Ankit Biswas

I’m a student at Harvard College working at the intersection of astrophysics and machine learning. My research in astrophysics involves applying computational techniques to discover elusive objects both small and very large (relatively), from detecting the coldest brown dwarfs (~failed stars) with novel ML models, to searching for the earliest stars using Bayesian optimization, to reconstructing the large-scale structure of the early universe with GANs trained on mock simulations of the universe. Simultaneously, I do work in mechanistic interpretability, originally as a means to understand the astrophysical ML models I built (and why they perform better than previous tools/probes) but now more so to analyze the internals of large language models and how they interpret our sensory world. My previous work in this has involved KANs for object detection, neuron attribution in flow matching, dimensional redundancy in text encoders, and the representation of perceptual spaces in LLMs.

Formally, I am part of the LAPIS Survey, working with Dr. Daniel Eisenstein and PI Dr. Sandro Tachella, where I contribute to both technical and science working groups on creating the first census of galaxies in the first 200 Myr of the universe and vastly expanding the current distance frontier. Simultaneously, I work in identifying internal representations of time within LLMs with Marc Kaufmann and Justin Shenk as a SPAR Research Fellow.

I have previously worked with Dr. Patrick Treuthardt, Dr. Ian Hewitt, and Alp Tezbasaran at the North Carolina Museum of Natural Sciences Astronomy & Astrophysics Lab on developing methods for the extraction of galaxy substructure and the detection of intermediate-mass black holes. I have also collaborated with Dr. Laura Dominé, Dr. Richard Cloete, and Dr. Wesley Andres Watters as part of the Galileo Project research team at the Harvard & Smithsonian Center for Astrophysic as a research assistants, where I assisted in the comissioning and calibration of one of the primary telescopes and helped develop computer-vision foundation models to detect anomalous aerial transients. Alongside Dr. Zheng Cai at the Tsinghua University High-z Team, I’ve also been the co-Principal Investigator on observing-time proposals for extragalactic targets at Keck, the VLT, and JWST.

Outside research, I enjoy youth education advocacy and have worked in various science communication initiatives. Currently, I run the Youth Research Initiative, where we mentor middle-to-high school kids around the world in learning the scientific method and completing research projects. My work has been featured in Bloomberg, NASA, and various other venues.

I’m eternally grateful to all the mentors and friends that have helped me so far on my journey—feel free to reach out at my email if you have something interesting!

recent news

technical experience

projects

technical blog 2026

Can LLMs taste (literally)? 

Ankit Biswas

Mech-interp look at how models represent taste

writeupcode

paper 2026

Degrading Dimensions in Text Embeddings are Domain Specialists 

Ankit Biswas

Under Review at ACL ARR

absTruncating text embeddings is an area of key interest for satisfying memory and compute constraints. This is feasible for many models, as they embed dimensions that individually degrade performance on downstream tasks. We explore what information these degrading dimensions encode and why they hurt task performance with five encoder models across 25 retrieval/classification tasks. We demonstrate that degrading dimensions encode a mix of both monosemantic and polysemantic domain-specific features, finding that strongly-activating dimensions are "domain detectors," discriminating in-domain queries from a general corpus but being useless/harmful otherwise. Thus, we hypothesize that degrading dimensions are specialists for minority domains and harm performance on tasks of other domains, the "degrading" being a sign-averaging artifact of the domain being rare in the task mix. We confirm this, flipping 90.7% of degrading dimensions to improving by re-weighting the evaluation toward each dimension's specialist domain, against 21.2% under a random-domain control. Finally, we train a learned router to ablate dimensions based on the query embedding, outperforming both baseline performance and a blanket ablation of all degrading dimensions on seen domains, suggesting that dimensionality reduction may be adaptable based on task. bib@article{biswas2026degrading, author = {Biswas, Ankit}, title = {Degrading Dimensions in Text Embeddings are Domain Specialists}, journal = {Under Review at ACL ARR}, year = {2026}, url = {https://openreview.net/forum?id=5LQlv1eJ2S} } pdfcode

technical project 2026

arXiv Daemon 

Ankit Biswas

Fetches the best papers from arXiv email notices for me to read each morning

code

paper 2025

Atmospheric model-trained machine learning selection and classification of ultracool TY dwarfs 

