Julian Asilis

About me

I am a fifth-year PhD student in the USC Theory Group, fortunate to be advised by Vatsal Sharan. These days, I am primarily interested in post-training, including task adaptation, learned representations, and the role of data.

I am grateful to be supported by an NSF Graduate Research Fellowship, and to have spent the summer of 2025 at Pinterest as a machine learning intern. Before the PhD, I worked as a quantitative researcher at the hedge fund AQR and studied math as an undergrad at Harvard.

I'll be graduating in December 2026 and am currently on the industry job market!

Selected Research

Does Random Search Discover Task Experts? Revisiting Neural Thickets
with Guanzhe Hong and Vatsal Sharan.
(arXiv coming soon!)

Resa: Transparent Reasoning Models via SAEs
with Shangshang Wang, Ömer Faruk Akgül, Enes Burak Bilgin, Ollie Liu, Deqing Fu, and Willie Neiswanger.
Conference on Language Modeling (COLM 2026).
[arxiv] | [pdf]

Textual Steering Vectors Can Improve Visual Understanding in Multimodal Large Language Models
with Woody Gan, Deqing Fu, Ollie Liu, Dani Yogatama, Vatsal Sharan, Robin Jia, and Willie Neiswanger.
Association for Computational Linguistics (ACL 2026).
[arxiv] | [pdf]

Tina: Tiny Reasoning Models via LoRA
with Shangshang Wang, Ömer Faruk Akgül, Enes Burak Bilgin, Ollie Liu, and Willie Neiswanger.
International Conference on Learning Representations (ICLR 2026).
[arxiv] | [pdf]

Additional Research

Algorithmic Principles For Multiclass Learning Are Hard To Come By
with Shaddin Dughmi, Vatsal Sharan, Alec Sun, Shang-Hua Teng, and Chang Wang.
In submission, 2026.
[arxiv] | [pdf]

When Clean Data Hurts: Learning With Monotone Corruptions
with Shaddin Dughmi and Chirag Pabbaraju
In submission, 2026.
[arxiv] | [pdf]

Semi-Random Graphs, Robust Asymmetry, and Reconstruction
with Xi Chen, Dutch Hansen, and Shang-Hua Teng.
Innovations in Theoretical Computer Science (ITCS 2026).
[arxiv] | [pdf]

On Agnostic PAC Learning in the Small Error Regime
with Mikael Møller Høgsgaard and Grigoris Velegkas.
Neural Information Processing Systems (NeurIPS 2025).
Spotlight paper
[arxiv] | [pdf]

Local Regularizers Are Not Transductive Learners
with Sky Jafar and Shaddin Dughmi.
Conference on Learning Theory (COLT 2025).
[arxiv] | [pdf]

Understanding Aggregations of Proper Learners in Multiclass Classification
with Mikael Møller Høgsgaard and Grigoris Velegkas.
Algorithmic Learning Theory (ALT 2025).
[arxiv] | [pdf]

Proper Learnability and the Role of Unlabeled Data
with Siddartha Devic, Shaddin Dughmi, Vatsal Sharan, and Shang-Hua Teng.
Algorithmic Learning Theory (ALT 2025).
[arxiv] | [pdf]

Transductive Learning Is Compact
with Siddartha Devic, Shaddin Dughmi, Vatsal Sharan, and Shang-Hua Teng.
Neural Information Processing Systems (NeurIPS 2024).
[arxiv] | [pdf] | [slides]

Open Problem: Can Local Regularization Learn All Multiclass Problems?
with Siddartha Devic, Shaddin Dughmi, Vatsal Sharan, and Shang-Hua Teng.
Open Problem @ Conference on Learning Theory (COLT 2024).
[pdf] | [slides]

Regularization and Optimal Multiclass Learning
with Siddartha Devic, Shaddin Dughmi, Vatsal Sharan, and Shang-Hua Teng.
Conference on Learning Theory (COLT 2024).
[arxiv] | [pdf] | [slides]

Computable PAC Learning of Continuous Features
with Nate Ackerman, Jieqi Di, Cameron Freer, and Jean-Baptiste Tristan.
Logic in Computer Science (LICS 2022).
[pdf]

Thesis

Probability Monads, under the direction of Michael Hopkins.

Teaching

Computer Science I: CSCI 1101 @ BC, Spring 2022 Head TA. Notes here.

Applied Machine Learning: CSCI 3340.01 @ BC, Fall 2021 Teaching assistant. Notes here.

Sets, Groups, and Topology: Math 101 @ Harvard, Spring 2020 Course assistant. Partial notes here.

Real Analysis I: Math 112 @ Harvard, Spring 2019 Course assistant. Notes here.

Abstract Algebra I: Math 122 @ Harvard, Fall 2018 Course Assistant. Notes here, taken by Vaughan McDonald.

Course Notes

Advanced algorithms: CS 670 @ USC. Shortest paths, spanning trees, matroids, Fibonacci heaps, dynamic programming, max-flow, hardness.

Combinatorial analysis: Math 532 @ USC. (Exponential) generating functions, inclusion and exclusion, Mobius inversion, set and number partitions.

Algebraic geometry: Math 137 @ Harvard. Algebraic sets, Nullstellensatz (again), local rings, DVRs. Notes for first half of the course.

Commutative algebra: Math 221 @ Harvard. Localization, Nullstellensatz, Tor, Nakayama, dimension theory.

Category theory: Math 99r @ Harvard. Functors, natural transformations, Yoneda, (co)limits.

Theme by orderedlist