Ilia Shumailov is an AI researcher affiliated with the University of Oxford whose published work spans training-data degradation, model security, and the economics of inference hardware. He is named as an author on three of the papers summarized in this collection.
Research
Shumailov is the first author of "The Curse of Recursion: Training on Generated Data Makes Models Forget" (2023), later published in Nature as "AI models collapse when trained on recursively generated data" (2024). That paper coined the term model collapse for the degradation that occurs when generative models are trained on their own outputs across successive generations, and it remains the canonical reference for the phenomenon (Model Collapse, Training Data Walls).
He is one of three equal-contributor authors, alongside Alexander Panfilov and David Schmotz, of "Stealing Reasoning Traces from Proprietary LLM APIs" (arXiv:2608.09867, August 10, 2026), which reports that the encrypted chain-of-thought blocks returned by Anthropic, OpenAI and Google APIs are interchangeable across sessions, users and models within a provider's ecosystem, and that injecting a frontier model's encrypted trace into a weaker sibling model causes the sibling to transcribe it in plaintext (Stealing Reasoning Traces from Proprietary LLM APIs).
He is a co-author of "DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs" (arXiv:2608.14614, July 10, 2026), which reports one year of production operation of a 128-GPU cluster built entirely from second-hand components and measures the cost and carbon trade-offs of repurposing retired datacenter GPUs for LLM inference (DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs). His affiliation on that paper is the University of Oxford; his co-authors are from the University of Cambridge and Quettaflop AI.