Ankit Biswas

Monthly Notices of the Royal Astronomical Society, 541(2)

absAn ensemble machine learning framework trained on atmospheric models to select and classify ultracool T- and Y-type brown dwarfs, enabling the discovery of new substellar candidates. bib@article{biswas2025tydwarfs, author = {Biswas, Ankit}, title = {Atmospheric model-trained machine learning selection and classification of ultracool TY dwarfs}, journal = {Monthly Notices of the Royal Astronomical Society}, volume = {541}, number = {2}, pages = {1670--1681}, year = {2025}, publisher = {Oxford University Press} } pdf

paper 2025

Commissioning an all-sky infrared camera array for detection of airborne objects 

Laura Dominé, Ankit Biswas, Richard Cloete, and 13 more authorsAlex Delacroix, Andriy Fedorenko, Lucas Jacaruso, Ezra Kelderman, Eric Keto, Sarah Little, Abraham Loeb, Eric Masson, Mike Prior, Forrest Schultz, Matthew Szenher, Wesley Andrés Watters, Abigail White

Sensors, 25(3)

absDesign and commissioning of an all-sky infrared camera array, developed by the Galileo Project, for the systematic detection and characterization of airborne objects. bib@article{domine2025allsky, author = {Domin{\'e}, Laura and Biswas, Ankit and Cloete, Richard and others}, title = {Commissioning an all-sky infrared camera array for detection of airborne objects}, journal = {Sensors}, volume = {25}, number = {3}, pages = {783}, year = {2025}, publisher = {MDPI} } pdf

conference 2025

Reconstructing the Cosmic Web Over the First 2 Gyr via Deep Learning Decomposition of the Cosmic Infrared Background 

Ankit Biswas

American Astronomical Society Meeting

absA deep-learning method to decompose the Cosmic Infrared Background and reconstruct the high-redshift large-scale structure of the universe during the first two billion years. bib@inproceedings{biswas2025cosmicweb, author = {Biswas, Ankit}, title = {Reconstructing the Cosmic Web Over the First 2 Gyr via Deep Learning Decomposition of the Cosmic Infrared Background}, booktitle = {American Astronomical Society Meeting Abstracts}, volume = {246}, pages = {418--03}, month = {June}, year = {2025} }

conference 2025

Discovery of the 5 Oldest Population III LBG Candidates with a Population-Synthesis Approach 

Ankit Biswas

American Association for the Advancement of Science

absIdentification of five of the oldest Lyman-break galaxy candidates hosting Population III stars, using a population-synthesis approach to probe the Epoch of Reionization. bib@inproceedings{biswas2024popIII, author = {Biswas, Ankit}, title = {Discovery of the 5 Oldest Population III LBG Candidates with a Population-Synthesis Approach}, booktitle = {American Junior Academy of Science}, year = {2024} }

conference 2024

The Search for Intermediate-Mass Black Holes: Next Steps at NCMNS

Patrick Treuthardt, Ian Hewitt, Grace Ockert, Ankit Biswas

North Carolina Astronomers' Meeting

absAn overview of current research in the (pitch angle)–(supermassive black hole mass) relation at NCMNS. bib@inproceedings{treuthardt2024imbh, author = {Treuthardt, Patrick and Hewitt, Ian and Ockert, Grace and Biswas, Ankit}, title = {The Search for Intermediate-Mass Black Holes: Next Steps at NCMNS}, booktitle = {North Carolina Astronomers' Meeting}, year = {2024} }

media

NASA Science Editorial Team
People-Powered Research With a Twist — Jun 2026

Duke Today
A High School Perspective on Duke Research — Jan 2026

Vertices (self-authored)
Hunting the Cosmos in the World's Driest Desert — Jan 2026

Univ. of Notre Dame
AI collaboration to tag Amboseli photos — Oct 2025

Bloomberg
America's Leading Alien Hunters Depend on AI — May 2025

TEDxYouth@NCSSM
Visions of the End Times — Mar 2025

Society for Science
Five Questions with Ankit Biswas — Jun 2023