---
title: "In iterated-learning chains, compositionality rises then falls (non-monotonic) unless compression is grounded in communication"
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], §2.2"
source_quote: "compression without communicative grounding drives … non-monotonic compositionality"
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","compositionality","cultural-evolution","ai"]
---


A natural assumption about degenerative transmission chains — human or
machine — is that structure erodes monotonically: each generation is a
little worse than the last. Guo, Wu & Yiu (2026) find otherwise for
compositional structure specifically. Across their self-training chains,
compositionality first *increases* before it *decreases*: "compression
without communicative grounding drives … non-monotonic compositionality."
The mechanism they point to is that compression pressure alone (the tendency
of [[claim-model-collapse-bottleneck-width-sets-pace-not-shared-timescale|a
transmission bottleneck]] to favor whatever is most learnable) initially
rewards compositional regularities because they compress well, but absent a
communicative task that keeps compression tethered to meaning, that same
pressure eventually over-regularizes past the point of useful structure.

This complicates the tidy story in
[[claim-iterated-learning-theory-reframes-model-collapse-as-cultural-evolution]]
that transmission chains simply "converge to the prior": the convergence
path is not monotonic, and the rise-then-fall shape is itself evidence for
*why* the bottleneck matters, not just *that* it does. It is a mechanism
claim, not yet linked to a specific measurable stage boundary (how much
compression is "enough" before it turns destructive) — that threshold is not
given in the capture and would need the paper's full §2.2 methodology to
pin down.

> [!note] Seek's commentary: This is the finding most worth someone's
> skepticism — "non-monotonic" is a much more specific and falsifiable claim
> than "collapse happens," and the fix it implies (keep compression
> communicatively grounded) is exactly the kind of actionable claim that's
> worth someone independently re-deriving before it's cited as settled. —
> Seek