The Science of Intelligence

Intelligence / Reasoning

1. The Turing Test gives a definition for whether an agent is at least as intelligent as humans. There are many subtleties to this definition (“average’ human vs. collective of human intelligence, etc). However, how would we know if an AI is strictly more intelligent than humans? How do we know if one AI is more intelligent than another? Relatedly, if we had an answer to the last question, could this be leveraged so that an AI can train an improved (more intelligent) version of itself? For more concrete details talk to mentor! (with Dan Mitropolsky mitropol@mit.edu) 
2. Very recent work has shown that LMs can solve difficult math problems with appropriate prompting. Can we stress test this ability / framework / algorithm? What happens if we apply this framework to unsolved mathematical problems? (with Tommy Poggio tp@csail.mit.edu, and Dan mitropol@mit.edu) 
3. LMs may be good at modeling language, but can fail at very simple tasks (e.g. outputting sentences with the same number of words in each sentence). Can we systematically evaluate LMs on such tasks and come up with a dataset / challenge set for evaluating going forward? (with Liu Ziyin ziyinl@mit.edu and Dan mitropol@mit.edu)
4. Reasoning in Diffusion versus Autoregressive Generation: When a model generates text left-to-right, it reasons under a strict causal schedule: each token is a commitment that constrains all future steps. In contrast, models that can revise or place tokens “anywhere and everywhere at once” (e.g. diffusion-style decoders) reason under a global schedule: partial hypotheses can be updated in parallel to satisfy long-range constraints. This project asks how these output schedules shape reasoning behavior, sample efficiency, and generalization. (With Brian Cheung cheungb@mit.edu)
5. Recent advances in reasoning-focused training have significantly improved the accuracy of large language models (LLMs). However, despite these gains, LLMs remain limited in their ability to reliably communicate uncertainty. They are often overconfident and prone to making guesses when unsure. In formal terms, they are poorly calibrated. This limitation is particularly concerning in high-stakes domains such as healthcare and law, where confidently delivered but incorrect outputs can have harmful or even dangerous consequences. This project will investigate methods to evaluate, analyze, and enhance the confidence estimation of LLMs. The ultimate goal of this work is to both deepen our understanding of how LLMs express confidence and to develop techniques that improve the accuracy of this expression. (with Mehul Damani mehul42@mit.edu) 
   

Linguistics

1. Exploring algorithmic accounts of noisy-channel language processing. Past work in psycholinguistics suggests that human language processing for atypical utterances is consistent with rational, Bayesian inference (taking into account a prior over intended messages as well as the likelihood of different kinds of errors). Yet questions remain about the algorithmic-level account for this type of “noisy-channel inference” — how do comprehenders search the large space of possible alternatives, and how do linear order and constraints on cognitive resources affect the types of inferences that they make? The tools of probabilistic inference allow us to build implemented models of noisy-channel processing that combine large language models for next word prediction with symbolic error models and customizable inference algorithms, allowing us to explore this space and compare model predictions to human behavior. Future work can explore modeling choices, or apply this model to new datasets, to provide insights about the algorithms humans employ for robust language comprehension. (with Thomas Clark thclark@mit.edu) 

CogSci paper: https://escholarship.org/uc/item/9kr1b1gm 

CogSci presentation: https://us06web.zoom.us/rec/play/6TARhN2TvyYb1KfhwzdN9BZhJ3v9fKBqsQjygN5kj-bWZ1wt9jSurzapPlsFlv8ICXWEymLbMCaxq9bJ.atKg7z9anLCPDgTD 

GitHub: https://github.com/thomashikaru/noisy_channel_model 

Thomas’ tutorial for different class

2. Modeling the effects of distinctiveness and confusability in memory for linguistic content

Past work in memorability has explored the features that make various kinds of stimuli (such as photographs, faces, sounds, or linguistic items) more or less memorable. In addition to the intrinsic properties of a stimulus, there are context effects on memorability (e.g. the presence of similar distractor items will drive up memory errors). One theory of memory is that stimuli are stored in memory as lossy representations within some representation space, and that making a familiarity judgment consists of doing a type of probabilistic inference, conditional on a query and the set of stored noisy representations, of whether the query has previously been seen. We have collected a large dataset of familiarity judgments for short sentences in a memory experiment, and a possible project would involve modeling the trial-by-trial responses, in order to predict when sentences are correctly recognized, false rejected, or falsely reported as familiar. (with Thomas Clark thclark@mit.edu)

3. Language is to a large extent predictable and systematic — we can figure out the meaning of a phrase if we know the meanings of each word in the phrase. Idioms, however, do not have predictable meanings and yet are commonly found in language. This project focuses on modeling the life cycle of idioms using data from historical text corpora, looking specifically at the role of word and phrase frequency. When an idiom first arises, what is the relationship between the relative frequencies of the individual words, the extent to which we independently expect those words to occur together, and the idiomatic meaning of the phrase? And as these relationships change over time, do we observe idioms dying out, becoming more established, and/or shifting their meanings? (With Michaela Socolof msocolof@mit.edu)

