Leon Chlon
I'm AI Research Lead in the Fundamentals team at PhysicsX, building in-context learning models that predict surface fields such as pressure from a handful of example simulations, without retraining. Across physics and language models, my research asks when models genuinely learn from their context and when they take shortcuts: why language models hallucinate, and how structured inductive biases, symmetry-aware learning, and a Bayesian perspective can make them more reliable and interpretable.
I also work on world models, multimodal learning, and physics-informed ML, where principled modeling bridges theory with scalable systems. Previously Visiting Fellow at the University of Oxford (Torr Vision Group). I founded Hassana Labs, a research organization opening doors in AI for researchers from marginalized communities, and wrote an open-source book, Information Geometry for Generative Models, released free with donations going to Lebanese refugees.
News
- 2026 Spoke at AAIF London (Agentic AI Foundation) on Training-Free Self-Improving Agents, alongside speakers from Google DeepMind, Imperial, Prolific and Vercel; 687 people registered. Slides.
- 2026 Joined PhysicsX as AI Research Lead (Fundamentals team), building in-context learning models for physics.
- 2026 Interviewed by WIRED's Isabella Ward about cracking AI.
- 2026 Two new preprints as PI: exact attention responses from RoPE derivatives (with Prof. Gregory Gutin, Royal Holloway), and robot world models that change their predictions when the same actions are written differently.
- 2026 Student researchers I supervised as PI have gone on to PhDs at UCL and the University of Barcelona.
- 2026 Predictable Compression Failures published at ICML 2026.
- 2026 Won $300K in backing from Microsoft, Google, and NVIDIA to fund an AI-research sabbatical at Oxford's Torr Vision Group.
- 2026 HallBayes featured at NVIDIA GTC 2026; integrated into PyTorch Geometric (1.7k★, 160+ forks).
- 2026 HP invited me to review its new ZGX Fury workstation.
- 2025 Co-designed CUDA INT4 kernels with Prof. Omri Weinstein (Hebrew University of Jerusalem) and Unsloth; cut quantization penalty by ~9% on GPTQ/GGUF (Qwen3.5).
- 2025 HP & NVIDIA-sponsored Triton kernels for Blackwell GB10 (sm_121); 6.1× over PyTorch baselines.
- 2025 Released Mezzanine, a post-training toolkit adopted by DeepMind, CERN, DAMTP, and CRUKCI; powered the first pixel-free I-JEPA for robotics.
- 2025 HallBayes passed 1,000 GitHub stars in under a month; one of Hassana Labs' open-source releases raised £500 for a UNICEF fundraiser for Sudan.
Selected publications
-
Predictable Compression Failures: Order Sensitivity and Information Budgeting for Evidence-Grounded Binary Adjudication
-
Exact Finite Attention Responses From RoPE Derivatives
-
Robot World Models Are Not Invariant to How the Actions Are Written
-
LLMs are Bayesian, in Expectation, not in Realization
Open source
- hallbayes ★ 1.7k Model-agnostic post-training to suppress LLM hallucinations. 10,000+ industry integrations, including deepset and PwC; integrated into PyTorch Geometric. Featured at NVIDIA GTC 2026.
- ntkmirror ★ 351 LoRA-free, forward-pass fine-tuning for Hugging Face causal language models: frozen-model activation control for lightweight in-context learning and fine-tuning.
- mezzanine ★ 143 Post-training toolkit. Adopted by DeepMind, CERN, DAMTP, and CRUKCI. First pixel-free I-JEPA for robotics; 2.1× retrieval-rank improvement on LeRobot ALOHA.
- triton-blackwell ★ 2 Fused Triton transformer kernels for Blackwell GB10 (sm_121); 6.1× over PyTorch baselines.
- berry ★ 14 Information-budget MCP server for hallucination detection in agent workflows.
Other public research code: ITO (★65) · teta (★33) · factuality-slice (★31) · aecf (★4).
Core contributions to DeepMind Optax and HuggingFace; 4× speedup in SAM image processing. EWOR Fellowship (0.1% acceptance rate).
Selected work
- PhysicsX · AI Research Lead, Fundamentals 2026–present In-context learning models that predict surface fields from a handful of example simulations, without retraining. Developed the Gaussian SDPA now used across the team; geometry and field encoders; evaluations that separate real in-context learning from shortcuts, including on DrivAerML.
- University of Oxford · AI Research Lead, Visiting Fellow 2025–2026 Torr Vision Group. 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 Production MMoE on TensorFlow + ONNX Runtime; 43% p50-latency reduction on workloads serving 50M+ requests/day.
- TikTok · Post-training, Ranking 2023–2025 Bayesian-optimised MMoE retraining across a 500M+ DAU ranking system; double-digit gains in cross-platform sharing.
- World Bank Group · Post-training 2023–2025 Fine-tuned LLaMA-2 70B LoRa adapters across 300k socioeconomic indicators; reduced forecast error by 35% sMAPE; immediate adoption across 15 cross-country teams.
- AI reliability advisory 2023–2025 Consultations with senior stakeholders at PwC, McKinsey, NVIDIA, and Novartis, and with government policymakers including SDAIA (Saudi Arabia).
- Tailor Bio · Founding AI Engineer (0→1) 2022–2023 Cambridge spinout. Led the founding ML team; drug-discovery stack on AWS (PyTorch, RDKit, Neo4j) training GNN foundation models with custom sparse mat-mul + gradient checkpointing. 3× throughput; 3 lead compounds validated for Series B.
- Uber / Careem · Lead ML Engineer, Dynamic Pricing 2022 Marketplace pricing for 100M+ users via algorithmic game theory; 2.7× efficiency over legacy; $2M+ revenue impact.
- Meta · Senior Research Scientist, AI Safety 2020–2022 Bayesian RL for crash detection across Instagram / WhatsApp / Facebook (10–15% better detection); FAISS / HNSW vector embeddings improving hate-speech recall by 15% at billion-post scale.
- McKinsey & Company · Senior Data Scientist 2018–2020 Gradient-boosting credit-risk models for Tier-1 banks covering $100B+ exposures; >40% better early-warning detection; Basel III / IFRS 9 strategies for C-suite at three major institutions.
Talks & writing
Mentoring & teaching
Education
- Harvard Medical School / MIT 2018 Postdoctoral Fellow, Machine Learning
- University of Cambridge 2017 PhD, Machine Learning
- University of Cambridge 2014 MPhil, Theoretical Physics