Nature: AI models collapse when trained on recursively generated data
alex_verem · x · 2026-09-18
The original Nature paper (Shumailov et al., vol. 631, 2024) behind viral model-collapse threads: indiscriminate training on model-generated content causes irreversible defects, erasing the tails of the original distribution. 819k accesses, 878 citations.
More from Research
- Idea: replicate JEV-style probability estimates via softmax of LLM logprobs — ngxson · 2026-09-18
- msign Is Wrong for Full-Rank SVD Power Iteration, but Correct for Top-k Subspaces — YouJiacheng · 2026-09-18
- Human Organoids Fill 91.9% of Cortex in Engineered Mice, Nature Paper Reports — ctjlewis · 2026-09-18
- Lambda's EdiVal-Agent uses an agentic judge for image editing evals, hitting 81.3% human agreement — TheZachMueller · 2026-09-18
- Tahoe-100M cell state dataset published in Cell, downloaded 600K+ times — anshulkundaje · 2026-09-18
- AlphaGenome author pushes back: variant prioritization gains are real contributions — anshulkundaje · 2026-09-18