Jialin Song

Burnaby, BC, Canada

jialin_song [at] sfu.ca

I am a third-year Ph.D. student in Computing Science at Simon Fraser University, where I work on the safety and reliability of machine learning systems, advised by Prof. Wuyang Chen.

My current research has two threads. The first is large language model safety: I build scalable benchmarks for stress-testing models against multi-turn conversational attacks, and study why safety training that holds up in a single turn erodes over a longer dialogue. The second is scientific machine learning: I develop data-efficient operator learning methods, physics-informed surrogates for high-performance computing workloads, and language models that formalize and reason about partial differential equations.

Before SFU I was a machine learning engineer at the International Computer Science Institute in Berkeley, where I worked on robust and scientific machine learning with Prof. Michael Mahoney. I spent the summer of 2025 and 2026 as a research intern with the Deep Learning Group at Microsoft Research, Redmond, WA, working with Xiaodong Liu and Weiwei Yang. I hold an M.Eng. in EECS from UC Berkeley and a B.S. in Computer Science and Robotics Engineering from Worcester Polytechnic Institute.

Feel free to reach out if you would like to talk about any of this.

news

Oct 2026	Two blog posts about our work are up on the Microsoft Research blog: post one and post two.
Sep 2026	The MultiBreak dataset is now released on Hugging Face.
May 2026	Started a research internship at Microsoft Research in Redmond, WA.
Apr 2026	MultiBreak was accepted to ICML 2026.

selected publications

ICML

MultiBreak: A Scalable and Diverse Multi-turn Jailbreak Benchmark for Evaluating LLM Safety

Jialin Song, X. Liu, W. Yang, and 4 more authors

In International Conference on Machine Learning (ICML), 2026

Abs arXiv

A scalable multi-turn jailbreak benchmark built with active-learning-based data expansion across diverse harm categories, used to stress-test the safety behaviour of large language models and to expose persistent vulnerabilities to conversational attacks.
ICML

PDE-Controller: LLMs for Autoformalization and Reasoning of PDEs

M. Soroco^*, Jialin Song^*, M. Xia, and 3 more authors

In International Conference on Machine Learning (ICML), 2025

Abs arXiv

A large language model that converts natural-language descriptions of partial differential equations into formal temporal logic and executable control code, trained with a back-translation and auto-verification pipeline for data generation.
NeurIPS

Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning

W. Chen^*, Jialin Song^*, P. Ren, and 3 more authors

In Advances in Neural Information Processing Systems (NeurIPS), 2024

Abs arXiv

An unsupervised pretraining and in-context learning pipeline that reduces the simulation data needed to train neural operators, studied across masking ratios, subsampling regimes and several PDE systems.