Hey! I'm Ali 👋
I'm an incoming CS PhD Student at Northwestern, building cross-domain representation learning methods for robotics and computational biology.
Loading mosaic
Mosaic
Tiles are from robotics datasets I collected or trained policies on.
About Me
Often, we try to improve model performance by building domain-specific methods tuned to particular modalities or benchmarks. I'm interested in building cross-domain methods along two axes: learning stable noise-invariant latent representations and designing model architectures that support reliable long-horizon inference in noisy environments. I develop these methods in both robotics and computational biology. In both settings, I work on self-supervised objectives that learn robust latents and on reward-guided inference methods that correct long-horizon trajectory drift.
I work on these problems at Northwestern University and am fortunate to be advised by Prof. Han Liu and Prof. Zhaoran Wang. My work on virtual cell models and cellular reprogramming is conducted in collaboration with the Chan Zuckerberg Biohub. I like to granularly understand the systems that my methods interact with, so I build accessible robotics hardware. Outside of research, I enjoy mountain biking and swimming.
News
2 updates
Latest
ICML 2026: Two Papers Accepted: On Structured State-Space Duality + Virtual Cells Need Context, Not Just Scale
Starting my CS PhD at Northwestern University
Milestone
Publications & Projects
WorkStatusDescription
Publications & Preprints
Humanity's Last Exam
Long Phan et al.
ML
A large-scale benchmark that stress-tests multimodal foundation models on PhD-level questions across science, engineering, and the humanities.
On Structured State-Space Duality
Jerry Yao-Chieh Hu, Xiwen Zhang, Ali ElSheikh, Weimin Wu, Han Liu
ML
ICML 2026
Extends structured state-space duality to diagonal SSMs and characterizes when an SSM admits a 1-semiseparable masked attention dual.
Cell-JEPA: Latent Representation Learning for Single-Cell Transcriptomics
Ali ElSheikh, Rui-Xi Wang, Weimin Wu, Yibo Wen, et al.
BIO
Latent representation learning for single-cell transcriptomics with a JEPA-style objective trained on 5.8M scRNA-seq cells to learn robust cell embeddings.
Virtual Cells Need Context, Not Just Scale
Payam Dibaeinia, Sudarshan Babu, Mei Knudson, Ali ElSheikh, et al.
BIO
ICML 2026
Position paper arguing virtual cell models need broader biological context coverage and causal transportability, not just larger model capacity.
Manuscripts in Progress
Harnessing PRDM1-PGC1α Axis to Enhance CAR T Cell Therapy
BIO
PRDM1 knockout boosts CD19 CAR T expansion and persistence via PGC1α-driven mitochondrial fitness.
Projects & Software
aArm & SO-100
ROBOT
Built and upgraded the SO-100 platform with new electronics and four additional camera feeds. Designing aArm, a 7-DoF robot arm with QDD actuators, and building its control stack.
QDD Actuator
ROBOT
Inspired by OpenQDD v1; reworked electronics and added a 10:1 helical reducer producing ~20 Nm peak holding torque with a lower-cost FOC driver and magnetic encoder.
Franka Panda Pipetting Apparatus
ROBOT
Equipped a Franka Panda arm with a pipetter and built a custom liquid-handling apparatus for automated pipetting experiments.
GELLO Arm Simulation Teleoperation
ROBOT
Built a GELLO-style leader arm for controlling robot policies in simulation, providing low-cost kinesthetic teleoperation data for simulated manipulation tasks.
Tool-Using Agents
ML
Adapted the VeRL framework to optimize LLM agents for tool-use behavior and strict output formatting via RL post-training.
Options & Portfolio Lab
ML
Toolkit for constructing option spreads and approximating risk-neutral distributions with an ML pipeline for portfolio imputation and hierarchical risk parity.