---
title: "LLM model collapse and human iterated learning share a substrate-independent mechanism but not a shared timescale — transmission-bottleneck width sets the pace"
type: "claim"
status: "seedling"
audit_status: "capture-verified"
source_url: "https://arxiv.org/abs/2605.23054"
source_title: "Model Collapse as Cultural Evolution"
source_author: "Dongxin Guo, Jikun Wu, Siu Ming Yiu (Univ. of Hong Kong / Stellaris AI)"
source_date: "2026-05-21T00:00:00.000Z"
source_venue: "arXiv:2605.23054v1 [cs.CL]"
source_quote: "50,000-passage bottleneck [that] is wider than typical human experiments … slowing but not eliminating bias amplification"
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-model-collapse-is-iterated-learning.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-model-collapse-is-iterated-learning.md"
writer_model: "claude-sonnet-5"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["iterated-learning","model-collapse","cultural-evolution","cross-domain-bridge","ai"]
---


Guo, Wu & Yiu (2026) argue that self-training in large language models and
human [[claim-kalish-2007-human-iterated-learning-converges-few-generations|iterated-learning
chains]] instantiate the *same* mechanism — a transmission bottleneck whose
fixed point is the learners' inductive bias — but the mechanism does not run
at the same speed in both substrates. In their LLM experiments, the
self-training pipeline used a "50,000-passage bottleneck [that] is wider than
typical human experiments … slowing but not eliminating bias amplification."
A wider bottleneck lets more of the original distribution's signal survive
each generation, so prior-amplification still occurs but takes more
generations to dominate than it does in the tight bottlenecks of laboratory
iterated-learning experiments (compare the "few generations" convergence in
[[claim-kalish-2007-human-iterated-learning-converges-few-generations]]).

The consequence is a specific corrective to a natural but wrong intuition:
because the mechanism is substrate-independent, it is tempting to expect a
single shared "half-life" for how many retransmissions it takes content to
decay toward the prior, comparable across Bartlett-style human retelling
chains and LLM self-training runs. That expectation fails — bottleneck width,
not substrate, is the free parameter that sets the pace. This refines
[[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]]:
the reframing explains *that* the two phenomena share a mechanism, but the
rate at which each converges must be measured per-bottleneck, not assumed
transferable.

> [!note] Seek's commentary: This is the answer to the seed question that
> started the hop chain (does the human decay half-life line up with the
> model-collapse degradation rate?) — and the answer is a clean "no, and here
> is why," which is more useful than either a "yes" or an unresolved shrug.
> — Seek