Nicolás Astorga

Nicolás Astorga

Ph.D. student in Machine Learning

University of Cambridge | van der Schaar Lab

About Me

I’m a Ph.D. student at the University of Cambridge advised by Prof. Mihaela van der Schaar. My research explores Reasoning, Bayesian experimental design, and Optimization in the context of LLMs. Previously, I completed a two M.Sc. in Electrical Engineering and Computer Science at the University of Chile. I also worked as an ML Engineer at ALeRCE, and interned at Harvard IACS. My research contributions have been published in leading conferences, including NeurIPS, ICML, ICLR (Spotlight), ECCV, and AISTATS.

I am really interested in LLM research and leveraging their inductive biases to drive exploration and exploitation, and then using the collected experience to improve those biases (via in-context learning or training). This exploration–exploitation is with a goal in mind, making search methods and efficient experimentation also interesting. These ideas admit many implementations, which is part of the fun. I love that LLMs—and deep learning more broadly—are built to search and learn.

Interests
  • Large Language Models (LLMs)
  • Reasoning & RL with LLMs
  • Bayesian Experimental Design, Active Learning & Bayesian Optimization
  • Autoformulation for optimization with LLMs
  • Generative Models & Variational Inference
Education

  • Ph.D. in Machine Learning

    University of Cambridge (2023–present)

  • Dual M.Sc. — Electrical Engineering; Computer Science

    University of Chile (2020–2023)

  • B.Sc. — Computer, Electrical & Mechanical Engineering (Three Major)

    University of Chile (2013–2019)

Publications (Highlighted)

Auto‑formulation of Mathematical Optimisation Models Using Large Language Models
Automatically formulating optimisation models from natural language using LLMs. Accepted at ICML 2025.
Nicolás Astorga, Tennison Liu, Y. Xiao, Mihaela van der Schaar
Jun 1, 2025
arxiv
Auto‑formulation of Mathematical Optimisation Models Using Large Language Models
Active Task Disambiguation with LLMs
This paper formalizes task ambiguity and frames task disambiguation as Bayesian Experimental Design, yielding effective clarifying question selection for LLMs. Accepted at ICLR 2025.
Katarzyna Kobalczyk, Nicolás Astorga, Tennison Liu, Mihaela van der Schaar
Jan 22, 2025
arXiv
Active Task Disambiguation with LLMs
Active Learning with LLMs for Partially Observed and Cost‑Aware Scenarios
Active learning strategies for partially observed, cost‑aware settings via LLMs. Accepted at NeurIPS 2024.
Nicolás Astorga, Tennison Liu, N. Seedat, Mihaela van der Schaar
Dec 1, 2024
arxiv
Active Learning with LLMs for Partially Observed and Cost‑Aware Scenarios
Large Language Models to Enhance Bayesian Optimisation
Leveraging LLMs to improve Bayesian optimisation. Accepted at ICLR 2024.
Tennison Liu, Nicolás Astorga, N. Seedat, Mihaela van der Schaar
May 1, 2024
arxiv
Large Language Models to Enhance Bayesian Optimisation

Publications (All)

Conferences — First author

  • N. Astorga*, T. Liu*, Y. Xiao, M. van der Schaar (2025). Auto-formulation of Mathematical Optimisation Models Using Large Language Models. ICML 2025. *Equal contribution.
  • K. Kobalczyk*, N. Astorga*, T. Liu, M. van der Schaar (2025). Active Task Disambiguation with Large Language Models. ICLR 2025 (Spotlight). *Equal contribution.
  • N. Astorga, T. Liu, N. Seedat, M. van der Schaar (2024). Active Learning with LLMs for Partially Observed and Cost-Aware Scenarios. NeurIPS 2024.
  • T. Liu*, N. Astorga*, N. Seedat, M. van der Schaar (2024). Large Language Models to Enhance Bayesian Optimisation. ICLR 2024. *Equal contribution.
  • N. Astorga, P. Huijse, P. Protopapas, P. Estévez (2020). MPCC: Matching Priors and Conditionals for Clustering. ECCV 2020, Glasgow.
  • N. Astorga, P. Huijse, P. A. Estévez, F. Förster (2018). Clustering of Astronomical Transient Candidates Using Deep Variational Embedding. IJCNN 2018, Rio de Janeiro.

Conferences — Second author

  • S. Ruhrberg, N. Astorga, M. van der Schaar (2025). Timely Clinical Diagnosis through Active Test Selection. NeurIPS 2025.
  • H. Amad, N. Astorga, J.-M. van der Schaar (2025). Continuously Updating Digital Twins Using Large Language Models. AISTATS 2025.
  • J. Piskorz, N. Astorga, J. Berrevoets, M. van der Schaar (2025). Active Feature Acquisition for Personalised Treatment Assignment. ICML 2025.

Journals

  • G. Cabrera-Vives, D. Moreno-Cartagena, N. Astorga, I. Reyes-Jainaga, et al. (2024). ATAT: Astronomical Transformer for Time Series and Tabular Data. Astronomy & Astrophysics.
  • M. Pérez-Carrasco, G. Cabrera-Vives, L. Hernández-García, F. Förster, N. Astorga, et al. (2023). Alert Classification for the ALeRCE Broker System: The Anomaly Detector. The Astronomical Journal.
  • F. Förster, G. Cabrera-Vives, E. Castillo-Navarrete, P. A. Estévez, N. Astorga, et al. (2021). The Automatic Learning for the Rapid Classification of Events (ALeRCE) Alert Broker. The Astronomical Journal.
  • C. Modarres, N. Astorga, E. Droguett, V. Meruane (2018). Convolutional Neural Networks for Automated Damage Recognition and Damage Type Identification. Structural Control and Health Monitoring.

Workshop

  • H. Sun, T. Pouplin, N. Astorga, T. Liu, M. van der Schaar (2024). Improving LLM Generation with Inverse and Forward Alignment: Reward Modelling, Prompting, Fine-Tuning, and Inference-Time Optimisation. NeurIPS 2024 Workshop on System-2 Reasoning at Scale.

Professional Experience

 
 
 
 
 
ALeRCE – Automatic Learning for the Rapid Classification of Events
Machine Learning Engineer
ALeRCE – Automatic Learning for the Rapid Classification of Events
February 2022 – April 2023 Santiago, Chile
  • Deployed production‑grade ML models via Kubernetes to classify LSST astronomical alerts in real time.
  • Built distributed PySpark pipelines to curate >30M light‑curve observations from multiple catalogues.
  • Collaborated in the ELAsTiCC challenge; proposed a Transformer‑based model for tabular/time‑series data; work accepted at Astronomy & Astrophysics.
 
 
 
 
 
Harvard University — Institute for Applied Computational Science
Research Intern
Harvard University — Institute for Applied Computational Science
January 2019 – August 2019 Cambridge, MA, USA
  • Proposed MPCC, a GAN–VAE hybrid clustering framework (ECCV 2020) leveraging forward KL divergence and extending BigGAN.
 
 
 
 
 
University of Chile — Lab. of Computational Intelligence
Research Assistant
University of Chile — Lab. of Computational Intelligence
March 2016 – December 2023 Santiago, Chile
  • Developed a VAE‑based clustering method for astronomical transient detection (IJCNN 2018).
  • Integrated normalising flows into variational embeddings, improving ELBO by ≥10%.
  • Matched fully supervised performance using Gaussian processes in a semi‑supervised setting with only 10% labeled data.

Contact

  • nja46@cam.ac.uk
  • Centre for Mathematical Sciences, Wilberforce Rd, Cambridge, CB3 0WA,