Licong Xu | Institute of Astronomy, Cambridge

Bio

Hi! I’m Licong Xu, a PhD student at the University of Cambridge, based at the Institute of Astronomy and the Kavli Institute for Cosmology, working at the intersection of cosmology, scientific computing, and artificial intelligence. I completed my BA and MSci in Natural Sciences at Cambridge, specialising in astrophysics.

My research focuses on extracting cosmological and astrophysical information from secondary anisotropies of the cosmic microwave background, particularly the thermal Sunyaev–Zel’dovich (tSZ) effect and galaxy clusters. I develop and study different statistical summaries of these signals, including cluster number counts and the tSZ power spectrum, and investigate how they can be used to probe both cosmology and the astrophysics of baryonic feedback. A major theme of my work is developing end-to-end analysis pipelines and inference methods applied to cosmological simulations, connecting theoretical models and simulations to observable signals and their physical interpretation through approaches including Bayesian likelihood-based inference and simulation-based inference (SBI).

I am also broadly interested in machine learning and AI for scientific discovery. I make extensive use of modern AI systems, including open-source models, coding agents, and agentic workflows, across my research and software development. Beyond using AI as a tool, I have also worked on agentic AI for scientific discovery, developing systems in which LLM agents can plan, use tools, execute code, and collaborate on scientific research tasks.

News

Sep 2026
Talk at Cosmology from Home: “Towards a Joint Analysis of tSZ Cluster Counts and the Masked Power Spectrum: Opportunities and Limitations.”
Sep 2026
The Denario project was accepted at Physical Review X.
Jul 2026
Our simulation-based inference analysis of the tSZ power spectrum is published in Physical Review D (arXiv:2606.14622).
Jul 2026
Competing with AI Scientists was accepted at Communications AI & Computing (Nature Portfolio).
Mar 2026
Presented work on autonomous discovery in cosmology at the 60th Rencontres de Moriond (arXiv:2605.14791).
Dec 2025
First place and Special Jury Prize in the Fair Universe Weak Lensing Uncertainty Challenge at NeurIPS 2025, using CMBAgent.
Jul 2025
Two papers at the ML4Astro workshop at ICML 2025, including a Spotlight on RAG agents (arXiv:2507.07257, arXiv:2507.07155).

Articles

L. Xu, Í. Zubeldia, J. Alvey, B. Bolliet, and A. Challinor

Impact of non-Gaussian likelihood on cosmological constraints from the thermal Sunyaev–Zel’dovich power spectrum: a simulation-based inference analysis

Phys. Rev. D 114, 043528 (2026)

journal arXiv:2606.14622 pdf
L. Xu and T. Borrett

Beyond AI as Assistants: Toward Autonomous Discovery in Cosmology

60th Rencontres de Moriond on Cosmology (2026)

arXiv:2605.14791 pdf
T. Borrett, L. Xu, A. Nilipour, B. Bolliet, et al.

Competing with AI Scientists: Agent-Driven Approach to Astrophysics Research

Accepted at Communications AI & Computing (Nature Portfolio)

arXiv:2604.09621 pdf
F. Villaescusa-Navarro, B. Bolliet, P. Villanueva-Domingo, et al. (incl. L. Xu)

The Denario project: Deep knowledge AI agents for scientific discovery

Accepted at Phys. Rev. X

arXiv:2510.26887 pdf code
L. Xu, M. Sarkar, A. I. Lonappan, Í. Zubeldia, et al.

Open Source Planning & Control System with Language Agents for Autonomous Scientific Discovery

ML4Astro workshop at ICML 2025

arXiv:2507.07257 pdf code
X. Xu, B. Bolliet, A. Dimitrov, A. Laverick, F. Villaescusa-Navarro, L. Xu, and Í. Zubeldia

Evaluating Retrieval-Augmented Generation Agents for Autonomous Scientific Discovery in Astrophysics

Spotlight, ML4Astro workshop at ICML 2025

arXiv:2507.07155 pdf

Code

Coding agents now write a large share of the software behind my analyses. With the physics and the tests specified, an agent can deliver a faster version of an existing tool in time for me to use it in the analysis itself. I plan to release this workflow (working in progress ...).

CMBAgent

Open-source Planning & Control multi-agent system for autonomous scientific research. More details in the paper arXiv:2507.07257.
Denario

Modular multi-agent system for scientific research assistance. More details in the paper arXiv:2510.26887.
hmfast

Differentiable JAX-native halo-model code for cosmological observables, including the tSZ power spectrum. The tszsbi codebase for the simulation-based inference analysis in Xu et al., Phys. Rev. D 114, 043528 (2026) is now merged here.
JXPaint

GPU Compton-y map painter in Python and JAX, written with coding agents as a port of XGPaint. Bit-for-bit with the reference painter and about 590× faster, and used to paint the tSZ maps in my analyses.
GMaster

GPU implementation of NaMaster for pseudo-C_ℓ spectra of masked HEALPix maps, written with coding agents. It accelerates masked power-spectrum estimation and is used in the joint tSZ analysis.