looking for these higher order signatures or summaries of the whole processing pipeline Is the model showing signs of sensitivity to competitor interference effects or dual stage processing? And I think there are still a lot of open questions that I'm really excited to investigate further.

12:20Connecting those ideas again back to your paper. As you mentioned, the paper mainly starts to link the internal working of transformers models to how human cognition unfolds step by step. as you mentioned, how the field is treating these models. And to do so in this state where you don't only look at whether they both get it right or wrong or arrive at the same answer, you track human processing by very unique behavioral measures, which we will hopefully talk about later and a couple of questions later. So the goal here, to my understanding, please correct me if I'm wrong, is to figure out if looking inside these AI models and their processing journey will tell us more about how big things start to struggle while trying to get to that final answer or so to say the final output.

13:10We can start here maybe. The first guiding question that pops to my mind for this paper is how do you even measure the internal processing effort in a model? It doesn't quote-unquote but feel effort as you feel it, right? Right. Yeah, that's such a great question. And honestly, it's one of those questions that keeps me up at night. Like, what does effort even mean for a model? I guess, first of all, I'll say, like, no matter what kind of input a model gets, so whether it's like the world's hardest math problem or the model is just predicting what word comes next in a very simple sentence, like no matter what, the same thing is always happening in the model.

13:56It's like a bunch of matrix multiplications and operations across a fixed number of layers. So in that sense, it's a very deep question of can there be signatures of like deeper or shallower thinking within that one forward pass of a model? And even if the model can't feel effort in the way that a human can, there's still ways to meaningfully say that there are like different amounts of computation and responses to different kinds of input. So there are a few ways to think about this. One way to think about effort from models, which is similar to what we do in the paper, is to kind of think of effort as like meaningful computation.

14:38So, for example, let's say that you ask a model to do a simple math problem like 2 plus 5. And maybe the model is fairly confident in the correct response already by like layer number 5. and then the rest of the layers are just doing something that's like propagating that correct response across the remaining layers in the model. So even though the model is doing the computation across a fixed number of layers, the model is only really doing something interesting in the first few layers. So maybe we can think of that as requiring fairly little effort. And of course, it's not actually that simple, but I think this kind of layer-based view of effort has been the kind of like assumed working definition of effort in a lot of works on like early exiting or saving compute at test time and so on.

15:28I think there's another way to also think about effort, which is perhaps less localized. It's going to more like macro scale. And this is related to this issue of task demands. Let's say I want you to do a simple math problem like two plus five. so I could just test your ability to complete this math problem by writing down the equation 2 plus 5 and I have you write down the answer it's pretty simple but on the other hand I could also test your ability to solve 2 plus 5 by asking you to like balance on one leg and tell me like if Mary has two pedometers and five pomelos how many objects does Mary have so I bet the second one would be a lot harder and maybe you'd get it wrong, maybe you'd take longer to respond, or maybe you'd feel like less sure of your answer, even though the underlying math question is the same.

16:24And we know that humans are sensitive to these kinds of task demands, especially children. And in recent work that I've done with my collaborator, Mike Frank here at Stanford, we showed that language models are also sensitive to these kinds of demands associated with doing tasks. So depending on how you phrase a problem, it can make the task harder or easier for the model, so to speak. And so I think this is another way to think about effort in an AI model at more of like a macro scale. Going back to the first approach you mentioned, the layer-based view on effort, this paper focuses on layer time dynamics within transformers.

17:07Can you explain what that means in simpler terms maybe. Also, what is the importance of the Transformers architecture in absorbing these changes? Because you mentioned that the specific structure of these models are very important to follow through with these effortful actions, if you were to say. How does this layer time dynamic and Transformers come into play for this type of action? yeah so i mean i guess i should say off the bat that we could have done this analysis on like an rnn or a different kind of model like it wasn't super specific to transformer model but just in practice the like most performant ai models in language and vision domains that we were looking at today were based on transformers so that's what we went with and that's where most of the mechanistic interpretability is but the idea of layer time is kind of like like this analog of real time for humans in the sense that you know each layer is getting information from previous layers and so it's kind of unfolding sequentially and so we can start to draw these parallels between layers and actual clock time in humans and think about things like how early does a model get the right answer or how much time quote unquote is the model spending considering an incorrect answer versus a correct answer and so on.

