Emergent-Language Doctrine

From the AI Research Compass wiki

Emergent-Language Doctrine: Agents should invent their own communication through shared tasks, so language is grounded in use rather than copied from human text.

Beliefs

The Emergent-Language Doctrine argues that a language learned by imitating the internet is borrowed, not owned. Human language arose because creatures needed to coordinate in a shared world, so the doctrine recreates that pressure: put agents in referential games, cooperative navigation or negotiation, give them a channel of discrete symbols, and let a protocol emerge. Study which conditions make that protocol compositional, efficient and interpretable, and you learn what language is for, not just what it looks like.

It sits low in the Harness half, because meaning lives in a population of interacting agents and their environment rather than in one network, and well toward the World side: it is about language, but language that must be grounded in perception and action. Its beyond-revolutionary shade, the most radical on the chart, reflects that it treats today’s corpus-trained models as a shortcut that skips the interesting part.

Arguments

  • Referential games between neural agents reliably produce communication protocols that let them solve tasks no single agent could solve alone.
  • Grounded multi-agent populations have developed compositional, word-like symbols when environmental pressure and channel constraints favoured them.
  • Emergent communication is a controlled laboratory for questions about the origins of compositionality that cannot be tested on fixed human corpora.

Critiques

  • Emergent protocols are often degenerate, exploiting low-level cues in ways that look nothing like human language and resist interpretation.
  • The resulting “languages” have tiny vocabularies and serve toy tasks, far from the expressive power language models already exhibit.
  • Agents that do not speak human language are of limited practical use, since people eventually have to understand them.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +0.65, y ≈ −0.68, paradigm ≈ +1.00.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data