Shiye Su

Learning to See Through Language: An Exploration of Language Modeling's Effect on Visual Representations

Mark Endo, Shiye Su, Yuhui Zhang, Serena Yeung-Levy

Does language modeling actually improve vision, or just bend visual features toward text? We find that visual-instruction-fine-tuning can sharpen semantics and learn stronger representations for downstream applications, but can also erode fine-grained geometry. We propose a lightweight adapter to turn frozen MLLMs into competitive visual backbones.

Rethinking Visual Understanding in Text-to-Image Reward Modeling

Shiye Su, Xiaohan Wang, Ludwig Schmidt, Serena Yeung-Levy

We find that widely-used reward models for text-to-image generation are remarkably poor visual reasoners, with these capabilities degrading during preference tuning. We leverage existing multimodal learning data to create a novel type of preference pairs that differ along the text rather than image axis, and use them to achieve improved prompt adherence.

Biologically-Constrained Virtual Cell Modeling via Reinforcement Learning

Dongxia Wu*, Shiye Su*, Yuhui Zhang*, Elaine Sui, Emma Lundberg, Emily Fox, Serena Yeung-Levy

We advance cellular perturbation modeling beyond "visually realistic" generations towards "biologically meaningful" ones by introducing a suite of biogically-informed rewards. Used as additional evaluators, training signals for reinforcement post-training, and verifiers for test time scaling, these rewards yield the new state-of-the-art CellFluxRL.

Stochastic Perturbations Improve Distribution-to-Distribution Generative Models

Shiye Su, Yuhui Zhang, Linqi Zhou, Rajesh Ranganath, Serena Yeung-Levy

CVPR 2026 Findings

We identify sparse supervision as a key challenge in the distribution-to-distribution learning and improve performance using simple and effective perturbations during training.

OpenThoughts: Data Recipes for Reasoning Models

Etash Guha, Ryan Marten, Sedrick Keh, Negin Raoof, Georgios Smyrnis, Hritik Bansal, Marianna Nezhurina, Jean Mercat, Trung Vu, Zayne Sprague, Ashima Suvarna, Benjamin Feuer, Liangyu Chen, Zaid Khan, Eric Frankel, Sachin Grover, Caroline Choi, Niklas Muennighoff, Shiye Su, Wanjia Zhao, John Yang, Shreyas Pimpalgaonkar, Kartik Sharma, Charlie Cheng-Jie Ji, Yichuan Deng, Sarah Pratt, Vivek Ramanujan, Jon Saad-Falcon, Jeffrey Li, Achal Dave, Alon Albalak, Kushal Arora, Blake Wulfe, Chinmay Hegde, Greg Durrett, Sewoong Oh, Mohit Bansal, Saadia Gabriel, Aditya Grover, Kai-Wei Chang, Vaishaal Shankar, Aaron Gokaslan, Mike A. Merrill, Tatsunori Hashimoto, Yejin Choi, Jenia Jitsev, Reinhard Heckel, Maheswaran Sathiamoorthy, Alexandros G. Dimakis, Ludwig Schmidt

ICLR 2026 Oral; NeurIPS 2025, Workshop on Foundations of Reasoning in Language Models

We create open-source datasets for training reasoning models and investigate each step of the data geneartion pipeline with 1,000+ controlled experiments, yielding the state-of-the-art OpenThoughts3 model.

Generative Neural Networks for Kerr Combs

Janet Zhong, Eran Lustig, Shiye Su, Louise Schul, Jamison Sloan, Congyue Deng, Jelena Vuckovic, Shanhui Fan

NeurIPS 2025, Machine Learning and the Physical Sciences Workshop

We apply generative modeling to the inverse design of Kerr micro-resonators, finding high-bandwidth steady-state solitons as given by Lugiato-Lefever equation simulations.

Explaining Hypergraph Neural Networks: From Local Explanations to Global Concepts

Shiye Su, Iulia Duta, Lucie Charlotte Magister, Pietro Liò

We design an architecture-agnostic, post-hoc method to produce faithful and concise explanations for hypergraph neural networks.

Commute-Time-Optimised Graphs for GNNs

Igor Sterner*, Shiye Su*, Petar Veličković

ICML 2024, GRaM workshop (PMLR)

We develop a graph rewiring procedure based on Cayley expanders, reducing oversquashing and incorporating user priors.

ALMANACS: A Simulatability Benchmark for Language Model Explainability

Edmund Mills, Shiye Su, Stuart Russell, Scott Emmons

We introduce a simulatability benchmark for language model interpretability, comprising structured, idiosyncratic questions with test-time distributional shift.

Fitness Aware Human Motion Generation with Fine-Tuning

Kiril Bikov, Shiye Su, Deepro Choudhury, Zhilin Guo, Weihao Xia, Mehmet Salih Çeliktenyıldız, Chenliang Zhou, Param Hanji, Cengiz Oztireli

NeurIPS 2025, FITML Workshop

We condition a human motion diffusion model on subject fitness, learning from a motion-capture dataset of graded Functional Movement Screen actions.

Chaos and Measurement-Induced Criticality on Stabiliser Circuits

Shiye Su

My undergraduate thesis was advised by Sarang Gopalakrishnan and David Huse. We studied the entanglement transition in a many-body system evolving under Clifford gates and projective measurements, as well as the spectral and thermalisation properties of Clifford gates.

Shiye Su

About

Research

Awards

Misc