AI4AI Lab

The AI4AI Lab (AI for Actionable Impact) is a machine learning research group at Imperial College London. We build AI that people can act on: models whose outputs can be used to make real decisions in high-stakes settings.

Meet the team Publications →

Research

CausalityWhat happens when we act

InterpretabilityWhy a model decides as it does

UncertaintyWhat we don't know

PeopleWho acts on the decision

Actionability rests on three cornerstones that inform one another, with people at the centre. We work on three central questions:

People · I How can humans and AI work together for consequential decisions? (learning to defer, concept-based models, inverse reinforcement learning)
U How do we interact with the environment, while avoiding risk but learning as quickly as possible? (meta-learning, Bayesian optimisation, model selection, uncertainty quantification)
C How should AI act when its decisions change the world it learns from? (off-policy evaluation, distribution shift, continual learning, causal discovery, causal inference and experimental design)

These questions are relevant for small-data decision-making (e.g. clinical decision systems) and for large-scale foundation models.

Recent news

Oct 2026	A student team supervised by Sonali reached the global top 10 of the Silicon Labs IoT Challenge 2026, the highest-placed UK team, with ArterialAge, a wearable prototype for earlier detection of heart failure. The team pitched in the final in Hanoi. Congratulations to Pranav Sharma, Abhi Kapoor, Krish Thakrar, Anjali Ketan and Joshika Manikandan. Read Pranav’s story on Imperial News.
Sep 2026	Sonali gave an invited talk, Dharma, Demonstration and the Alignment of Machines, at Hinduism and AI: A Roundtable at the University of Cambridge, a full-day meeting on what Hindu thought can contribute to the design, use and governance of intelligent systems. The roundtable was convened by the Stanford Law AI Initiative, Cambridge’s Centre for Human-Inspired Artificial Intelligence and the BAPS Swaminarayan Research Institute.
Aug 2026	Sonali gave an invited talk at the MCML workshop on causal machine learning in Munich.
Jul 2026	Two papers out: Reward Shaping Control Variates for Off-Policy Evaluation Under Sparse Rewards at ICML 2026, and our blue sky ideas paper Causal Methods for LLM Development and Evaluation at KDD 2026.
Jun 2026	Reproducible clinical archetypes in acute respiratory failure: a multi-cohort trajectory analysis was accepted at Intensive Care Medicine.
May 2026	Off by a beat: the effects of temporal misalignment in reinforcement learning for sepsis treatment was accepted at npj Digital Medicine.
Apr 2026	The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives was presented at ICLR 2026.

The team

Full team page →

Sonali Parbhoo

Assistant Professor, Lab lead

Dominic Marshall

MD-PhD student

Ritam Majumdar

PhD student

Nikita Narayanan

PhD student

Elizaveta Sheremetyeva

PhD student

Haoyang Hong

PhD student

Yuxin Wang

Visiting PhD student, LMU Munich

Hannah Liu

Visiting Researcher, Imperial College London

Pranav Sharma

Undergraduate researcher

Selected publications

All publications →

arXiv

Continual Learning without Continual Training

Nikita Narayanan, Ritam Majumdar, and Sonali Parbhoo

arXiv preprint arXiv:2610.10379, 2026

PDF Bib

@article{Narayanan2026Continual,
  title = {Continual Learning without Continual Training},
  author = {Narayanan, Nikita and Majumdar, Ritam and Parbhoo, Sonali},
  journal = {arXiv preprint arXiv:2610.10379},
  year = {2026},
}

KDD

Causal methods for LLM development and evaluation

Dennis Frauen, Marie Brockschmidt, Konstantin Hess, and 10 more authors

In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, 2026

PDF Bib

@inproceedings{Frauen2026Causal,
  title = {Causal methods for LLM development and evaluation},
  author = {Frauen, Dennis and Brockschmidt, Marie and Hess, Konstantin and Ma, Haorui and Ma, Yuchen and Maarouf, Abdurahman and Schröder, Maresa and Schweisthal, Jonas and Wang, Yuxin and Deviyani, Athiya and Parbhoo, Sonali and Krishnan, Rahul G and Feuerriegel, Stefan},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2},
  pages = {13170--13177},
  year = {2026},
}

ICML

Reward Shaping Control Variates for Off-Policy Evaluation Under Sparse Rewards

Ritam Majumdar, Finale Doshi-Velez, and Sonali Parbhoo

In Forty-third International Conference on Machine Learning, 2026

PDF Bib

@inproceedings{Majumdar2026Reward,
  title = {Reward Shaping Control Variates for Off-Policy Evaluation Under Sparse Rewards},
  author = {Majumdar, Ritam and Doshi-Velez, Finale and Parbhoo, Sonali},
  booktitle = {Forty-third International Conference on Machine Learning},
  year = {2026},
}

Intensive Care Med

Reproducible clinical archetypes in acute respiratory failure: a multi-cohort trajectory analysis

Dominic C Marshall, Ashleigh D Green, Matthieu Komorowski, and 3 more authors

Intensive Care Medicine, 2026

PDF Bib

@article{Marshall2026Reproducible,
  title = {Reproducible clinical archetypes in acute respiratory failure: a multi-cohort trajectory analysis},
  author = {Marshall, Dominic C and Green, Ashleigh D and Komorowski, Matthieu and Patel, Brijesh V and Parbhoo, Sonali and Antcliffe, David B},
  journal = {Intensive Care Medicine},
  pages = {1--12},
  publisher = {Springer Berlin Heidelberg},
  year = {2026},
}

npj Digit Med

Off by a beat: the effects of temporal misalignment in reinforcement learning for sepsis treatment

Shengpu Tang, Jiayu Yao, Jenna Wiens, and 1 more author

NPJ Digital Medicine, 2026

PDF Bib

@article{Tang2026Off,
  title = {Off by a beat: the effects of temporal misalignment in reinforcement learning for sepsis treatment},
  author = {Tang, Shengpu and Yao, Jiayu and Wiens, Jenna and Parbhoo, Sonali},
  journal = {NPJ Digital Medicine},
  volume = {9},
  number = {1},
  pages = {360},
  publisher = {Nature Publishing Group UK},
  year = {2026},
}

ICLR

The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives

Matthieu Bou, Nyal Patel, Arjun Jagota, and 2 more authors

In The Fourteenth International Conference on Learning Representations, 2026

PDF Bib Code

@inproceedings{Bou2026AlignmentAuditor,
  title = {The Alignment Auditor: A Bayesian Framework for Verifying and Refining LLM Objectives},
  author = {Bou, Matthieu and Patel, Nyal and Jagota, Arjun and Krishna, Satyapriya and Parbhoo, Sonali},
  booktitle = {The Fourteenth International Conference on Learning Representations},
  year = {2026},
}