Leon Chlon
AI Research Lead, PhysicsX. Founder, Hassana Labs.
I study when language models genuinely learn from their context and when they take shortcuts: why they hallucinate, and how structured inductive biases, symmetry-aware learning and a Bayesian perspective can make them more reliable and easier to interpret. I also work on world models, multimodal learning and physics-informed machine learning.
I’m AI Research Lead in the Fundamentals team at PhysicsX. Before that I was a Visiting Fellow in the Torr Vision Group at the University of Oxford. I founded Hassana Labs, an independent non-profit AI research lab that opens doors for researchers from marginalised communities, and wrote a free book, Information Geometry for Generative Models, with donations going to Lebanese refugees.
Selected papers
-
When the evidence doesn’t carry enough information to support an answer, a model fills the gap with something plausible; the paper turns this into a rule for when to answer and when to abstain. Code
-
Exact Finite Attention Responses From RoPE Derivatives
An exact formula for how a model’s attention responds when parts of its input are moved, removed or changed.
-
Robot World Models Are Not Invariant to How the Actions Are Written
A robot world model trained on one way of writing actions breaks when it’s handed the same actions written another, equivalent way.
-
LLMs are Bayesian in Expectation, Not Realization
Averaged over the order of their examples, language models come close to ideal Bayesian reasoning; any single ordering can drift from it.
-
Information Geometry for Generative Models
Thirteen chapters, from compression and Bayesian prediction to transformers and diffusion. Originally bound for Oxford University Press and released free instead; donations support displaced Lebanese refugees.
Software
-
Berry formerly HallBayes
Checks whether an AI’s answer is supported by the evidence it cites, and flags the claims that aren’t. About 1,700 GitHub stars; featured at NVIDIA GTC 2026; deepset built a Haystack cookbook on it.
-
Adapts Hugging Face language models with small controllers in the forward pass instead of LoRA. Controllers fitted separately combine by addition. About 350 GitHub stars.
-
Measures how much a model’s predictions swing across symmetry-equivalent inputs, then distils a student model that averages the swing away. Recipes for robotics, weather, particle physics and fast physics surrogates.
-
Makes Triton work on NVIDIA’s GB10 (Blackwell) desktop hardware until official support arrives.
More research code: ITO, factuality-slice and aecf. I’ve also contributed to Google DeepMind’s Optax and to Hugging Face, including a 4× speed-up in SAM image processing.
News
- Sep 2026
Spoke at the Agentic AI Foundation in London on training-free self-improving agents, alongside speakers from Google DeepMind, Imperial, Prolific and Vercel. Slides
- Sep 2026
Two preprints: exact attention responses from RoPE derivatives, with Julie Huang, Maggie Chlon and Gregory Gutin (Royal Holloway), and robot world models that change their predictions when the same actions are written differently, with Ahmed Karim.
- Aug 2026
Quoted in WIRED on how the watermarks in AI-written text can be removed.
- Jun 2026
Predictable Compression Failures published at ICML 2026.
- 2026
Joined PhysicsX as AI Research Lead in the Fundamentals team.
- 2026
Berry, then called HallBayes, featured at NVIDIA GTC.
- 2026
Backing from Microsoft, Google and NVIDIA, worth $300K, for an AI-research sabbatical in Oxford’s Torr Vision Group.
- Dec 2025
One of Hassana Labs’ open-source releases raised £500 for a UNICEF fundraiser for Sudan.
- Sep 2025
HallBayes passed 1,000 GitHub stars in under a month.
- 2025
Triton kernels for NVIDIA’s Blackwell GB10, sponsored by HP and NVIDIA.
- 2025
Co-designed CUDA INT4 kernels with Omri Weinstein (Hebrew University of Jerusalem) and Unsloth, cutting the quantisation penalty by about 9% on GPTQ and GGUF (Qwen3.5).
Work
-
PhysicsX, AI Research Lead, Fundamentals since 2026
-
University of Oxford, AI Research Lead and Visiting Fellow, Torr Vision Group 2025–2026
Built HallBayes, Mezzanine and Triton kernels for Blackwell GB10; co-supervised joint PhD and lab projects across MIT, Cambridge and Oxford.
-
Apple, Content Ranking 2023–2025
A production multi-gate mixture-of-experts model on TensorFlow and ONNX Runtime; 43% lower median latency on workloads serving more than 50 million requests a day.
-
TikTok, Post-training, Ranking 2023–2025
Bayesian-optimised retraining of a mixture-of-experts ranking system with more than 500 million daily active users; double-digit gains in cross-platform sharing.
-
World Bank Group, Post-training 2023–2025
Fine-tuned LoRA adapters for Llama 2 70B across 300,000 socioeconomic indicators, cutting forecast error by 35% (sMAPE); adopted by 15 cross-country teams.
-
AI reliability advisory 2023–2025
Advised senior stakeholders at PwC, McKinsey, NVIDIA and Novartis, and government policymakers including SDAIA in Saudi Arabia.
-
Tailor Bio, Founding AI Engineer 2022–2023
A Cambridge spin-out. Led the founding machine-learning team and built the drug-discovery stack on AWS (PyTorch, RDKit, Neo4j), training graph neural network foundation models with custom sparse matrix multiplication and gradient checkpointing for three times the throughput. Three lead compounds were validated for the Series B.
-
Uber / Careem, Lead ML Engineer, Dynamic Pricing 2022
Marketplace pricing for more than 100 million users using algorithmic game theory: 2.7 times as efficient as the previous system, with more than $2M in revenue impact.
-
Meta, Senior Research Scientist, AI Safety 2020–2022
Bayesian reinforcement learning for crash detection across Instagram, WhatsApp and Facebook (10–15% better detection), and vector search with FAISS and HNSW that raised hate-speech recall by 15% at billion-post scale.
-
McKinsey & Company, Senior Data Scientist 2018–2020
Gradient-boosted credit-risk models for tier-one banks covering more than $100B in exposures, with over 40% better early warning; Basel III and IFRS 9 strategy for executives at three major institutions.
Talks and writing
- Talks at the Agentic AI Foundation, London (2026, slides); ICAIRE, Riyadh (2025); the AE Global Summit on Open Problems for AI, London (2025); and the UCL Data Science Society (2025). Invited talks at Imperial College London, the American University of Beirut and the University of Sharjah.
- Quoted by Isabella Ward in WIRED (August 2026).
- I explain research for a general audience on LinkedIn, Instagram and TikTok, with about 190,000 followers across the three.
Teaching and mentoring
- At Hassana Labs I’ve supervised three student researchers: one is now a PhD student at UCL, one at the University of Barcelona, and the third is finishing PhD applications. The lab’s research fellowships are for people from marginalised communities.
- Research mentoring of PhD and master’s students at MIT and Harvard, leading to co-authored papers.
- Teaching assistant for MIT 9.073/HST.460, Statistics for Neuroscience Research (Emery Brown), and guest teaching for Harvard SEAS ES 201/APMTH 231, Decision Theory (Demba Ba).
- A free careers workshop in Beirut (2024), a free AI careers course (2025) and a 24-hour AI research hackathon in London (2026).
Education
Harvard Medical School and MIT, Postdoctoral Fellow, Machine Learning 2018
University of Cambridge, PhD, Machine Learning 2017
University of Cambridge, MPhil, Theoretical Physics 2014
EWOR Fellowship