Ilia Shumailov

Researcher associated with Oxford, first author of shumailov-2024-model-collapse (Nature 2024) and of the earlier arXiv lineage paper “Curse of Recursion” (arXiv:2305.17493, 2023). The Nature article lists him with Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal; the article’s correspondence line reads “Correspondence to Ilia Shumailov, Zakhar Shumaylov or Yarin Gal.”

Also appears as a co-author on three unrelated NeurIPS 2022 papers (architecture adaptation, stochastic pre-processing defenses, vote-histogram leakage), which is how the widespread “NeurIPS 2022 curse of recursion” citation got tangled: the recursion papers are 2023 (arXiv) and 2024 (Nature), and a 2022 proceedings paper under that name was not found in this vault’s checks.

He also sits third in the author list of “Stealing Reasoning Traces from Proprietary LLM APIs” (arXiv:2608.09867, August 2026, Alexander Panfilov first author; abs page read 2026-10-04). The paper shows encrypted chain-of-thought blocks from major providers are interchangeable across sessions and decodable through weaker models in the same ecosystem. graepel-2026-llms-dont-reason cites it as evidence about chain-of-thought reports; see chain-of-thought-faithfulness for what the paper does and does not establish.

Sources

• Crossref record for DOI 10.1038/s41586-024-07566-y (author list), checked 2026-09 by the research run recorded in outputs/llm-degradation-toward-the-mean.provenance.md.
• NeurIPS 2022 proceedings listing (2,671 titles) grepped by a research worker 2026-09.