nucleation

CLAIM-066: Kirby, Cornish and Smith show structure emerging under transmission with an experimenter's filter, not filtering by measured inductive biases, and Griffiths and Kalish's convergence to the prior holds only for sampling learners

2 min read

Properties9
StatusSuperseded — by ARG-035 — Griffiths and Kalish's convergence theorem and Kirby, Cornish and Smith's chains show that serial transmission is filtered by the transmitters' inductive biases
Version2
Rolegranted
Warrantinterpretive
Tagslinguistics, cognition, compositionality
Linepragmatic-transport
Workswhat-survives-translation
GroundsTHEORY-155 — Iterated learning by Bayesian agents who sample from the posterior converges, in an ergodic chain, to the shared prior, LIT-771 — Cumulative cultural evolution in the laboratory: An experimental approach to the origins of structure in human language, LIT-769 — Language Evolution by Iterated Learning With Bayesian Agents
SupportsCLAIM-234 — The record's iterated-learning readings bear on lexical, structural and functional transmission, and none measures stance, footing or speech-act function along a chain

Graph View

The claim

Kirby et al. measured no prior and use no Bayesian learners. Without the homonym filter (Experiment 1) the languages collapsed to two to five words that underspecified meaning; compositional structure appeared only when the experimenter removed homonyms before each generation (Experiment 2), which is a selection on the data, the case Griffiths and Kalish’s analysis explicitly excludes. Griffiths and Kalish show convergence to the prior for learners who sample from the posterior; learners who take the most probable language depend also on how much data passes between generations, and can amplify weak biases (THEORY-155).

What it does not say

It does not undercut the manuscript’s §8 conclusion that a stable endpoint can be far from the ancestor: for sampling learners the stationary distribution is the prior, independent of the starting language. It says the three citations are not one finding, and that which learner model applies decides whether the endpoint reflects the learners or the transmission filter.

A chain without a filter

Kalish, Griffiths and Lewandowsky (LIT-820) supply the unfiltered chain with a bias stated in advance that Kirby, Cornish and Smith lack. They still cannot tell sampling learners from MAP learners.

Note of 2026-10-28: moved to ARG-035

This claim says what two sources show, and that the manuscript cites them as showing one thing. Its reading is already THEORY-155’s, homonym filter and MAP case included, so nothing goes to a THEORY, and its objection moved to ARG-035, from THEORY-155 and LIT-771 to CLAIM-090, form source, Proposed. CLAIM-234, the Active claim that rested on it, moved to an argument the same day, so ADR-061’s split rule, applied after the moves, does not hold it. The text above is unchanged.

Backlinks