18:35I believe just the whole layer structure is also very important too, to follow the action following two steps by steps, right? Because I'm assuming there are some controversies around having layer-based steps in human cognition, but with these models, because their structures are based on layers, I think, does it make it easier for you to just like what's going on at what step exactly? Yeah, that's a great question. So human processing isn't just a serial, very discrete step-by-step process, of course. I think one thing about our approach that differentiates it from previous approaches is that we're not making these direct parallels of slices of time, so to speak, between humans and models.

19:22In contrast to prior work, for example, showing that in object recognition kind of vision models, earlier stages of processing seem to map on to earlier regions of brains that do visual processing and so on. So that's a very specific commitment. We're not claiming that there's like a direct parallel between layer two in a model versus some brain region or whatever for humans. And so we're looking more at the general shape of the curve, how much time is being spent preferring the intuitive answer or correct answer? Is there a two-stage process? I mean, I think there are pros and cons to that approach, of course.

19:58So like, of course, it'd be great to have a much more direct parallel between models and brains or minds. But also at the same time, we can use our approach to investigate links between model and human processing for like much more complicated tasks where we don't have existing theories of like a very specific kind of pipeline in brains or in minds. And so it kind of opens up a much larger set of questions that we can start to ask, even though we have less direct parallels. That's fascinating that now we are at a stage where we can actually make those connections. Your paper mentions several studies across different cognitive domains to point to these similar processing challenges or behaviors.

20:43Can you highlight a couple of examples? And I know that these studies are titled and they have very different umbrella terms like factual information or exemplary categorization, logical reasoning. Can you touch upon some of these so that we can better understand how you actually draw those parallels? Yeah, definitely. So one example domain that we looked at is capital cities, and specifically capital cities, which are not the biggest or most famous city in the country or state that it belongs to. So for example, the capital of Illinois is Springfield. But when you ask people what the capital of Illinois is, they're often tempted to say Chicago.

21:29Yeah, I think I would say Chicago. Yeah, exactly. Right. And so in this domain, it's really interesting because you get these very salient, competing, intuitive answers that are incorrect. And so what we did in our human experiments is we measured, well, number one, we measured the accuracy of people's responses. Did they get the correct answer or not? How much time do they take to respond? And we also had finer grained measures of their uncertainty while they type out their responses. So we measure like how often do they press the backspace key or like how many keystrokes do they make relative to their final answer or what time do they begin typing their final answer without ever pressing backspace again.

22:19So there are all these like really fine grained measures that we can get to kind of try to find these like indicators of processing load or difficulty that go beyond just accuracy. Another domain we looked at which is completely different is categorizing atypical animals. So an example here is tell me if a whale is a mammal or a fish. And you know even though like most of us know or we learned in middle school or something that whales are mammals, obviously whales have a lot of properties that are characteristic of fish. And the cool thing here is that people's responses in this task were measured using mouse tracking.

22:58So even if people basically always click on the correct answer, so like they'll always click on mammal, in this case you can see like through their mouse tracking trajectories that the way that they move their mouse will often first curve toward the answer fish. And you can take this curve and get a bunch of measures that reflect how much uncertainty or like decision conflict people are facing. How big is this curve? Or what time point did the person's mouse start to accelerate the most toward the correct answer? And you can get these really cool temporal spatially fine-grained measures of decision conflict that's very different from like the typing domain that we looked at or some of the other domains that we tested as well.

23:47How would you connect these decision conflict behaviors in humans to models, for example? What were the things you compared within these challenges? Yeah, so first we kind of looked at whether models would show these signatures of competitor interference effects the way that we would expect for humans. In the cases where you have capital cities, for example, that are not the same as the most famous city, do you see in the model's layers that it first predicts the salient incorrect answer? Like it would first predict Chicago and then eventually predict Springfield for Illinois. Versus if you have a capital city that is the same as the most famous or popular city like Paris and France, you would not expect to see any kind of like competitor interference or like two-stage processing kind of effect.

24:44So that was one kind of measure that we looked at across these domains. And then we have a bunch of other measures that were a bit more exploratory, but that captured different aspects of the model's kind of like decision process, so to speak. Yeah, so we can look at the uncertainty over time. So we can measure like the entropy of the model's distribution at each layer. We can look at the difference in confidence between the correct and correct answers. And we also tried this relatively novel measure of boosting across layers. And that kind of tells you like how much each layer is contributing to boosting the probability of the correct answer versus lowering the probability of the incorrect answer and so on.

