Sahel Iqbal
I’m a postdoctoral researcher at the Department of Statistics at the University of Oxford where I work with Tom Rainforth. My research develops methods for reliable decision-making and inference in machine learning systems, especially in settings where models must reason under uncertainty and make effective use of computation. I work across reinforcement learning, Bayesian inference, experimental design, and Monte Carlo methods, with recent interests in inference-time techniques for steering diffusion models and large language models.
For details regarding my research, see my Google Scholar profile. I can be contacted on X and at my email [firstname].[lastname]@stats.ox.ac.uk.
Previously, I completed my PhD with Simo Särkkä at Aalto University, Finland, and before that, I studied physics at NISER, India. Outside work, my time is mostly taken up by reading, lifting weights, and writing JAX code. The projects that I’m actively working on are available on GitHub.
Recent Posts
Asynchronous Data Copies in CuTe DSL
Part 1 of a multi-part series on GPU kernel development with NVIDIA’s CuTe domain-specific language.
Steering Language Models with Sequential Monte Carlo
How to give your language model the blues with SMC.
Using LaTeX Snippets in Markdown Files in Neovim
How I adapted my LuaSnip LaTeX snippets to work in Markdown for easier math note-taking in Neovim.
Recent News
2026-05: Our paper Online Bayesian experimental design for partially observed dynamical systems has been accepted to ICML 2026.
2026-03: Our paper Maximin robust Bayesian experimental design is now out on arXiv. We propose a worst-case decision-making rule for dealing with misspecified models in Bayesian experimental design.
2026-01: We released a new JAX library for inference in state-space models called
cuthbert! Checkout the tweet and the code.2025-09: Our paper Sequential Monte Carlo for Policy Optimization in Continuous POMDPs has been accepted to NeurIPS 2025.
2025-07: If you’re attending MCM 2025, my coauthor Adrien Corenflos will be giving a talk on our joint work on Bayesian experimental design.