Code, pre-registrations with frozen thresholds, result summaries and manuscript for a minimal controlled model of cultural transmission in a Lewis referential game. Two small neural agents invent a language over 64 objects; each generation is a fresh pair that first learns a limited set of its parent's (object, message) pairs from a persistent record and then trains on the game. A limited transmitted record breaks symmetry among comparably structured reconstructions available to independent learners, and carries the developmental state in which it was captured (the Snapshot Effect). CPU-only PyTorch; one-command replication of the pre-registered tests on fresh seeds (replicate.py). Every pre-registered outcome, including failed predictions, is reported. Designed, coded, run and written with generative AI tools under the direction of the author, who takes full responsibility for the result. Code: MIT. Text and figures: CC BY 4.0. The manuscript is also published as a separate preprint record (see Related works).
Ariel Elboim· Zenodo (CERN European Organi...· 0 citations
Code, pre-registrations with frozen thresholds, result summaries and manuscript for a minimal controlled model of cultural transmission in a Lewis referential game. Two small neural agents invent a language over 64 objects; each generation is a fresh pair that first learns a limited set of its parent's (object, message) pairs from a persistent record and then trains on the game. A limited transmitted record breaks symmetry among comparably structured reconstructions available to independent learners, and carries the developmental state in which it was captured (the Snapshot Effect). CPU-only PyTorch; one-command replication of the pre-registered tests on fresh seeds (replicate.py). Every pre-registered outcome, including failed predictions, is reported. Designed, coded, run and written with generative AI tools under the direction of the author, who takes full responsibility for the result. Code: MIT. Text and figures: CC BY 4.0. The manuscript is also published as a separate preprint record (see Related works).
Ariel Elboim· Zenodo (CERN European Organi...· 0 citations
In iterated learning of emergent languages, cultural transmission can align semantic category systems and steer learners toward structured, efficient regions of language space. We show that limited transmitted evidence can also break symmetry among comparably structured solutions, biasing independent learners toward a shared, historically contingent reconstruction despite substantial item-level turnover. The target and strength of coordination are empirically distinguishable: developmental provenance can shift what learners reconstruct without necessarily changing how strongly they coordinate. In a controlled referential-game model, anchors captured from an immature snapshot of a parent's language steer descendants toward the immature organization; in the main GRU learner, this occurs without detectable weakening of coordination but at a cost to systematicity. We call this the Snapshot Effect. A second architecture and a minimal reconstruction model recover the same symmetry-breaking and provenance-target pattern. The local propagation mechanism is regime-dependent: in the compressed languages of the GRU learner, class-matched anchors preferentially stabilize untaught neighbouring meanings that use the same convention, whereas this mechanism does not carry over to a near-injective learner. Here, symmetry breaking refers to historical evidence resolving underdetermination among multiple viable structured reconstructions, not to competition among labels. Culture need not create structure. It makes one structured possibility historical. Working paper, version 1.0. 13 pages plus a 17-page supplement (full methods, every pre-registered outcome including failures, a minimal reconstruction model, figure legends). Code, registrations and a one-command replication driver: https://doi.org/10.5281/zenodo.22305564. Generative AI tools assisted with code, execution, review and editing; the author takes full responsibility.
Ariel Elboim· Zenodo (CERN European Organi...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