25:32So we have a bunch of different measures of this in models. There are consistent findings across diverse tasks and modalities strongly suggesting that these similarities aren't coincidental. As you mentioned, like, for example, these different metrics used in models and different methods you just talked about that you applied to human minds, the similarities between them, you found that these are not happening randomly. The models trained on general purpose objectives like next word prediction, they seem to have inadvertently developed processing strategies that eco-human cognitive processes. What are some implications for both AI and cognitive science now that you found that we can actually track these similar traces of effort in humans and models and they're actually responsive to the same stimuli.

26:32Yeah. Yeah. So from the AI perspective, I think it's important to develop a better understanding of what kinds of things are easy or hard for a model to do. And once we have a way of measuring that, I think there are a lot of potential implications and applications of that. So for example, we can start to design better evaluations of abilities. And this is very near and dear to my heart as a cognitive scientist who's always thinking about how to actually, you know, design evaluations with better construct validity. So, you know, if our goal is to be adversarial to a model and we want to know whether it can do a certain kind of ability or task, no matter what kind of setting it's in, then maybe we want to make the evaluation hard for the model, like as hard as possible.

27:26And so if we have these measures, then that kind of helps us gauge where we are in terms of how much load it's putting onto a model. And if we're trying to be scientists and we're trying to understand in a very pure sense, is the model capable of this ability at all, then we want to maybe make the task as, I mean, easy is not the right word because it's not about making the task easier, but it's about making the measurement more pure or more direct in a sense. And so I think getting a sense of like how models process inputs and taking that high level inspiration from cognitive science to understand what might be like good measures of processing difficulty in models, I think that can really help with designing evaluations.

28:11And another kind of AI perspective is looking at methods for test time compute efficiency, for example. So if we have a sense of what kind of processing pipeline is going on under the hood in response to like difficult inputs or easy inputs from the model, then we can start to think about, can we predict in which settings a model will require more compute or less compute in a forward pass. And maybe we can create methods that adapt to the inputs in a way that's most efficient. And people are, of course, already working on this, but I think having more human-inspired measures could be a way of getting at this issue.

28:55These findings are very surprising to me. I know that AI models are very developed, and we can use them as like, you know, models of human mind, etc. But like making a decision, struggling with that decision, I think those seem very human-like challenges and seeing the traces of these and models. I think it's like, I don't know, it gets me very excited. When you started this project, were you expecting these results or were you expecting them at this level of confidence? Yeah, that's a great question too. I definitely was, I wasn't fully surprised that there is some sense of, you know, initially being tricked by an incorrect answer, for example, because we know that models are very sensitive to like frequency effects and things in the training data.

29:45And so maybe it's reasonable to expect that they would, you know, in early layers preferring things that are more frequent in the training data or using some other kind of heuristics or biases. Right. But I think what surprised me is how this is generalized across the human modalities so much like going from, you know, a simple like capital cities task, which, you know, is very kind of natural for a language model to do versus predicting like people's mouse movements, which just feels like a completely different kind of modality. that was like super striking to me in the most recent version of the paper we tested 18 different language models and two different vision models you know this wasn't limited to just one model we found this across many different models and there are some interesting interactions between model size and the ability to predict human behaviors I think just like showing that this finding was not just limited to one domain was really cool and really exciting to me And I mean, I think this also speaks to like maybe some of the follow up directions that I'd love to do with this project and getting more at the cognitive science perspective of all these things.

31:02I think that this work so far is kind of like a proof of concept. So we aren't really speaking directly to cognitive theories yet, but I think our results suggest there's some level of alignment between human and machine processing. and it's not just one model it's not just one domain so it's there but why is it there right so that's like the the next question we want to tackle in our next kind of series of experiments um so i think like in future work we can start to ask questions like what properties of a model make it more aligned with humans in this processing patterns so for example if we change the way a model is trained like its data or its training objective or also putting constraints on the architecture itself, right?

31:47I think those are really interesting questions to see if those kinds of manipulations will causally affect the way that a model aligns or predicts human behavior. And ideally, the kind of downstream goal of all of this is to use aspects of model processing and these tools from mechanistic interpretability to make new predictions about human behavior and test those experimentally. I'm looking forward to reading more about them in the future, hopefully. And thank you so much for giving us these insights into your paper. Before we wrap up our conversation, for our audience, especially for students who are listening to us and who are also interested in cognitive science and technology, what advice would you give them for bridging the gap between these two?