Portrait photo of Saurav Jha, IVADO postdoctoral researcher at MILA

I am an ivado postdoctoral researcher at MILA, working with Sarath Chandar. My recent research interests span vision-language-action (VLA) world models, (continual/temporal) learning on streaming data, and real-world verification of generative model outputs.

I completed my PhD at UNSW Sydney in August 2025, where I was advised by Lina Yao and Dong Gong. During the latter half of my PhD, I worked as an applied research scientist at openstream.ai, and as a research intern at Sony (hosted by Shiqi Yang and Shusuke Takahashi ) and Tencent (hosted by Shengju Qian).

Prior to my PhD, I worked on continual learning with Joost van de Weijer, and did my Erasmus Mundus Joint Master's Degree (EMJMD) in Advanced Systems Dependability at the University of St Andrews, the UK and l'Université de Lorraine, France. During my master's, I interned in Emmanuel Vincent's group at Inria Nancy. I once wrote this medium blog documenting my EMJMD experience to help guide future aspirants.

In my free time, I enjoy learning to travel, snorkelling, hiking, and binge-watching. I grew up in eastern Nepal, and every couple of years, I like to plan week-long treks in the Nepalese Himalayas.

I am always keen on hearing from students and collaborators who are interested in my research. Please feel free to reach out with any ideas/questions.

news

Experience

IVADO postdoctoral fellow (Sep 2025 - Present)

MILA • Montréal, Canada 🇨🇦

Research aligned with advancing Canada's R3AI initiative.

Applied AI Scientist (May 2025 - Aug 2025)

OpenStream.ai • Melbourne, Australia 🇦🇺

Infra-focus: Developed production-grade conversational LLM agents for enterprise clients.

ML-focus: Implemented & shipped a POC for neuro-symbolic verification of multi-agent systems.

AI Research Intern (Sep 2024 - Mar 2025)

LightSpeed Studios, Tencent • Sydney, Australia 🇦🇺

Worked on controllable image generation and preference optimization for multi-modal LLMs.

Research Scientist Intern (May 2024 - Aug 2024)

Creative AI Lab, Sony Group Corporation • Tokyo, Japan 🇯🇵

Worked on continual personalization of pre-trained text-to-image diffusion models.

Research Assistant (Sep 2021 - Jan 2022)

Computer Vision Centre, Universitat Autònoma de Barcelona • Barcelona, Spain 🇪🇸

Worked on rehearsal-free continual learning for Vision Transformers (ViTs).

Research Intern (Mar 2021 - Jul 2021)

Multispeech group, Inria Nancy • Nancy, France 🇫🇷

Worked on learning domain-specific language models for speech recognition.

Machine Learning Engineer (Jun 2018 - Jul 2019)

FactSet Research Systems Inc. • Hyderabad, India 🇮🇳

Worked on improving FactSet's named entity recognition service with acronym disambiguation and neural topic modeling.

Awards & Recognition

Academic Services

tutoring at UNSW

Selected Publications

Probing the Effectiveness of World Models for Spatial Reasoning

Probing the Effectiveness of World Models for Spatial Reasoning through Test-Time Scaling

Saurav Jha, M. Jehanzeb Mirza, Wei Lin, Shiqi Yang, Sarath Chandar

World Modeling Workshop 2026

Paper Code

We propose Verification through Spatial Assertions (ViSA), a proposer-solver method that enables faithful test-time verification of world model views for enhancing the spatial reasoning in existing VLMs.

Mining Your Own Secrets

Mining Your Own Secrets: Diffusion Classifier Scores for Continual Personalization of Text-to-Image Diffusion Models

Saurav Jha, Shiqi Yang, Masaki Ishii, Meng Zhao, Christian Simon, Jehanzeb Mirza, Dong Gong, Lina Yao, Shusuke Takahashi, Yuki Mitsufuji

ICLR 2025

Paper Project page

We propose using diffusion classifier scores for regularizing the parameter-space and function-space of text-to-image diffusion models, to achieve continual personalization.

CLAP4CLIP

CLAP4CLIP: Continual LeArning with Probabilistic finetuning for Vision-Language Models

Saurav Jha, Dong Gong, Lina Yao

NeurIPS 2024

Paper Code

Our work proposes Continual LeArning with Probabilistic finetuning (CLAP) - a probabilistic modeling frame- work over visual-guided text features per task, thus providing more calibrated CL finetuning.

NPCL

NPCL: Neural Processes for Uncertainty-Aware Continual Learning

Saurav Jha, Dong Gong, He Zhao, Lina Yao

NeurIPS 2023

Paper Code

We propose a neural process-based continual learning approach with task-specific modules arranged in a hierarchical latent variable model. We tailor regularizers on the learned latent distributions to alleviate forgetting.

Towards Exemplar-Free Continual Learning

Towards Exemplar-Free Continual Learning in Vision Transformers: an Account of Attention, Functional and Weight Regularization

Francesco Pelosin*, Saurav Jha*, Andrea Torsello, Bogdan Raducanu, Joost van de Weijer

CVPR 2022 Workshop on Continual Learning (CLVision)

Paper Code

We investigate the continual learning of Vision Transformers (ViT) for the challenging exemplar-free scenario, with special focus on how to efficiently distill the knowledge of its crucial self-attention mechanism.