Papers

1. Robot World Models Are Not Invariant to How the Actions Are Written

   Ahmed Karim, Leon Chlon. arXiv preprint, September 2026. arXiv:2609.23252

   A robot world model trained on one way of writing actions breaks when it’s handed the same actions written another, equivalent way.

   World models and science

2. Exact Finite Attention Responses From RoPE Derivatives

   Julie Huang, Maggie Chlon, Gregory Gutin, Leon Chlon. arXiv preprint, September 2026. arXiv:2609.14127

   An exact formula for how a model’s attention responds when parts of its input are moved, removed or changed.

   In-context learning

3. Predictable Compression Failures: Order Sensitivity and Information Budgeting for Evidence-Grounded Binary Adjudication

   Leon Chlon, Ahmed Karim, Maggie Chlon, MarcAntonio Awada. ICML 2026. arXiv:2509.11208

   Treats hallucination as a measurable shortfall of information, and gives a rule for when a model should answer or abstain.

   Hallucination and verification

4. LLMs are Bayesian in Expectation, Not Realization

   Leon Chlon, Zein Khamis, Fatima Sheaib, Maggie Chlon, Mahdi El Zein, MarcAntonio M. Awada. arXiv preprint, 2025, revised 2026. arXiv:2507.11768

   Averaged over the order of their examples, language models come close to ideal Bayesian reasoning; any single ordering can drift from it.

   In-context learning

5. Information Geometry for Generative Models

   Leon Chlon. Open textbook, 2026.

   A free 13-chapter textbook on the maths behind generative AI, from compression and Bayesian prediction to transformers and diffusion.

Talk slides are on the About page. For a paper we haven’t listed, or a PDF that won’t open, email us.