CtD: Composition through Decomposition in Emergent Communication

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Carmeli, Boaz, Meir, Ron, Belinkov, Yonatan
Natura: Preprint
Pubblicazione: 2026
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866917205159968768
author Carmeli, Boaz
Meir, Ron
Belinkov, Yonatan
author_facet Carmeli, Boaz
Meir, Ron
Belinkov, Yonatan
contents Compositionality is a cognitive mechanism that allows humans to systematically combine known concepts in novel ways. This study demonstrates how artificial neural agents acquire and utilize compositional generalization to describe previously unseen images. Our method, termed "Composition through Decomposition", involves two sequential training steps. In the 'Decompose' step, the agents learn to decompose an image into basic concepts using a codebook acquired during interaction in a multi-target coordination game. Subsequently, in the 'Compose' step, the agents employ this codebook to describe novel images by composing basic concepts into complex phrases. Remarkably, we observe cases where generalization in the `Compose' step is achieved zero-shot, without the need for additional training.
format Preprint
id arxiv_https___arxiv_org_abs_2601_10169
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle CtD: Composition through Decomposition in Emergent Communication
Carmeli, Boaz
Meir, Ron
Belinkov, Yonatan
Artificial Intelligence
Compositionality is a cognitive mechanism that allows humans to systematically combine known concepts in novel ways. This study demonstrates how artificial neural agents acquire and utilize compositional generalization to describe previously unseen images. Our method, termed "Composition through Decomposition", involves two sequential training steps. In the 'Decompose' step, the agents learn to decompose an image into basic concepts using a codebook acquired during interaction in a multi-target coordination game. Subsequently, in the 'Compose' step, the agents employ this codebook to describe novel images by composing basic concepts into complex phrases. Remarkably, we observe cases where generalization in the `Compose' step is achieved zero-shot, without the need for additional training.
title CtD: Composition through Decomposition in Emergent Communication
topic Artificial Intelligence
url https://arxiv.org/abs/2601.10169