About
I work to bridge the gap between benchmark performance and real-world AI capabilities. My research tests whether our experiments support the claims we make about AI, and builds methods whose gains hold up, across evaluation, post-training and world models.
I'm currently working on world models at Samsung AI Research, Cambridge, and recently submitted my PhD at the University of Edinburgh, advised by Timothy Hospedales and Henry Gouk. Along the way I held an A*STAR PhD Research Fellowship working on Singapore's S$70M National Multimodal LLM Programme, interned at Samsung AI Research, and co-founded FastAutomate, a GUI-automation agent startup.
I'm looking for research roles in evaluation, post-training and world models from March 2027.
Research areas
- World models & hierarchical planning current Hierarchical planners that commit only to sub-goals the world model predicts are reachable. Metro-WM, under review at ICLR 2027
- Test-time personalized alignment Steering LLMs to individual users at test time with reward-guided decoding. Showed that reward-model accuracy does not predict how well the model follows each user, and proposed policy accuracy to measure it directly. EMNLP 2026
- Safety-critical few-shot adaptation of foundation models Weight-space learning that gives certified performance guarantees when adapting foundation models from a few examples. ICML 2026
- Evaluating training algorithms for pre-training How evaluation protocols decide whether a new optimizer looks better than it is; a case study of VeLO's 4,000 TPU-months. NeurIPS 2023 ICBINB Workshop
News
- Oct 2026Submitted my PhD thesis at the University of Edinburgh.
- Sep 2026Metro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals is under review at ICLR 2027.
- Aug 2026The Reward Model Selection Crisis in Personalized Alignment accepted at EMNLP 2026.
- Jun 2026Joined Samsung AI Research, Cambridge as a Research Scientist Intern, working on neuro-symbolic hierarchical planning with latent world models.
- Apr 2026Weight Space Learning for Certifiable Few-shot Transfer Learning (STEEL) accepted at ICML 2026.
Older news
- Nov 2025FedP2EFT: Federated Learning to Personalize PEFT for Multilingual LLMs at AAAI 2026.
- Mar 2025Weight Space Learning for Certifiable Few-shot Transfer Learning selected for a Spotlight & Oral at the ICLR 2025 Weight Space Learning Workshop.
- Feb 2025Started the A*STAR PhD Research Fellowship in Singapore.
- Aug 2024Joined Samsung AI Research, Cambridge as a Machine Learning Research Engineer (Research Scientist Intern).
- 2024Liouna presented at the ICML 2024 WANT Workshop.
Selected publications
-
Under reviewMetro-WM: Long-Horizon Latent Planning with Realisable Sub-Goals Under review at ICLR 2027 Paper Code
-
EMNLP 2026
-
ICML 2026Weight Space Learning for Certifiable Few-shot Transfer Learning ICML 2026. ICLR 2025 Weight Space Learning Workshop (Spotlight & Oral) Paper
-
AAAI 2026FedP2EFT: Federated Learning to Personalize PEFT for Multilingual LLMs AAAI 2026 (Poster) Paper Code
-
PreprintEEVEE and GATE: Finding the Right Benchmarks and How to Run Them Seamlessly Preprint, submitted to the NeurIPS 2024 Datasets and Benchmarks Track Paper Code
-
ICML 2024 WorkshopLiouna: Biologically Plausible Learning for Efficient Pre-Training of Transferrable Deep Models ICML 2024 WANT Workshop Paper
-
NeurIPS 2023 WorkshopIs Scaling Learned Optimizers Worth It? Evaluating the Value of VeLO's 4000 TPU Months NeurIPS 2023 ICBINB Workshop Paper
Experience
Research
-
Jun 2026 – Feb 2027Research Scientist InternSamsung AI Research, Cambridge
Details
- Researching neuro-symbolic hierarchical planning with latent world models for long-horizon embodied robotics tasks.
- Co-authored Metro-WM, a long-horizon latent planning method that constrains high-level planning to realisable sub-goals (under review at ICLR 2027).
-
Feb 2025 – Feb 2026PhD Student FellowA*STAR, Singapore
Details
- Led research on personalized alignment for LLMs, showing that standard reward-model evaluation fails to predict downstream behaviour; published at EMNLP 2026.
- Proposed policy accuracy, a new evaluation metric for whether reward-guided LLMs correctly follow individual user preferences.
-
Aug 2024 – Feb 2025Research Scientist InternSamsung AI Research, Cambridge
Details
- Developed STEEL, the first framework to give certifiable performance guarantees for foundation models in few-shot transfer learning; published at ICML 2026.
- Scaled generative models over adapter parameter spaces to 1B parameters across vision and language tasks.
Industry
-
Jun 2021 – May 2023Co-Founder and Chief AI Officer
Details
- Co-founded an AI-powered GUI automation agent startup, leading technical strategy, product development and business operations.
- Delivered client engagements including HR automation pipelines.
- Built the team from the ground up, recruiting, training and managing engineers while balancing hands-on technical work with client delivery.
-
Jun 2021 – May 2023ML Department ManagerTMentors
Details
- Delivered a full-stack data science project for L’Azurdé, the largest jewellery e-commerce platform in the MENA region.
- Built end-to-end BI solutions for enterprise clients, including ETL pipelines, PowerBI integrations and real-time dashboards.
- Led team and project management, solution architecture, hiring, client relations, tender writing, DevOps and deployment.
-
Sep 2017 – Sep 2018Student Researcher, Software Engineering / Acting QA ManagerSharp Life Science · Oxford Science Park, UK
Details
- Designed a surrogate lumped-parameter model for electrowetting dynamics, significantly speeding up prototyping.
- Applied graph theory to validate, schedule and design laboratory workflows; built an extensible, temporally queryable DAG data structure.
- Designed and deployed a production domain-specific dataflow programming language in 10 weeks, still in active use.
- Introduced an integration testing framework across all software products; acted as QA Manager for the 2nd-generation platform release.
Education
-
Oct 2022 – Oct 2026PhD in Artificial IntelligenceUniversity of Edinburgh · Edinburgh, UKThesis submitted October 2026. Supervised by Prof. Timothy Hospedales and Dr. Henry Gouk.
-
Feb 2025 – Feb 2026PhD Research FellowshipA*STAR, Singapore
-
Sep 2020 – Aug 2021MSc in Artificial Intelligence, DistinctionUniversity of Edinburgh · Edinburgh, UK
-
Sep 2015 – Jun 2019BEng Computer Systems Engineering with a Year in Industry, First Class HonoursUniversity of Sheffield · Sheffield, UK