4. It is often said that current LLMs have “solved” language (specifically in terms of syntax). But LLMs have not been systematically stress-rested in their knowledge of language; one can find simple counterexamples to this (talk to mentors for details)! Can we develop a robust programme of linguistic competence challenges to test LM’s knowledge of English (or other language?) The idea is that this should test generalization to novel scenarios where we can see if the model accurately applies phonological / morphological / syntactic rules of the language (with Dan Mitropolsky mitropol@mit.edu)

Formal languages / logic

1. In ongoing work with Prof. Poggio and visiting student Laura Ying Schulz, we have been researching how transformers learn Context-Free-Grammars, both as a small, tractable proxy for how LMs learn language, and independently for its theoretical interest (what is the “learning landscape” like for learning CFGs?). Many open problems in this new area – a) comparing training on languages of different “complexity levels”, b) grammar induction (given a LM, can one “deduce” the underlying grammar)?  (with Dan Mitropolsky mitropol@mit.edu)

Representations / multimodality

8. As AI systems gain capability, do their internal representations become less human-like—and can that divergence be advantageous? This project investigates beneficial misalignment: the hypothesis that increasing competence often pushes models toward representations that depart from human or biological patterns yet improve task outcomes. Rather than asking whether machines should align to humans, we test the stronger claim that more intelligence ≠ more human-like intelligence, and that less alignment can sometimes be instrumentally better. (With Brian Cheung cheungb@mit.edu)
9. On the platonic ideal hypothesis, specifically vision and language alignment. Recently, a body of work has increasingly shown evidence for the “platonic hypothesis”– that ML models trained in various domains (eg vision and language) have similar “alignment” between similar objects (e.g. an apple and an orange are aligned *both* in vision models and in language). First, we can investigate the rigor of this claim, including any interesting counterexamples. Then, for cases with high alignment or a lack thereof, a mechanistic *theory* of why this occurs is lacking on the language side. For example, can we say something about the linguistic distribution of pairs of objects to predict alignment? (with Dan mitropol@mit.edu, Brian Cheung cheungb@mit.edu)
   

Neuroscience 

1. Epigenetics of dementia: ‘Do you remember me?’ This is a question dementia patients often get asked and the weight of the silence that follows continues to burden the society, especially with a rapidly ageing population. The interplay of environment and genetics (i.e. epigenetics) is crucial in the pathogenesis of neurodegenerative diseases. With the disease onset of Alzheimer’s disease (AD), Vascular dementia (VaD), mixed dementia, Post-stroke dementia (PSD) and Post-stroke no dementia (PSND) being primarily sporadic in nature coupled with DNA methylation changes observed in their pathogenesis, a pivotal question is raised. Can DNA methylation changes be used as a tool to differentiate between the different dementia types? With DNA methylation sequencing data collected from dementia patients’ brain tissues (n=48), the goal is to mine the big data in hand to identify similarities and differences in DNA methylation patterns. One could potentially use the DNA methylation patterns as a litmus test to delineate the types of dementia. (Contact Sharmelee Selvaraji sharms09@mit.edu)
2. Brain TreeBank data exploration: the InfoLab has collected 40 hours of intracranial neural activity, recorded while subjects watched movies. For each movie, we have the dialogue and various visual features aligned to the neural recordings. This project will focus on annotating a new feature in the movie, e.g., facial expression, and then training a decoder to localize processing of this feature. There are also annotated features, for which no decoding has yet been run, e.g., labeled scene decoding. See the BrainBERT paper for an example of decoding with this data (Contact Christopher Wang czw@mit.edu)
3. Models of Hippocampus – there is a rich, recent body of work on theoretical models of hippocampus (in particular see work by Ila Fiete’s lab– someone in her lab could make a good mentor or co-mentor!) However, the role of recurrence in hippocampus is unknown, even though hippocampus is highly recurrent. Survey the field and investigate / suggest possibilities. Maybe the assembly model (NEMO) could be relevant. (Contact Tommy Poggio and Dan Mitropolsky, we may help find further contacts)
   
   Motor / Robotics
4. How do we walk without thinking – And which sensory inputs really matter? In humans, walking begins as a highly sensory-driven process, where toddlers rely heavily on vision, touch, and balance cues, but gradually transitions to an automatic skill supported by internal models (Malone et al., 2025). Similarly, work in robotics has shown that some sensory signals can be removed without significant loss of performance, while others are indispensable (Yu et al., 2023). Building on these insights, we will use MuJoCo simulations and deep reinforcement learning to train humanoid and quadruped agents to walk, systematically analyzing and selectively removing sensory inputs to determine which are most critical, and when. By doing so, we will provide insights into how humans gradually reduce their reliance on sensory input during motor learning, and in robotics, by guiding the design of more efficient and robust controllers that can walk or run with less dependence on sensory feedback. (Contact Naomi Chaix-Eichel naomichx@mit.edu)
   
   

Possible directions / idea “stubs” below

The idea of LM inversion (i.e. “context prediction”), is this the same thing as a model knowing semantics? What is the complexity of inversion? 

Tomer’s work and vision augmentations… a metric for augmentation comparison, and automatic augmentation 

HRM experiments, “apple benchmark” 

Can a custom token improve attention in a transformer?

Attention provides state vector. Keep it separate? Transformer and HRMs.