A language model internally reasons about a task before outputting an answer.
|
AI Alignment and Interpretability
My AI research focuses on understanding and evaluating the internal reasoning processes of large language models. I study cases where models perform hidden computation using behavioral experiments and causal interventions to investigate what information is represented inside the model.
More broadly, I am interested in making AI models safe and trustworthy. As models become more capable and are deployed in more complex settings, it is increasingly important to understand not only what answers they produce, but how those answers arise. I work on developing methods that can reveal when AI systems are reasoning reliably, when they are relying on shortcuts or hidden computation, and how we can build monitoring tools that remain useful even when surface-level explanations are incomplete or misleading. |
A Milky-Way-mass dark-matter halo from the Caterpillar simulation suite and its subhalos (circled)
|
Galaxy Formation and Nuclear Astrophysics
My astrophysics research explored how galaxies form and evolve across cosmic time, from the first galaxies in the early Universe to the Milky Way as it exists today. A central goal of my work was to understand how the chemical elements are produced and dispersed, especially the heaviest elements created in rare, energetic events (e.g., rapid neutron-capture enrichment). This research requires high-performance computing and large-scale cosmological simulations.
I led the development of ultra–high-resolution, star-by-star cosmological hydrodynamic simulations of the first galaxies, designed to model feedback from individual stars and to track detailed elemental abundances self-consistently. |
STELLAR HALO MODELING: Unlocking the History of the Galaxy
Anatomy of the Milky Way: Bulge, thin disk, thick disk, stellar halo
The motions and chemical composition of the stars currently present in the extended outskirts (called the “stellar halo”) of the Milky Way preserve a record of the Galaxy’s formation history. Large galaxies form by "eating" (aka, merging with) many smaller galaxies to grow larger. While most of the stars in the center and disk of a galaxy are formed in situ, the stars in the stellar halo primarily originated from the many small galaxies that the central host galaxy accreted over billions of years. Astronomers have recognized the incredible potential of the information stored in the halo stars and are in the process of collecting revolutionary amounts of data on the Milky Way stellar halo. Large dedicated satellite missions such as Gaia (which aims to make a precise 5D map of the stars in our Galaxy) and spectroscopic surveys such as APOGEE are producing positions, 3D velocities, and element abundances for up to a billion Galactic stars.
Galactic archaeologists like myself work to interpret all of this data to learn the history of our Galaxy. By combining large observational datasets with cutting-edge galaxy simulations, we characterize the merger history of the Milky Way.
US Department of Energy | Computational Science Graduate Fellowship Program Review 2022 | Washington DC, USA
UNDERSTANDING ALL THE LITTLE STUFF
Stars forming in simulations of early dwarf galaxies (Brauer 2023)
I specialize in the smallest, earliest galaxies that merged into the Milky Way.
Stars from the smallest galaxies (“ultra-faint dwarf galaxies") are relics from the era of the first stars and galaxies, preserving clean signatures of early chemical enrichment. They are also poorly understood ingredients in the formation history of the Milky Way. Currently, we are obtaining a massive amount of chemical abundance data of stars in ultra-faint dwarfs in the Local Group and of metal-poor stars in the Milky Way’s stellar halo that partially originated in ultra-faint dwarfs. The problem is that chemical abundance observations are outpacing theoretical models. We have a wealth of data, but we cannot leverage the full detail within it. Current state-of-the-art chemical evolution models cannot model the full abundance distributions because they assume instantaneous formation of clusters of stars with homogeneous mixing of yields. New, more detailed models are necessary to fully utilize the data to explore complex galaxy formation processes including metal mixing, hierarchical galaxy merging, bursty star formation, and variations across different galaxies.
To study precisely how stars formed in the earliest galaxies, I am producing the first cosmological simulation of dwarf galaxies with individual stars, detailed chemical yields, and highly-resolved metal mixing. This will allow star-by-star comparisons with high-resolution stellar abundance observations. Such a simulation has never been run before because of the difficulty of simultaneously resolving galaxies and individual stars and gas, but for the first time, advances in high performance computing and massive amounts of detailed abundance data make this work both feasible and necessary.
International Astronomical Union Symposium 377 | Kuala Lumpur, Malaysia