The CPG attended the International Conference on Machine Learning (ICML) in Vancouver on 14-19 of July 2025. Igor Shilov and Matthieu Meeus both presented posters on “memorisation in LLMs. Their work explores the effects of repetition in text data, and the influence of near-duplicates on memorisation. Their work raises important questions on contamination in LLM benchmarks, and the value of using distrribution-level analysis for understanding memorisation.
