How can a group of individuals who share no common language eventually agree on what to call the world around them? In traditional AI, we often solve this by hard-coding a shared ontology or providing a centralized dictionary. However, in the real world—and in robust multi-agent systems—meaning is not a static lookup table; it is a dynamic, self-organizing convention.
Researchers like Luc Steels have pioneered a "selectionist" approach to this problem. Instead of pre-programming definitions, agents "bootstrap" their own communication systems through iterative, grounded interactions known as Language Games.
The Mechanics of the Language Game
The core idea, as detailed in The Origins of Syntax in Visually Grounded Robotic Agents, is a decentralized feedback loop. Imagine two robots in a room with several objects.
- The Interaction: One agent (the speaker) chooses a target object and "invents" or selects a word to describe it. The second agent (the hearer) attempts to identify the target based on its own internal dictionary.
- The Feedback: If the hearer points to the correct object, the interaction is a success. If not, the speaker reveals the intended target.
- The Update: Both agents update their internal "scores" for that word-object mapping.
This process is entirely local. There is no central authority telling the agents which words are "correct." Global coherence emerges purely from the accumulation of these successful local interactions.
Lateral Inhibition: The Engine of Consensus
If every agent is inventing their own words, how does the population avoid a "Tower of Babel" scenario where everyone uses different synonyms? The solution is a mechanism called lateral inhibition, described by Luc Steels and Frédéric Kaplan in Situated Grounded Word Semantics.
Lateral inhibition acts as a "winner-take-all" reinforcement strategy:
- Success: When a word successfully communicates a meaning, the strength of that specific mapping increases. Crucially, the strength of all competing words for that same meaning is decreased.
- Failure: If a word leads to a misunderstanding, its strength is penalized.
This creates a competitive ecosystem of words. Over time, a single dominant term for an object "inhibits" its rivals into extinction. As Luc Steels and Michael Spranger observe in How Experience of the Body Shapes Language about Space, populations often experience a "lexical overshoot"—a burst of many different names—followed by a rapid collapse into a shared, efficient vocabulary.
Scaling to Syntax and Logic
The beauty of the language game is that it isn't limited to simple nouns. The same selectionist principles can evolve complex grammar. Josefina Sierra-Santibáñez has demonstrated that agents can evolve recursive language systems to express propositional logic, such as "not both up and to the left" (An Agent-Based Model of the Emergence and Transmission of a Language System for the Expression of Logical Combinations).
In these models, agents don't just track word-object pairs; they track the success of grammatical rules. If a specific word order or auxiliary word consistently helps a hearer parse a complex logical formula, that rule's score increases. This shows that structural complexity can emerge from the same bottom-up social pressure that creates simple names.
Grounding and Uncertainty
For these languages to be meaningful, they must be "grounded" in the physical world or a shared task. Roman Miletitch and colleagues showed that foraging robot swarms can autonomously name resources in their environment, effectively creating a collective map through communication (Emergent naming of resources in a foraging robot swarm).
However, communication is rarely perfect. Modern research by Robert X. D. Hawkins and Thomas L. Griffiths suggests that agents (and humans) maintain "lexical uncertainty"—a probabilistic map of what a partner might mean (Shades of confusion: Lexical uncertainty modulates ad hoc coordination in an interactive communication task). This uncertainty allows agents to be flexible, rapidly forming ad hoc conventions to resolve misunderstandings on the fly.
Why This Matters
This research direction shifts our understanding of language from a fixed code to a living, adaptive system. Whether it is deaf communities standardizing STEM signs through bottom-up consensus (Daniel S. Otero et al., ASL-STEM Forum) or future 6G networks where AI agents "negotiate" their own semantic protocols (Mehdi Bennis, Semantic Communication meets System 2 ML), the principle remains the same:
Meaning is not given; it is negotiated through the success and failure of shared action.