Portrait of Riyasat Ohib

Riyasat Ohib

I’m a Research Scientist at Google DeepMind in New York. I did my PhD at Georgia Tech with Vince Calhoun and Sergey Plis, working on sparse learning in deep models.

I study how neural networks represent what they learn, how those representations shape behavior, and how understanding them can make models more efficient, interpretable, and controllable. I first explored these questions in research internships at Google DeepMind, Cohere, Dolby Laboratories, and Meta FAIR. Behind all of it is one question: how learning works, in brains and in machines.

When I’m not doing research, I’m usually playing soccer, traveling, hiking, or reading about science, history, and how the world works. I keep a list of what I’ve been reading.

  1. Teaser figure from Concept Spaces in the Residual Stream of Diffusion Transformers

    CVPR HOW 2026

    Concept Spaces in the Residual Stream of Diffusion Transformers

    Riyasat OhibMeera HahnMani Malek

    Training-free steering of diffusion transformers by intervening on concept directions in the residual stream, often matching or beating prompting with smoother control.

    PDF HTML

  2. Teaser figure from SSFL: Discovering Sparse Unified Subnetworks at Initialization for Efficient Federated Learning

    TMLR 2026

    SSFL: Discovering Sparse Unified Subnetworks at Initialization for Efficient Federated Learning

    Riyasat OhibBishal ThapaliyaGintare Karolina DziugaiteJingyu LiuVince D. CalhounSergey Plis

    SSFL finds sparse subnetworks before federated training starts, reducing communication while preserving accuracy across non-IID clients.

    HTML Code

  3. Teaser figure from Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts

    COLM 2025

    Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts

    Samin Yeasar ArnobZhan SuMinseon KimOleksiy OstapenkoRiyasat OhibEsra'a SalehDoina PrecupLucas CacciaAlessandro Sordoni

    Sparse adapters provide simple, parameter-efficient task experts that merge more effectively than LoRA or full fine-tuning across as many as 20 NLP tasks.

    HTML

  4. Teaser figure from Efficient Reinforcement Learning by Discovering Neural Pathways

    NeurIPS 2024

    Efficient Reinforcement Learning by Discovering Neural Pathways

    Samin Yeasar ArnobRiyasat OhibSergey M. PlisAmy ZhangAlessandro SordoniDoina Precup

    Neural pathways identify compact subnetworks for reinforcement learning, retaining strong performance with less than 5% of a larger network's parameters and supporting multiple tasks within one model.

    PDF HTML

All publications

News

  1. Joined Google DeepMind as a Research Scientist in New York, working on understanding representations and how to control frontier multimodal models.

  2. Completed my PhD at Georgia Tech with Vince Calhoun and Sergey Plis. Dissertation: The Hidden Structure of Deep Models: Discovering and Exploiting Sparse Subnetworks.

  3. Learning to Query History accepted at the ICLR TSALM workshop, on retrieval-augmented classification under nonstationarity.

  4. Concept Spaces in the Residual Stream of Diffusion Transformers accepted at the CVPR HOW workshop, on steering generation through concept directions.

News archive