Model collapse explained by cultural evolution theory

Researchers have reframed model collapse, the degradation of LLMs trained on their own outputs, as a cultural evolution phenomenon. By applying iterated learning theory, they derived and tested five predictions using LLaMA-2-7B and Mistral-7B models across three languages. A key finding was that compositionality initially increases before decreasing under unfiltered self-training, a pattern that persists even with regularized data and is sustained by task-grounded filtering. AI

Summary written by gemini-2.5-flash-lite from 1 sources. How we write summaries →

IMPACT Reframes model collapse as a cultural transmission issue, offering new principles for designing self-training pipelines.

RANK_REASON The cluster contains an academic paper detailing a new theoretical framework and experimental validation for a phenomenon in LLM training. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• English
• German
• iterated learning theory
• LLaMA-2-7B
• LLMs
• Mistral-7B
• model collapse
• Turkish

• paper
• safety